http://wiki.math.uwaterloo.ca/statwiki/api.php?action=feedcontributions&user=Ka2khan&feedformat=atomstatwiki - User contributions [US]2023-01-28T19:11:37ZUser contributionsMediaWiki 1.28.3http://wiki.math.uwaterloo.ca/statwiki/index.php?title=A_Bayesian_Perspective_on_Generalization_and_Stochastic_Gradient_Descent&diff=42323A Bayesian Perspective on Generalization and Stochastic Gradient Descent2018-12-07T03:58:58Z<p>Ka2khan: /* Introduction */</p>
<hr />
<div>==Introduction==<br />
This paper shows how Bayesian principles can explain many recent observations in the deep learning literature, and provide practical new insights. This work builds on Zhang et al.(2016) who showed that deep convolutional neural networks continues to perform well on the training data even after random relabelling of the input data points, but its ability to generalize and performance on the test data points goes down after the random relabelling. The authors consider two questions: how can we predict if a minimum will generalize to the test set, and why does stochastic gradient descent find minima that generalize well? <br />
<br />
The paper shows that the same phenomenon occurs even in small linear models. These observations are explained by the Bayesian evidence, which penalizes sharp minima but is invariant to model parameterization. They also demonstrate that, when one holds the learning rate fixed, there is an optimum batch size which maximizes the test set accuracy.<br />
<br />
The authors propose that the noise introduced by small mini-batches drives the parameters towards minima whose evidence is large. Interpreting stochastic gradient descent as a stochastic differential equation, we identify the “noise scale” <math display="inline"> g \approx \epsilon N/B </math> where <math display="inline">ε</math> is the learning rate, <math display="inline">N</math> the training set size and <math display="inline">B</math> the batch size. Consequently the optimum batch size is proportional to both the learning rate and the size of the training set, <math display="inline">B_{opt} \propto \epsilon N</math>. The authors verify these predictions empirically.<br />
<br />
==Motivation and Related Work==<br />
Zhang et al. (2016) trained deep convolutional networks on ImageNet and CIFAR10, achieving excellent accuracy on both training and test sets. They then took the same input images, but randomized the labels, and found that while their networks were now unable to generalize to the test set, they still memorized the training labels. They claimed these results contradict learning theory, although this claim is disputed (Kawaguchi et al., 2017; Dziugaite & Roy, 2017). Nonetheless, their results beg the question; if our models can assign arbitrary labels to the training set, why do they work so well in practice? <br />
<br />
Meanwhile, Keskar et al. (2016) observed that if we hold the learning rate fixed and increase the batch size, the test accuracy usually falls. This striking result shows improving the estimate of the full-batch gradient can harm performance. Goyal et al. (2017) observed a linear scaling rule between batch size and learning rate in a deep ResNet, while Hoffer et al. (2017) proposed a square root rule on theoretical grounds.<br />
<br />
Many authors have suggested “broad minima” whose curvature is small may generalize better than “sharp minima” whose curvature is large (Chaudhari et al., 2016; Hochreiter & Schmidhuber, 1997). Indeed, Dziugaite & Roy (2017) argued the results of Zhang et al. (2016) can be understood using “nonvacuous” PAC-Bayes generalization bounds which penalize sharp minima, while Keskar et al. (2016) showed stochastic gradient descent (SGD) finds wider minima as the batch size is reduced. However, Dinh et al. (2017) challenged this interpretation, by arguing that the curvature of a minimum can be arbitrarily increased by changing the model parameterization.<br />
<br />
==Contribution==<br />
<br />
The main contributions of this paper are to show that:<br />
* The results of Zhang et al. (2016) are not unique to deep learning; it is observed the same phenomenon in a small “over-parameterized” linear model. Overparameterization occurs when a model is able to effectively “remember” training data. This occurs when there are enough parameters that the system of equations ends up with an infinite number of possible solutions. One can see why this over-training would lead to poor results in test cases, as this “memorization” learns noise as opposed to the inherent structure of different classes. It is demonstrated that this phenomenon is straightforwardly understood by evaluating the Bayesian evidence in favor of each model, which penalizes sharp minima but is invariant to the model parameterization.<br />
* SGD integrates a stochastic differential equation whose “noise scale” <math>g &asymp; &epsilon;N/B</math>, where <math>\epsilon</math> is the learning rate, <math>N</math> training set size, and <math>B</math> batch size. Noise drives SGD away from sharp minima, and therefore there is an optimal batch size which maximizes the test set accuracy. This optimal batch size is '''proportional to the learning rate and training set size'''.<br />
<br />
Zhang et al. (2016) showed high training competency of neural networks under informative labels, but drastic overfitting on improper labels. This implies weak generalizability even when a small proportion of labels are improper. The authors show that generalization is strongly correlated with the Bayesian evidence, a weighted combination of the depth of a minimum (the cost function) and its breadth (the Occam factor). Bayesians tend to make distributional assumptions on gradient updates by adding isotropic Gaussian noise. This paper builds upon these Bayesian principles by driving SGD away from sharp minima, and towards broad minima (the more broad, the better generalization due to less influence from small perturbations within input). The stochastic differential equation used as a component of gradient updates effectively serves as injected noise that improves a network's generalizability.<br />
<br />
==Main Results==<br />
<br />
The weakly regularized model memorizes random labels, however, generalizes properly on informative labels. Besides, the predictions are overconfident. The authors also showed that the test accuracy peaks at an optimal batch size, if one holds the other SGD hyper-parameters constant. It is postulated that the optimum represents a tradeoff between depth and breadth in the Bayesian evidence. However it is the underlying scale of random fluctuations in the SGD dynamics which controls the tradeoff, not the batch size itself. Furthermore, this test accuracy peak shifts as the training set size rises. The authors observed that the best found batch size is proportional to the learning rate. This scaling rule allowed the authors to increase the learning rate by simultaneously increasing the batch size with no loss in test accuracy and no increase in computational cost, thus parallelism across multiple GPU's can be fully leveraged to easily decrease training time. The scaling rule could also be applied to production models by consequentially increasing the batch size as new training data is introduced.<br />
<br />
==Bayesian Model Comparison==<br />
<br />
===Introduction to Bayesian Statistics===<br />
Bayes' theorem is a fundamental theorem in Bayesian statistics, as it is used by Bayesian methods to update probabilities, which are degrees of belief, after obtaining new data. Given two events <math>A</math> and <math>B</math>, the conditional probability of <math>A</math> given <math>B </math> is true, Bayes theorem states that<br />
\begin{align*}\displaystyle P(A\mid B)={\frac {P(B\mid A)P(A)}{P(B)}}\end{align*}<br />
<br />
Bayesian networks are DAGs whose nodes represent variables in the Bayesian sense: they may be observable quantities, latent variables, unknown parameters or hypotheses. Edges represent conditional dependencies; nodes that are not connected (no path connects one node to another) represent variables that are conditionally independent of each other. Each node is associated with a probability function that takes, as input, a particular set of values for the node's parent variables, and gives (as output) the probability (or probability distribution, if applicable) of the variable represented by the node. For example, if <math>m </math> parent nodes represent <math>m </math> Boolean variables then the probability function could be represented by a table of <math>2^{m} </math> entries, one entry for each of the <math>2^{m} </math> possible parent combinations. <br />
<br />
===Bayesian Model Comparison in Neural Networks===<br />
MacKay (1992) applied Bayesian model comparison to neural networks. An overview is presented below. <br />
<br />
We first consider a classification model <math>M </math> with a single parameter <math>\omega </math>, training inputs <math>x </math> and training labels <math>y </math>. We can infer a posterior probability distribution over the parameter by applying Bayes theorem :<br />
<br />
\begin{align*}P(\omega\mid y,x;M) = \frac{P(y\mid \omega,x;M)P(\omega;M) }{P(y\mid x;M)}\end{align*}<br />
<br />
The likelihood, <math>P(y\mid \omega,x;M) = \Pi_i P(y_i\mid \omega,x_i;M) = e^{-H(\omega;M)} </math>, where <math>H(\omega;M) </math> denotes the cross-entropy of unique categorical labels. Using a Gaussian prior, <math>P(\omega;M) = \sqrt{\lambda/2\pi e^{-\lambda\omega^2/2}} </math>, and therefore the posterior probability density of the parameter given the training data, <math>P(\omega\mid y,x;M) \propto \sqrt{\lambda/2\pi e^{-C(\omega;M)}} </math>, where <math>C(\omega;M) = H(\omega;M) + \lambda\omega^2/2 </math> denotes the L2 regularized cross entropy, or “cost function”, and <math>\lambda </math> is the regularization coefficient. <br />
<br />
The value <math>\omega_0 </math> which minimizes the cost function lies at the maximum of this posterior. To predict an unknown label <math>y_t </math> of a new input <math>x_t </math>, we should compute the integral,<br />
<br />
\begin{align*} P(y_t\mid x_t,y,x;M) &= \int \frac{d\omega P(y_t\mid \omega,x_t;M)}{P(\omega\mid y,x;M)}\\ &= \frac{\int d \omega P(y_t \mid \omega ,x_t;M)e^{-C(\omega;M)}}{\int d \omega e^{-C(\omega;M)}} \end{align*}</math><br />
<br />
However, these integrals are dominated by the region near <math>\omega_0 </math> . We usually approximate <math>P(y_t\mid x_t,x,y;M) \approx P(y_t\mid \omega_0,x_t;M) </math>. Having minimized <math>C(\omega;M) </math> to find <math>\omega_0 </math>, we now wish to compare two different models and select the best one. We use the probability ratio<br />
<br />
\begin{align*}\frac{P(M_1\mid y,x)}{P (M_2\mid y, x)} = \frac{P(y\mid x;M_1) P(M_1)}{ P (y\mid x; M_2) P (M_2)} . \end{align*} <br />
<br />
The second factor on the right is the prior ratio, which describes which model is most plausible. To avoid unnecessary subjectivity, we usually set this to 1. Meanwhile the first factor on the right is the evidence ratio, which controls how much the training data changes our prior beliefs<br />
<br />
Germain et al. (2016) showed that maximizing the evidence (or “marginal likelihood”) minimizes a PAC-Bayes generalization bound. To compute it, we evaluate <br />
\begin{align*}P(y\mid x;M) &= \int d\omega P(y\mid \omega,x;M)P(\omega;M) \\ &=\sqrt{\frac{\lambda}{2\pi}}\int d \omega e^{C(\omega;M)}\end{align*}<br />
<br />
Notice that the evidence is computed by integrating out the parameters; and consequently it is invariant to the model parameterization. <br />
Since this integral is dominated by the region near the minimum <math>\omega_0 </math>, we can estimate the evidence by Taylor expanding <math>C(\omega; M) \approx C(\omega_0) + C′′(\omega_0)(\omega - \omega_0)^2/2</math>. This gives us<br />
<br />
\begin{align*} P(y\mid x;M) &\approx e^{-C(\omega_0)}\sqrt{\frac{\lambda}{2\pi}} \int d \omega e^{-C′′(\omega_0)(\omega - \omega_0)^2/2}\\ &= exp \big\{- C(\omega_0)-\frac{1}{2}\ln(C (\omega_0)/\lambda) \big\}.\end{align*}<br />
<br />
The evidence is controlled by the value of the cost function at the minimum, and by the logarithm of the ratio of the curvature about this minimum compared to the regularization constant. In models with many parameters <br />
\begin{align*} P(y\mid x;M) &\approx e^{-C(\omega_0)}\sqrt{\frac{\lambda}{2\pi}} \int d \omega e^{-C′′(\omega_0)(\omega - \omega_0)^2/2} \\ &= exp \big\{- C(\omega_0)-\frac{1}{2} \sum_{i=1}^p \ln (\lambda_i/\lambda) \big\}.\end{align*}<br />
<br />
Occam’s factor arises from the log ratio <math>\ln (\lambda_i/\lambda) </math> The Occam factor describes the fraction of the prior parameter space consistent with the data. Occam’s factor penalizes the amount of information the model must learn about the parameters to accurately model the training data. Since the fraction is always less than one, the authors propose to approximate <math>P(y\mid x;M) </math> away from local minima by only performing the summation over eigenvalues <math>\lambda_i \geq \lambda </math>.<br />
<br />
The authors compare evidence against a null model which assumes the labels are entirely random. This model has no parameters, and so the evidence is controlled by the likelihood alone. <math>P(y\mid x;NULL) = (1/n)^N = e^{-N \ln(n)} </math>, where <math>n </math> denotes the number of model classes and <math>N</math> the number of training labels. The evidence ratio :<br />
\begin{equation*}\frac{P(y\mid x;M) }{P(y\mid x;NULL) } = e ^{-E(\omega_0)} \end{equation*}<br />
<math>E(\omega_0) = C(\omega_0)-\frac{1}{2} \sum_{i=1}^p \ln (\lambda_i/\lambda) - N\ln (n) </math> is the log evidence ratio in favor of the null model.<br />
The authors assign confidence to the predictions of a model iff <math>E(\omega_0 < 0 </math>.<br />
<br />
The evidence supports the intuition that broad minima generalize better than sharp minima, but unlike the curvature it does not depend on the model parameterization. Dinh et al. (2017) showed one can increase the Hessian eigenvalues by rescaling the parameters, but they must simultaneously rescale the regularization coefficients, otherwise the model changes. Since Occam’s factor arises from the log ratio, <math>\ln (\lambda_i/\lambda) </math> , these two effects cancel out. Note however that while the evidence itself is invariant to model parameterization, one can find reparameterizations which change the approximate evidence after the Laplace approximation. It is difficult to evaluate the evidence for deep networks, as we cannot compute the Hessian of millions of parameters. Additionally, neural networks exhibit many equivalent minima, since we can permute the hidden units without changing the model. To compute the evidence we must carefully account for this “degeneracy”. The authors argue these issues are not a major limitation, since the intuition they build studying the evidence in simple cases will be sufficient to explain the results of both Zhang et al. (2016) and Keskar et al. (2016).<br />
<br />
==Bayes Theorem and Generalization==<br />
Zhang et al. (2016) showed that deep neural networks generalize well on training inputs with informative labels, but the same model can overfit on the same input images when the labels are randomized; perfectly memorizing the training set. To demonstrate that these observations are not unique to deep network, the authors use logistic regression. They form a small balanced training set comprising 800 images from MNIST, of which half have true label “0” and half true label “1”. The test set is balanced, comprising 5000 MNIST images of zeros and 5000 MNIST images of ones. There are two tasks. In the first task, the labels of both the training and test sets are randomized. In the second task, the labels are informative, matching the true MNIST labels. The model has 784 weights and 1 bias.<br />
<br />
The accuracy of the model predictions on both the training and test sets is shown in figure 1. When trained on the informative labels, the model generalizes well to the test set, so long as it is weakly regularized. However the model also perfectly memorizes the random labels, replicating the obser- vations of Zhang et al. (2016) in deep networks. No significant improvement in model performance is observed as the regularization coefficient increases. For completeness, we also evaluate the mean margin between training examples and the decision boundary. For both random and informative labels, the margin drops significantly as we reduce the regularization coefficient. When weakly regularized, the mean margin is roughly 50% larger for informative labels than for random labels.<br />
<br />
[[File:bg1.png|800px|thumb|center|]]<br />
<br />
Now consider figure 2, where we plot the mean cross-entropy of the model predictions, evaluated on both training and test sets, as well as the Bayesian log evidence ratio defined in the previous section. Looking first at the random label experiment in figure 2a, while the cross-entropy on the training set vanishes when the model is weakly regularized, the cross-entropy on the test set explodes. Not only does the model make random predictions, but it is extremely confident in those predictions. As the regularization coefficient is increased the test set cross-entropy falls, settling at <math>ln(2)</math>, the cross-entropy of assigning equal probability to both classes. Now consider the Bayesian evidence, which we evaluate on the training set. The log evidence ratio is large and positive when the model is weakly regularized, indicating that the model is exponentially less plausible than assigning equal probabilities to each class. As the regularization parameter is increased, the log evidence ratio falls, but it is always positive, indicating that the model can never be expected to generalize well.<br />
Now consider figure 2b (informative labels). Once again, the training cross-entropy falls to zero when the model is weakly regularized, while the test cross-entropy is high. Even though the model makes accurate predictions, those predictions are overconfident. As the regularization coefficient increases, the test cross-entropy falls below ln 2, indicating that the model is successfully gener- alizing to the test set. Now consider the Bayesian evidence. The log evidence ratio is large and positive when the model is weakly regularized, but as the regularization coefficient increases, the log evidence ratio drops below zero, indicating that the model is exponentially more plausible than assigning equal probabilities to each class. As we further increase the regularization, the log evi- dence ratio rises to zero while the test cross-entropy rises to <math>ln(2)</math>. Test cross-entropy and Bayesian evidence are strongly correlated, with minima at the same regularization strength.<br />
<br />
Bayesian model comparison has explained our results in a logistic regression. Meanwhile, Krueger et al. (2017) showed the largest Hessian eigenvalue also increased when training on random labels in deep networks, implying the evidence is falling. We conclude that Bayesian model comparison is quantitatively consistent with the results of Zhang et al. (2016) in linear models where we can compute the evidence, and qualitatively consistent with their results in deep networks where we cannot. Dziugaite & Roy (2017) recently demonstrated the results of Zhang et al. (2016) can also be understood by minimising a PAC-Bayes generalization bound which penalizes sharp minima.<br />
[[File:bg2.png|800px|thumb|center|]]<br />
==Bayes Theorem and Stochastic Gradient Descent ==<br />
<br />
We showed above that generalization is strongly correlated with the Bayesian evidence, a weighted combination of the depth of a minimum (the cost function) and its breadth (the Occam factor). Consequently Bayesians often add isotropic Gaussian noise to the gradient (Welling & Teh, 2011). In appendix A, we show this drives the parameters towards broad minima whose evidence is large. The noise introduced by small batch training is not isotropic, and its covariance matrix is a function of the parameter values, but empirically Keskar et al. (2016) found it has similar effects, driving the SGD away from sharp minima. This paper therefore proposes Bayesian principles also account for the “generalization gap”, whereby the test set accuracy often falls as the SGD batch size is increased (holding all other hyper-parameters constant). Since the gradient drives the SGD towards deep minima, while noise drives the SGD towards broad minima, we expect the test set performance to show a peak at an optimal batch size, which balances these competing contributions to the evidence.<br />
We were unable to observe a generalization gap in linear models (since linear models are convex there are no sharp minima to avoid). Instead we consider a shallow neural network with 800 hidden units and RELU hidden activations, trained on MNIST without regularization. We use SGD with a momentum parameter of 0.9. Unless otherwise stated, we use a constant learning rate of 1.0 which does not depend on the batch size or decay during training. Furthermore, we train on just 1000 images, selected at random from the MNIST training set. This enables us to compare small batch to full batch training. We emphasize that we are not trying to achieve optimal performance, but to study a simple model which shows a generalization gap between small and large batch training.<br />
In figure 3, we exhibit the evolution of the test accuracy and test cross-entropy during training. Our small batches are composed of 30 images, randomly sampled from the training set. Looking first at figure 3a, small batch training takes longer to converge, but after a thousand gradient updates a clear generalization gap in model accuracy emerges between small and large training batches. Now consider figure 3b. While the test cross-entropy for small batch training is lower at the end of training; the cross-entropy of both small and large training batches is increasing, indicative of over-fitting. Both models exhibit a minimum test cross-entropy, although after different numbers of gradient updates. Intriguingly, we show in appendix B that the generalization gap between small and large batch training shrinks significantly when we introduce L2 regularization.<br />
<br />
[[File:bg3.png|800px|thumb|center|]]<br />
<br />
From now on we focus on the test set accuracy (since this converges as the number of gradient updates increases). In figure 4a, we exhibit training curves for a range of batch sizes between 1 and 1000. We find that the model cannot train when the batch size <math>B \leq 10</math>. In figure 4b we plot the mean test set accuracy after 10,000 training steps. A clear peak emerges, indicating that there is indeed an optimum batch size which maximizes the test accuracy, consistent with Bayesian intuition. The results of Keskar et al. (2016) focused on the decay in test accuracy above this optimum batch size.<br />
[[File:bg4.png|800px|thumb|center|]]<br />
<br />
==Stochastic Differential Equations and Scaling Rules==<br />
The results showed above indicate that the test accuracy peaks at an optimal batch size, if one holds the other SGD hyper-parameters constant. It is argued that this peak arises from the tradeoff between depth and breadth in the Bayesian evidence. However it is not the batch size itself which controls this tradeoff, but the underlying scale of random fluctuations in the SGD dynamics. The following content identifies this SGD “noise scale”, and uses it to derive three scaling rules which predict how the optimal batch size depends on the learning rate, training set size and momentum coefficient. <br />
First, interpret gradient update, as the discrete update of a stochastic differential equation <br />
\begin{equation*}\frac{d\omega}{dt} = \frac{dC}{d\omega} + \eta(t)\end{equation*}<br />
<math>\eta</math> represents noise <math>\langle \eta(t) \rangle = 0</math> and <math> \langle \eta (t)\eta (t')\rangle = gF (\omega)\delta (t-t')</math>.<br />
<math>t</math> is a continous variable, and <math>F(\omega)</math> matrix describing the gradient covariances.<br />
The SGD noise scale is taken to be <math>g \approx \epsilon N/B</math> where <math>\epsilon</math> is the learning rate, <math>N</math> training set size and <math>B</math> the batch size.<br />
[[File:bg5.png|800px|thumb|center|]]<br />
[[File:bg6.png|800px|thumb|center|]]<br />
[[File:bg7.png|800px|thumb|center|]]<br />
The noise scale falls when the batch B<br />
size increases, consistent with our earlier observation of an optimal batch size Bopt while holding the other hyper-parameters fixed. Notice that one would equivalently observe an optimal learning rate if one held the batch size constant. A similar analysis of the SGD was recently performed by Mandt et al. (2017), although their treatment only holds near local minima where the covariances <math>F (ω)</math> are stationary. Our analysis holds throughout training, which is necessary since Keskar et al. (2016) found that the beneficial influence of noise was most pronounced at the start of training.<br />
When we vary the learning rate or the training set size, we should keep the noise scale fixed, which implies that <math>Bopt ∝ εN</math>. In figure 5a, we plot the test accuracy as a function of batch size after <math>(10000/ε)</math> training steps, for a range of learning rates. Exactly as predicted, the peak moves to the right as <math>ε</math> increases. Additionally, the peak test accuracy achieved at a given learning rate does not begin to fall until <math>ε ∼ 3</math>, indicating that there is no significant discretization error in integrating the stochastic differential equation below this point. Above this point, the discretization error begins to dominate and the peak test accuracy falls rapidly. In figure 5b, we plot the best observed batch size as a function of learning rate, observing a clear linear trend, <math>Bopt ∝ ε</math>. The error bars indicate the distance from the best observed batch size to the next batch size sampled in our experiments.<br />
<br />
This scaling rule allows us to increase the learning rate with no loss in test accuracy and no increase in computational cost, simply by simultaneously increasing the batch size. We can then exploit increased parallelism across multiple GPUs, reducing model training times (Goyal et al., 2017). A similar scaling rule was independently proposed by Jastrzebski et al. (2017) and Chaudhari & Soatto (2017), although neither work identifies the existence of an optimal noise scale. A number of authors have proposed adjusting the batch size adaptively during training (Friedlander & Schmidt, 2012; Byrd et al., 2012; De et al., 2017), while Balles et al. (2016) proposed linearly coupling the learning rate and batch size within this framework. In Smith et al. (2017), we show empirically that decaying the learning rate during training and increasing the batch size during training are equivalent.<br />
In figure 6a we exhibit the test set accuracy as a function of batch size, for a range of training set sizes after 10000 steps (<math>ε = 1</math> everywhere). Once again, the peak shifts right as the training set size rises, although the generalization gap becomes less pronounced as the training set size increases. In figure 6b, we plot the best observed batch size as a function of training set size; observing another linear trend, <math>Bopt ∝ N</math>. This scaling rule could be applied to production models, progressively growing the batch size as new training data is collected. We expect production datasets to grow considerably over time, and consequently large batch training is likely to become increasingly common.<br />
<math>B(1−m)</math> scale of conventional SGD as <math>m → 0</math>. When <math>m > 0</math>, we obtain an additional scaling rule <math>Bopt ∝ 1/(1 − m)</math>. This scaling rule predicts that the optimal batch size will increase when the momentum coefficient is increased. In figure 7a we plot the test set performance as a function of batch size after 10000 gradient updates (<math>ε = 1</math> everywhere), for a range of momentum coefficients. In figure 7b, we plot the best observed batch size as a function of the momentum coefficient, and fit our results to the scaling rule above; obtaining remarkably good agreement.<br />
<br />
==Critiques==<br />
<br />
#Bayesian statistics is not provably, at present, a theory that can be used to explain why a learning algorithm works. The Bayesian theory is too optimistic: we introduce a prior and model and then trust both implicitly. Relative to any particular prior and model (likelihood), the Bayesian posterior is the optimal summary of the data, but if either part is misspecified, then the Bayesian posterior carries no optimality guarantee. The prior is chosen for convenience here. <br />
#No discussions with respect to the analysis of information bottleneck which also discuss the generalization ability of the model. <br />
#No discussion on real online learning with streaming data where the total number of data points are unknown?<br />
#The paper presents how mini-batch noises with SGD can improve the performance of neural networks. However, the usefulness of the approach can be described and analyzed in greater details, if the author could provide the performance for various well-known real-life data.<br />
<br />
==Conclusion==<br />
<br />
The paper showed that mini-batch noise helps SGD to go away from sharp minima, and provided an evidence that there is an optimal optimum batch size for a maximum the test accuracy. Based on interpreting SGD as integrating stochastic differential equation, this batch size is proportional to the learning rate and the training set size. Moreover, the authors shown that <math>Bopt \propto 1/(1 − m) </math>, where <math>m</math> is the momentum coefficient. More analysis was done on the relation between the learning rate, effective learning rate, and batch size is presented in ICLR 2018, where the authors proved by experiments that all the benefits of decaying the learning rate are achieved by increasing the batch size in addition to reducing the number of parameter updates dramatically, and also were able use literature parameters without the need of any hyper parameter tuning (Samuel L. Smith, Pieter-Jan Kindermans, Chris Ying, Quoc V. Le).<br />
<br />
==References==<br />
<br />
#Alessandro Achille and Stefano Soatto. On the emergence of invariance and disentangling in deep representations. arXiv preprint arXiv:1706.01350, 2017.<br />
#Lukas Balles, Javier Romero, and Philipp Hennig. Coupling adaptive batch sizes with learning rates. arXiv preprint arXiv:1612.05086, 2016.<br />
#Richard H Byrd, Gillian M Chin, Jorge Nocedal, and Yuchen Wu. Sample size selection in optimization methods for machine learning. Mathematical programming, 134(1):127–155, 2012. <br />
#Pratik Chaudhari and Stefano Soatto. Stochastic gradient descent performs variational inference converges to limit cycles for deep networks. arXiv preprint arXiv:1710.11029, 2017.<br />
#Pratik Chaudhari, Anna Choromanska, Stefano Soatto, and Yann LeCun. Entropy-SGD: Biasing gradient descent into wide valleys. arXiv preprint arXiv:1611.01838, 2016.<br />
#Soham De, Abhay Yadav, David Jacobs, and Tom Goldstein. Automated inference with adaptive batches. In Artificial Intelligence and Statistics, pp. 1504–1513, 2017.<br />
#Laurent Dinh, Razvan Pascanu, Samy Bengio, and Yoshua Bengio. Sharp minima can generalize for deep nets. arXiv preprint arXiv:1703.04933, 2017.<br />
#Gintare Karolina Dziugaite and Daniel M Roy. Computing nonvacuous generalization bounds for deep (stochastic) neural networks with many more parameters than training data. arXiv preprint arXiv:1703.11008, 2017.<br />
#Michael P Friedlander and Mark Schmidt. Hybrid deterministic-stochastic methods for data fitting. SIAM Journal on Scientific Computing, 34(3):A1380–A1405, 2012.<br />
#Crispin W Gardiner. Handbook of Stochastic Methods, volume 4. Springer Berlin, 1985.<br />
#Pascal Germain, Francis Bach, Alexandre Lacoste, and Simon Lacoste-Julien. PAC-bayesian theory meets bayesian inference. In Advances in Neural Information Processing Systems, pp. 1884– 1892, 2016.<br />
#Priya Goyal, Piotr Dolla ́r, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, An- drew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch SGD: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017.<br />
#Stephen F Gull. Bayesian inductive inference and maximum entropy. In Maximum-entropy and Bayesian methods in science and engineering, pp. 53–74. Springer, 1988.<br />
#Geoffrey E Hinton and Drew Van Camp. Keeping the neural networks simple by minimizing the description length of the weights. In Proceedings of the sixth annual conference on Computational learning theory, pp. 5–13. ACM,1993.<br />
#Sepp Hochreiter and Ju ̈rgen Schmidhuber. Flat minima. Neural Computation, 9(1):1–42, 1997. Elad Hoffer, Itay Hubara, and Daniel Soudry. Train longer, generalize better: closing the generalization gap in large batch training of neural networks. arXiv preprint arXiv:1705.08741, 2017.<br />
#Stanisław Jastrzebski, Zachary Kenton, Devansh Arpit, Nicolas Ballas, Asja Fischer, Yoshua Bengio, and Amos Storkey. Three factors influencing minima in SGD. arXiv preprint arXiv:1711.04623, 2017.<br />
#Robert E Kass and Adrian E Raftery. Bayes factors. Journal of the american statistical association, 90(430):773–795, 1995.<br />
#Kenji Kawaguchi, Leslie Pack Kaelbling, and Yoshua Bengio. Generalization in deep learning. arXiv preprint arXiv:1710.05468, 2017.<br />
#Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Pe- ter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. arXiv preprint arXiv:1609.04836, 2016.<br />
#David Krueger, Nicolas Ballas, Stanislaw Jastrzebski, Devansh Arpit, Maxinder S Kanwal, Tegan Maharaj, Emmanuel Bengio, Asja Fischer, and Aaron Courville. Deep nets don’t learn via mem- orization. ICLR Workshop, 2017.<br />
#Qianxiao Li, Cheng Tai, and E Weinan. Stochastic modified equations and adaptive stochastic gradient algorithms. In International Conference on Machine Learning, pp. 2101–2110, 2017.<br />
#David JC MacKay. A practical bayesian framework for backpropagation networks. Neural compu- tation, 4(3):448–472, 1992.<br />
#Stephan Mandt, Matthew D Hoffman, and David M Blei. Stochastic gradient descent as approximate bayesian inference. arXiv preprint arXiv:1704.04289, 2017.<br />
#Ravid Shwartz-Ziv and Naftali Tishby. Opening the black box of deep neural networks via informa- tion. arXiv preprint arXiv:1703.00810, 2017.<br />
#Samuel L. Smith, Pieter-Jan Kindermans, and Quoc V. Le. Don’t decay the learning rate, increase the batch size. arXiv preprint arXiv:1711.00489, 2017.<br />
#Max Welling and Yee W Teh. Bayesian learning via stochastic gradient langevin dynamics. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pp. 681–688, 2011.<br />
#Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. arXiv preprint arXiv:1611.03530, 2016.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=A_Bayesian_Perspective_on_Generalization_and_Stochastic_Gradient_Descent&diff=42322A Bayesian Perspective on Generalization and Stochastic Gradient Descent2018-12-07T03:52:15Z<p>Ka2khan: /* Introduction */</p>
<hr />
<div>==Introduction==<br />
This paper shows how Bayesian principles can explain many recent observations in the deep learning literature, and provide practical new insights. This work builds on the work done by Zhang et al.(2016). In their work, they trained deep convolutional neural networks which achieved high accuracy on both the test and training data. They then randomly reassigned the labels for the same input images and found out that their model was now not performing well on the test set but could still easily memorize training data labels. Zhang et al.(2016) claimed that this finding is contradictory to the learning theory. This conclusion was disputed by others (Kawaguchi et al., 2017; Dziugaite & Roy, 2017). The authors consider two questions: how can we predict if a minimum will generalize to the test set, and why does stochastic gradient descent find minima that generalize well? <br />
<br />
The paper shows that the same phenomenon occurs even in small linear models. These observations are explained by the Bayesian evidence, which penalizes sharp minima but is invariant to model parameterization. They also demonstrate that, when one holds the learning rate fixed, there is an optimum batch size which maximizes the test set accuracy.<br />
<br />
The authors propose that the noise introduced by small mini-batches drives the parameters towards minima whose evidence is large. Interpreting stochastic gradient descent as a stochastic differential equation, we identify the “noise scale” <math display="inline"> g \approx \epsilon N/B </math> where <math display="inline">ε</math> is the learning rate, <math display="inline">N</math> the training set size and <math display="inline">B</math> the batch size. Consequently the optimum batch size is proportional to both the learning rate and the size of the training set, <math display="inline">B_{opt} \propto \epsilon N</math>. The authors verify these predictions empirically.<br />
<br />
==Motivation and Related Work==<br />
Zhang et al. (2016) trained deep convolutional networks on ImageNet and CIFAR10, achieving excellent accuracy on both training and test sets. They then took the same input images, but randomized the labels, and found that while their networks were now unable to generalize to the test set, they still memorized the training labels. They claimed these results contradict learning theory, although this claim is disputed (Kawaguchi et al., 2017; Dziugaite & Roy, 2017). Nonetheless, their results beg the question; if our models can assign arbitrary labels to the training set, why do they work so well in practice? <br />
<br />
Meanwhile, Keskar et al. (2016) observed that if we hold the learning rate fixed and increase the batch size, the test accuracy usually falls. This striking result shows improving the estimate of the full-batch gradient can harm performance. Goyal et al. (2017) observed a linear scaling rule between batch size and learning rate in a deep ResNet, while Hoffer et al. (2017) proposed a square root rule on theoretical grounds.<br />
<br />
Many authors have suggested “broad minima” whose curvature is small may generalize better than “sharp minima” whose curvature is large (Chaudhari et al., 2016; Hochreiter & Schmidhuber, 1997). Indeed, Dziugaite & Roy (2017) argued the results of Zhang et al. (2016) can be understood using “nonvacuous” PAC-Bayes generalization bounds which penalize sharp minima, while Keskar et al. (2016) showed stochastic gradient descent (SGD) finds wider minima as the batch size is reduced. However, Dinh et al. (2017) challenged this interpretation, by arguing that the curvature of a minimum can be arbitrarily increased by changing the model parameterization.<br />
<br />
==Contribution==<br />
<br />
The main contributions of this paper are to show that:<br />
* The results of Zhang et al. (2016) are not unique to deep learning; it is observed the same phenomenon in a small “over-parameterized” linear model. Overparameterization occurs when a model is able to effectively “remember” training data. This occurs when there are enough parameters that the system of equations ends up with an infinite number of possible solutions. One can see why this over-training would lead to poor results in test cases, as this “memorization” learns noise as opposed to the inherent structure of different classes. It is demonstrated that this phenomenon is straightforwardly understood by evaluating the Bayesian evidence in favor of each model, which penalizes sharp minima but is invariant to the model parameterization.<br />
* SGD integrates a stochastic differential equation whose “noise scale” <math>g &asymp; &epsilon;N/B</math>, where <math>\epsilon</math> is the learning rate, <math>N</math> training set size, and <math>B</math> batch size. Noise drives SGD away from sharp minima, and therefore there is an optimal batch size which maximizes the test set accuracy. This optimal batch size is '''proportional to the learning rate and training set size'''.<br />
<br />
Zhang et al. (2016) showed high training competency of neural networks under informative labels, but drastic overfitting on improper labels. This implies weak generalizability even when a small proportion of labels are improper. The authors show that generalization is strongly correlated with the Bayesian evidence, a weighted combination of the depth of a minimum (the cost function) and its breadth (the Occam factor). Bayesians tend to make distributional assumptions on gradient updates by adding isotropic Gaussian noise. This paper builds upon these Bayesian principles by driving SGD away from sharp minima, and towards broad minima (the more broad, the better generalization due to less influence from small perturbations within input). The stochastic differential equation used as a component of gradient updates effectively serves as injected noise that improves a network's generalizability.<br />
<br />
==Main Results==<br />
<br />
The weakly regularized model memorizes random labels, however, generalizes properly on informative labels. Besides, the predictions are overconfident. The authors also showed that the test accuracy peaks at an optimal batch size, if one holds the other SGD hyper-parameters constant. It is postulated that the optimum represents a tradeoff between depth and breadth in the Bayesian evidence. However it is the underlying scale of random fluctuations in the SGD dynamics which controls the tradeoff, not the batch size itself. Furthermore, this test accuracy peak shifts as the training set size rises. The authors observed that the best found batch size is proportional to the learning rate. This scaling rule allowed the authors to increase the learning rate by simultaneously increasing the batch size with no loss in test accuracy and no increase in computational cost, thus parallelism across multiple GPU's can be fully leveraged to easily decrease training time. The scaling rule could also be applied to production models by consequentially increasing the batch size as new training data is introduced.<br />
<br />
==Bayesian Model Comparison==<br />
<br />
===Introduction to Bayesian Statistics===<br />
Bayes' theorem is a fundamental theorem in Bayesian statistics, as it is used by Bayesian methods to update probabilities, which are degrees of belief, after obtaining new data. Given two events <math>A</math> and <math>B</math>, the conditional probability of <math>A</math> given <math>B </math> is true, Bayes theorem states that<br />
\begin{align*}\displaystyle P(A\mid B)={\frac {P(B\mid A)P(A)}{P(B)}}\end{align*}<br />
<br />
Bayesian networks are DAGs whose nodes represent variables in the Bayesian sense: they may be observable quantities, latent variables, unknown parameters or hypotheses. Edges represent conditional dependencies; nodes that are not connected (no path connects one node to another) represent variables that are conditionally independent of each other. Each node is associated with a probability function that takes, as input, a particular set of values for the node's parent variables, and gives (as output) the probability (or probability distribution, if applicable) of the variable represented by the node. For example, if <math>m </math> parent nodes represent <math>m </math> Boolean variables then the probability function could be represented by a table of <math>2^{m} </math> entries, one entry for each of the <math>2^{m} </math> possible parent combinations. <br />
<br />
===Bayesian Model Comparison in Neural Networks===<br />
MacKay (1992) applied Bayesian model comparison to neural networks. An overview is presented below. <br />
<br />
We first consider a classification model <math>M </math> with a single parameter <math>\omega </math>, training inputs <math>x </math> and training labels <math>y </math>. We can infer a posterior probability distribution over the parameter by applying Bayes theorem :<br />
<br />
\begin{align*}P(\omega\mid y,x;M) = \frac{P(y\mid \omega,x;M)P(\omega;M) }{P(y\mid x;M)}\end{align*}<br />
<br />
The likelihood, <math>P(y\mid \omega,x;M) = \Pi_i P(y_i\mid \omega,x_i;M) = e^{-H(\omega;M)} </math>, where <math>H(\omega;M) </math> denotes the cross-entropy of unique categorical labels. Using a Gaussian prior, <math>P(\omega;M) = \sqrt{\lambda/2\pi e^{-\lambda\omega^2/2}} </math>, and therefore the posterior probability density of the parameter given the training data, <math>P(\omega\mid y,x;M) \propto \sqrt{\lambda/2\pi e^{-C(\omega;M)}} </math>, where <math>C(\omega;M) = H(\omega;M) + \lambda\omega^2/2 </math> denotes the L2 regularized cross entropy, or “cost function”, and <math>\lambda </math> is the regularization coefficient. <br />
<br />
The value <math>\omega_0 </math> which minimizes the cost function lies at the maximum of this posterior. To predict an unknown label <math>y_t </math> of a new input <math>x_t </math>, we should compute the integral,<br />
<br />
\begin{align*} P(y_t\mid x_t,y,x;M) &= \int \frac{d\omega P(y_t\mid \omega,x_t;M)}{P(\omega\mid y,x;M)}\\ &= \frac{\int d \omega P(y_t \mid \omega ,x_t;M)e^{-C(\omega;M)}}{\int d \omega e^{-C(\omega;M)}} \end{align*}</math><br />
<br />
However, these integrals are dominated by the region near <math>\omega_0 </math> . We usually approximate <math>P(y_t\mid x_t,x,y;M) \approx P(y_t\mid \omega_0,x_t;M) </math>. Having minimized <math>C(\omega;M) </math> to find <math>\omega_0 </math>, we now wish to compare two different models and select the best one. We use the probability ratio<br />
<br />
\begin{align*}\frac{P(M_1\mid y,x)}{P (M_2\mid y, x)} = \frac{P(y\mid x;M_1) P(M_1)}{ P (y\mid x; M_2) P (M_2)} . \end{align*} <br />
<br />
The second factor on the right is the prior ratio, which describes which model is most plausible. To avoid unnecessary subjectivity, we usually set this to 1. Meanwhile the first factor on the right is the evidence ratio, which controls how much the training data changes our prior beliefs<br />
<br />
Germain et al. (2016) showed that maximizing the evidence (or “marginal likelihood”) minimizes a PAC-Bayes generalization bound. To compute it, we evaluate <br />
\begin{align*}P(y\mid x;M) &= \int d\omega P(y\mid \omega,x;M)P(\omega;M) \\ &=\sqrt{\frac{\lambda}{2\pi}}\int d \omega e^{C(\omega;M)}\end{align*}<br />
<br />
Notice that the evidence is computed by integrating out the parameters; and consequently it is invariant to the model parameterization. <br />
Since this integral is dominated by the region near the minimum <math>\omega_0 </math>, we can estimate the evidence by Taylor expanding <math>C(\omega; M) \approx C(\omega_0) + C′′(\omega_0)(\omega - \omega_0)^2/2</math>. This gives us<br />
<br />
\begin{align*} P(y\mid x;M) &\approx e^{-C(\omega_0)}\sqrt{\frac{\lambda}{2\pi}} \int d \omega e^{-C′′(\omega_0)(\omega - \omega_0)^2/2}\\ &= exp \big\{- C(\omega_0)-\frac{1}{2}\ln(C (\omega_0)/\lambda) \big\}.\end{align*}<br />
<br />
The evidence is controlled by the value of the cost function at the minimum, and by the logarithm of the ratio of the curvature about this minimum compared to the regularization constant. In models with many parameters <br />
\begin{align*} P(y\mid x;M) &\approx e^{-C(\omega_0)}\sqrt{\frac{\lambda}{2\pi}} \int d \omega e^{-C′′(\omega_0)(\omega - \omega_0)^2/2} \\ &= exp \big\{- C(\omega_0)-\frac{1}{2} \sum_{i=1}^p \ln (\lambda_i/\lambda) \big\}.\end{align*}<br />
<br />
Occam’s factor arises from the log ratio <math>\ln (\lambda_i/\lambda) </math> The Occam factor describes the fraction of the prior parameter space consistent with the data. Occam’s factor penalizes the amount of information the model must learn about the parameters to accurately model the training data. Since the fraction is always less than one, the authors propose to approximate <math>P(y\mid x;M) </math> away from local minima by only performing the summation over eigenvalues <math>\lambda_i \geq \lambda </math>.<br />
<br />
The authors compare evidence against a null model which assumes the labels are entirely random. This model has no parameters, and so the evidence is controlled by the likelihood alone. <math>P(y\mid x;NULL) = (1/n)^N = e^{-N \ln(n)} </math>, where <math>n </math> denotes the number of model classes and <math>N</math> the number of training labels. The evidence ratio :<br />
\begin{equation*}\frac{P(y\mid x;M) }{P(y\mid x;NULL) } = e ^{-E(\omega_0)} \end{equation*}<br />
<math>E(\omega_0) = C(\omega_0)-\frac{1}{2} \sum_{i=1}^p \ln (\lambda_i/\lambda) - N\ln (n) </math> is the log evidence ratio in favor of the null model.<br />
The authors assign confidence to the predictions of a model iff <math>E(\omega_0 < 0 </math>.<br />
<br />
The evidence supports the intuition that broad minima generalize better than sharp minima, but unlike the curvature it does not depend on the model parameterization. Dinh et al. (2017) showed one can increase the Hessian eigenvalues by rescaling the parameters, but they must simultaneously rescale the regularization coefficients, otherwise the model changes. Since Occam’s factor arises from the log ratio, <math>\ln (\lambda_i/\lambda) </math> , these two effects cancel out. Note however that while the evidence itself is invariant to model parameterization, one can find reparameterizations which change the approximate evidence after the Laplace approximation. It is difficult to evaluate the evidence for deep networks, as we cannot compute the Hessian of millions of parameters. Additionally, neural networks exhibit many equivalent minima, since we can permute the hidden units without changing the model. To compute the evidence we must carefully account for this “degeneracy”. The authors argue these issues are not a major limitation, since the intuition they build studying the evidence in simple cases will be sufficient to explain the results of both Zhang et al. (2016) and Keskar et al. (2016).<br />
<br />
==Bayes Theorem and Generalization==<br />
Zhang et al. (2016) showed that deep neural networks generalize well on training inputs with informative labels, but the same model can overfit on the same input images when the labels are randomized; perfectly memorizing the training set. To demonstrate that these observations are not unique to deep network, the authors use logistic regression. They form a small balanced training set comprising 800 images from MNIST, of which half have true label “0” and half true label “1”. The test set is balanced, comprising 5000 MNIST images of zeros and 5000 MNIST images of ones. There are two tasks. In the first task, the labels of both the training and test sets are randomized. In the second task, the labels are informative, matching the true MNIST labels. The model has 784 weights and 1 bias.<br />
<br />
The accuracy of the model predictions on both the training and test sets is shown in figure 1. When trained on the informative labels, the model generalizes well to the test set, so long as it is weakly regularized. However the model also perfectly memorizes the random labels, replicating the obser- vations of Zhang et al. (2016) in deep networks. No significant improvement in model performance is observed as the regularization coefficient increases. For completeness, we also evaluate the mean margin between training examples and the decision boundary. For both random and informative labels, the margin drops significantly as we reduce the regularization coefficient. When weakly regularized, the mean margin is roughly 50% larger for informative labels than for random labels.<br />
<br />
[[File:bg1.png|800px|thumb|center|]]<br />
<br />
Now consider figure 2, where we plot the mean cross-entropy of the model predictions, evaluated on both training and test sets, as well as the Bayesian log evidence ratio defined in the previous section. Looking first at the random label experiment in figure 2a, while the cross-entropy on the training set vanishes when the model is weakly regularized, the cross-entropy on the test set explodes. Not only does the model make random predictions, but it is extremely confident in those predictions. As the regularization coefficient is increased the test set cross-entropy falls, settling at <math>ln(2)</math>, the cross-entropy of assigning equal probability to both classes. Now consider the Bayesian evidence, which we evaluate on the training set. The log evidence ratio is large and positive when the model is weakly regularized, indicating that the model is exponentially less plausible than assigning equal probabilities to each class. As the regularization parameter is increased, the log evidence ratio falls, but it is always positive, indicating that the model can never be expected to generalize well.<br />
Now consider figure 2b (informative labels). Once again, the training cross-entropy falls to zero when the model is weakly regularized, while the test cross-entropy is high. Even though the model makes accurate predictions, those predictions are overconfident. As the regularization coefficient increases, the test cross-entropy falls below ln 2, indicating that the model is successfully gener- alizing to the test set. Now consider the Bayesian evidence. The log evidence ratio is large and positive when the model is weakly regularized, but as the regularization coefficient increases, the log evidence ratio drops below zero, indicating that the model is exponentially more plausible than assigning equal probabilities to each class. As we further increase the regularization, the log evi- dence ratio rises to zero while the test cross-entropy rises to <math>ln(2)</math>. Test cross-entropy and Bayesian evidence are strongly correlated, with minima at the same regularization strength.<br />
<br />
Bayesian model comparison has explained our results in a logistic regression. Meanwhile, Krueger et al. (2017) showed the largest Hessian eigenvalue also increased when training on random labels in deep networks, implying the evidence is falling. We conclude that Bayesian model comparison is quantitatively consistent with the results of Zhang et al. (2016) in linear models where we can compute the evidence, and qualitatively consistent with their results in deep networks where we cannot. Dziugaite & Roy (2017) recently demonstrated the results of Zhang et al. (2016) can also be understood by minimising a PAC-Bayes generalization bound which penalizes sharp minima.<br />
[[File:bg2.png|800px|thumb|center|]]<br />
==Bayes Theorem and Stochastic Gradient Descent ==<br />
<br />
We showed above that generalization is strongly correlated with the Bayesian evidence, a weighted combination of the depth of a minimum (the cost function) and its breadth (the Occam factor). Consequently Bayesians often add isotropic Gaussian noise to the gradient (Welling & Teh, 2011). In appendix A, we show this drives the parameters towards broad minima whose evidence is large. The noise introduced by small batch training is not isotropic, and its covariance matrix is a function of the parameter values, but empirically Keskar et al. (2016) found it has similar effects, driving the SGD away from sharp minima. This paper therefore proposes Bayesian principles also account for the “generalization gap”, whereby the test set accuracy often falls as the SGD batch size is increased (holding all other hyper-parameters constant). Since the gradient drives the SGD towards deep minima, while noise drives the SGD towards broad minima, we expect the test set performance to show a peak at an optimal batch size, which balances these competing contributions to the evidence.<br />
We were unable to observe a generalization gap in linear models (since linear models are convex there are no sharp minima to avoid). Instead we consider a shallow neural network with 800 hidden units and RELU hidden activations, trained on MNIST without regularization. We use SGD with a momentum parameter of 0.9. Unless otherwise stated, we use a constant learning rate of 1.0 which does not depend on the batch size or decay during training. Furthermore, we train on just 1000 images, selected at random from the MNIST training set. This enables us to compare small batch to full batch training. We emphasize that we are not trying to achieve optimal performance, but to study a simple model which shows a generalization gap between small and large batch training.<br />
In figure 3, we exhibit the evolution of the test accuracy and test cross-entropy during training. Our small batches are composed of 30 images, randomly sampled from the training set. Looking first at figure 3a, small batch training takes longer to converge, but after a thousand gradient updates a clear generalization gap in model accuracy emerges between small and large training batches. Now consider figure 3b. While the test cross-entropy for small batch training is lower at the end of training; the cross-entropy of both small and large training batches is increasing, indicative of over-fitting. Both models exhibit a minimum test cross-entropy, although after different numbers of gradient updates. Intriguingly, we show in appendix B that the generalization gap between small and large batch training shrinks significantly when we introduce L2 regularization.<br />
<br />
[[File:bg3.png|800px|thumb|center|]]<br />
<br />
From now on we focus on the test set accuracy (since this converges as the number of gradient updates increases). In figure 4a, we exhibit training curves for a range of batch sizes between 1 and 1000. We find that the model cannot train when the batch size <math>B \leq 10</math>. In figure 4b we plot the mean test set accuracy after 10,000 training steps. A clear peak emerges, indicating that there is indeed an optimum batch size which maximizes the test accuracy, consistent with Bayesian intuition. The results of Keskar et al. (2016) focused on the decay in test accuracy above this optimum batch size.<br />
[[File:bg4.png|800px|thumb|center|]]<br />
<br />
==Stochastic Differential Equations and Scaling Rules==<br />
The results showed above indicate that the test accuracy peaks at an optimal batch size, if one holds the other SGD hyper-parameters constant. It is argued that this peak arises from the tradeoff between depth and breadth in the Bayesian evidence. However it is not the batch size itself which controls this tradeoff, but the underlying scale of random fluctuations in the SGD dynamics. The following content identifies this SGD “noise scale”, and uses it to derive three scaling rules which predict how the optimal batch size depends on the learning rate, training set size and momentum coefficient. <br />
First, interpret gradient update, as the discrete update of a stochastic differential equation <br />
\begin{equation*}\frac{d\omega}{dt} = \frac{dC}{d\omega} + \eta(t)\end{equation*}<br />
<math>\eta</math> represents noise <math>\langle \eta(t) \rangle = 0</math> and <math> \langle \eta (t)\eta (t')\rangle = gF (\omega)\delta (t-t')</math>.<br />
<math>t</math> is a continous variable, and <math>F(\omega)</math> matrix describing the gradient covariances.<br />
The SGD noise scale is taken to be <math>g \approx \epsilon N/B</math> where <math>\epsilon</math> is the learning rate, <math>N</math> training set size and <math>B</math> the batch size.<br />
[[File:bg5.png|800px|thumb|center|]]<br />
[[File:bg6.png|800px|thumb|center|]]<br />
[[File:bg7.png|800px|thumb|center|]]<br />
The noise scale falls when the batch B<br />
size increases, consistent with our earlier observation of an optimal batch size Bopt while holding the other hyper-parameters fixed. Notice that one would equivalently observe an optimal learning rate if one held the batch size constant. A similar analysis of the SGD was recently performed by Mandt et al. (2017), although their treatment only holds near local minima where the covariances <math>F (ω)</math> are stationary. Our analysis holds throughout training, which is necessary since Keskar et al. (2016) found that the beneficial influence of noise was most pronounced at the start of training.<br />
When we vary the learning rate or the training set size, we should keep the noise scale fixed, which implies that <math>Bopt ∝ εN</math>. In figure 5a, we plot the test accuracy as a function of batch size after <math>(10000/ε)</math> training steps, for a range of learning rates. Exactly as predicted, the peak moves to the right as <math>ε</math> increases. Additionally, the peak test accuracy achieved at a given learning rate does not begin to fall until <math>ε ∼ 3</math>, indicating that there is no significant discretization error in integrating the stochastic differential equation below this point. Above this point, the discretization error begins to dominate and the peak test accuracy falls rapidly. In figure 5b, we plot the best observed batch size as a function of learning rate, observing a clear linear trend, <math>Bopt ∝ ε</math>. The error bars indicate the distance from the best observed batch size to the next batch size sampled in our experiments.<br />
<br />
This scaling rule allows us to increase the learning rate with no loss in test accuracy and no increase in computational cost, simply by simultaneously increasing the batch size. We can then exploit increased parallelism across multiple GPUs, reducing model training times (Goyal et al., 2017). A similar scaling rule was independently proposed by Jastrzebski et al. (2017) and Chaudhari & Soatto (2017), although neither work identifies the existence of an optimal noise scale. A number of authors have proposed adjusting the batch size adaptively during training (Friedlander & Schmidt, 2012; Byrd et al., 2012; De et al., 2017), while Balles et al. (2016) proposed linearly coupling the learning rate and batch size within this framework. In Smith et al. (2017), we show empirically that decaying the learning rate during training and increasing the batch size during training are equivalent.<br />
In figure 6a we exhibit the test set accuracy as a function of batch size, for a range of training set sizes after 10000 steps (<math>ε = 1</math> everywhere). Once again, the peak shifts right as the training set size rises, although the generalization gap becomes less pronounced as the training set size increases. In figure 6b, we plot the best observed batch size as a function of training set size; observing another linear trend, <math>Bopt ∝ N</math>. This scaling rule could be applied to production models, progressively growing the batch size as new training data is collected. We expect production datasets to grow considerably over time, and consequently large batch training is likely to become increasingly common.<br />
<math>B(1−m)</math> scale of conventional SGD as <math>m → 0</math>. When <math>m > 0</math>, we obtain an additional scaling rule <math>Bopt ∝ 1/(1 − m)</math>. This scaling rule predicts that the optimal batch size will increase when the momentum coefficient is increased. In figure 7a we plot the test set performance as a function of batch size after 10000 gradient updates (<math>ε = 1</math> everywhere), for a range of momentum coefficients. In figure 7b, we plot the best observed batch size as a function of the momentum coefficient, and fit our results to the scaling rule above; obtaining remarkably good agreement.<br />
<br />
==Critiques==<br />
<br />
#Bayesian statistics is not provably, at present, a theory that can be used to explain why a learning algorithm works. The Bayesian theory is too optimistic: we introduce a prior and model and then trust both implicitly. Relative to any particular prior and model (likelihood), the Bayesian posterior is the optimal summary of the data, but if either part is misspecified, then the Bayesian posterior carries no optimality guarantee. The prior is chosen for convenience here. <br />
#No discussions with respect to the analysis of information bottleneck which also discuss the generalization ability of the model. <br />
#No discussion on real online learning with streaming data where the total number of data points are unknown?<br />
#The paper presents how mini-batch noises with SGD can improve the performance of neural networks. However, the usefulness of the approach can be described and analyzed in greater details, if the author could provide the performance for various well-known real-life data.<br />
<br />
==Conclusion==<br />
<br />
The paper showed that mini-batch noise helps SGD to go away from sharp minima, and provided an evidence that there is an optimal optimum batch size for a maximum the test accuracy. Based on interpreting SGD as integrating stochastic differential equation, this batch size is proportional to the learning rate and the training set size. Moreover, the authors shown that <math>Bopt \propto 1/(1 − m) </math>, where <math>m</math> is the momentum coefficient. More analysis was done on the relation between the learning rate, effective learning rate, and batch size is presented in ICLR 2018, where the authors proved by experiments that all the benefits of decaying the learning rate are achieved by increasing the batch size in addition to reducing the number of parameter updates dramatically, and also were able use literature parameters without the need of any hyper parameter tuning (Samuel L. Smith, Pieter-Jan Kindermans, Chris Ying, Quoc V. Le).<br />
<br />
==References==<br />
<br />
#Alessandro Achille and Stefano Soatto. On the emergence of invariance and disentangling in deep representations. arXiv preprint arXiv:1706.01350, 2017.<br />
#Lukas Balles, Javier Romero, and Philipp Hennig. Coupling adaptive batch sizes with learning rates. arXiv preprint arXiv:1612.05086, 2016.<br />
#Richard H Byrd, Gillian M Chin, Jorge Nocedal, and Yuchen Wu. Sample size selection in optimization methods for machine learning. Mathematical programming, 134(1):127–155, 2012. <br />
#Pratik Chaudhari and Stefano Soatto. Stochastic gradient descent performs variational inference converges to limit cycles for deep networks. arXiv preprint arXiv:1710.11029, 2017.<br />
#Pratik Chaudhari, Anna Choromanska, Stefano Soatto, and Yann LeCun. Entropy-SGD: Biasing gradient descent into wide valleys. arXiv preprint arXiv:1611.01838, 2016.<br />
#Soham De, Abhay Yadav, David Jacobs, and Tom Goldstein. Automated inference with adaptive batches. In Artificial Intelligence and Statistics, pp. 1504–1513, 2017.<br />
#Laurent Dinh, Razvan Pascanu, Samy Bengio, and Yoshua Bengio. Sharp minima can generalize for deep nets. arXiv preprint arXiv:1703.04933, 2017.<br />
#Gintare Karolina Dziugaite and Daniel M Roy. Computing nonvacuous generalization bounds for deep (stochastic) neural networks with many more parameters than training data. arXiv preprint arXiv:1703.11008, 2017.<br />
#Michael P Friedlander and Mark Schmidt. Hybrid deterministic-stochastic methods for data fitting. SIAM Journal on Scientific Computing, 34(3):A1380–A1405, 2012.<br />
#Crispin W Gardiner. Handbook of Stochastic Methods, volume 4. Springer Berlin, 1985.<br />
#Pascal Germain, Francis Bach, Alexandre Lacoste, and Simon Lacoste-Julien. PAC-bayesian theory meets bayesian inference. In Advances in Neural Information Processing Systems, pp. 1884– 1892, 2016.<br />
#Priya Goyal, Piotr Dolla ́r, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, An- drew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch SGD: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017.<br />
#Stephen F Gull. Bayesian inductive inference and maximum entropy. In Maximum-entropy and Bayesian methods in science and engineering, pp. 53–74. Springer, 1988.<br />
#Geoffrey E Hinton and Drew Van Camp. Keeping the neural networks simple by minimizing the description length of the weights. In Proceedings of the sixth annual conference on Computational learning theory, pp. 5–13. ACM,1993.<br />
#Sepp Hochreiter and Ju ̈rgen Schmidhuber. Flat minima. Neural Computation, 9(1):1–42, 1997. Elad Hoffer, Itay Hubara, and Daniel Soudry. Train longer, generalize better: closing the generalization gap in large batch training of neural networks. arXiv preprint arXiv:1705.08741, 2017.<br />
#Stanisław Jastrzebski, Zachary Kenton, Devansh Arpit, Nicolas Ballas, Asja Fischer, Yoshua Bengio, and Amos Storkey. Three factors influencing minima in SGD. arXiv preprint arXiv:1711.04623, 2017.<br />
#Robert E Kass and Adrian E Raftery. Bayes factors. Journal of the american statistical association, 90(430):773–795, 1995.<br />
#Kenji Kawaguchi, Leslie Pack Kaelbling, and Yoshua Bengio. Generalization in deep learning. arXiv preprint arXiv:1710.05468, 2017.<br />
#Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Pe- ter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. arXiv preprint arXiv:1609.04836, 2016.<br />
#David Krueger, Nicolas Ballas, Stanislaw Jastrzebski, Devansh Arpit, Maxinder S Kanwal, Tegan Maharaj, Emmanuel Bengio, Asja Fischer, and Aaron Courville. Deep nets don’t learn via mem- orization. ICLR Workshop, 2017.<br />
#Qianxiao Li, Cheng Tai, and E Weinan. Stochastic modified equations and adaptive stochastic gradient algorithms. In International Conference on Machine Learning, pp. 2101–2110, 2017.<br />
#David JC MacKay. A practical bayesian framework for backpropagation networks. Neural compu- tation, 4(3):448–472, 1992.<br />
#Stephan Mandt, Matthew D Hoffman, and David M Blei. Stochastic gradient descent as approximate bayesian inference. arXiv preprint arXiv:1704.04289, 2017.<br />
#Ravid Shwartz-Ziv and Naftali Tishby. Opening the black box of deep neural networks via informa- tion. arXiv preprint arXiv:1703.00810, 2017.<br />
#Samuel L. Smith, Pieter-Jan Kindermans, and Quoc V. Le. Don’t decay the learning rate, increase the batch size. arXiv preprint arXiv:1711.00489, 2017.<br />
#Max Welling and Yee W Teh. Bayesian learning via stochastic gradient langevin dynamics. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pp. 681–688, 2011.<br />
#Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. arXiv preprint arXiv:1611.03530, 2016.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Learning_to_Teach&diff=42321Learning to Teach2018-12-07T02:25:51Z<p>Ka2khan: /* Problem Definition */</p>
<hr />
<div><br />
<br />
=Introduction=<br />
<br />
This paper proposed the "learning to teach" (L2T) framework with two intelligent agents: a student model/agent, corresponding to the learner in traditional machine learning algorithms, and a teacher model/agent, determining the appropriate data, loss function, and hypothesis space to facilitate the learning of the student model.<br />
<br />
In modern human society, the role of teaching is heavily implicated in our education system; the goal is to equip students with the necessary knowledge and skills in an efficient manner. This is the fundamental ''student'' and ''teacher'' framework on which education stands. However, in the field of artificial intelligence (AI) and specifically machine learning, researchers have focused most of their efforts on the ''student'' (ie. designing various optimization algorithms to enhance the learning ability of intelligent agents). The paper argues that a formal study on the role of ‘teaching’ in AI is required. Analogous to teaching in human society, the teaching framework can: select training data that corresponds to the appropriate teaching materials (e.g. textbooks selected for the right difficulty), design loss functions that correspond to targeted examinations, and define the hypothesis space that corresponds to imparting the proper methodologies. Furthermore, an optimization framework (instead of heuristics) should be used to update the teaching skills based on the feedback from students, so as to achieve teacher-student co-evolution.<br />
<br />
Thus, the training phase of L2T would have several episodes of interactions between the teacher and the student model. Based on the state information in each step, the teacher model would update the teaching actions so that the student model could perform better on the Machine Learning problem. The student model would then provide reward signals back to the teacher model. These reward signals are used by the teacher model as part of the Reinforcement Learning process to update its parameters. In this paper policy gradient algorithm is incorporated. This process is end-to-end trainable and the authors are convinced that once converged, the teacher model could be applied to new learning scenarios and even new students, without extra efforts on re-training.<br />
<br />
To demonstrate the practical value of the proposed approach, the '''training data scheduling''' problem is chosen as an example. The authors show that by using the proposed method to adaptively select the most<br />
suitable training data, they can significantly improve the accuracy and convergence speed of various neural networks including multi-layer perceptron (MLP), convolutional neural networks (CNNs)<br />
and recurrent neural networks (RNNs), for different applications including image classification and text understanding.<br />
Further more , the teacher model obtained by the paper from one task can be smoothly transferred to other tasks. As an example, the teacher model trained on MNIST with the MLP learner, one can achieve a satisfactory performance on CIFAR-10 only using roughly half<br />
of the training data to train a ResNet model as the student.<br />
<br />
=Related Work=<br />
The L2T framework connects with two emerging trends in machine learning. The first is the movement from simple to advanced learning. This includes meta-learning (Schmidhuber, 1987; Thrun & Pratt, 2012) which explores automatic learning by transferring learned knowledge from meta tasks [1]. This approach has been applied to few-shot learning scenarios and in designing general optimizers and neural network architectures. (Hochreiter et al., 2001; Andrychowicz et al., 2016; Li & Malik, 2016; Zoph & Le, 2017)<br />
<br />
The second is the teaching, which can be classified into either machine-teaching (Zhu, 2015) [2] or hardness based methods. The former seeks to construct a minimal training set for the student to learn a target model (ie. an oracle). The latter assumes an order of data from easy instances to hard ones, hardness being determined in different ways, is beneficial to the learning process. In curriculum learning (CL) (Bengio et al, 2009; Spitkovsky et al. 2010; Tsvetkov et al, 2016) [3] measures hardness through heuristics based understanding of the data while self-paced learning (SPL) (Kumar et al., 2010; Lee & Grauman, 2011; Jiang et al., 2014; Supancic & Ramanan, 2013) [4] measures hardness by loss on data. Another teaching method called 'pedagogical teaching' with applications in inverse reinforcement learning is quite close, in setup, to the method being proposed in this paper. The similarity can be observed in the way the teacher adjusts its behaviour to facilitate student learning and in the way the teacher communicates with the student. <br />
<br />
The limitations of these works include the lack of formal definition of the teaching problem whereas a learning problem has been formal mathematical definition. This makes it difficult to differentiate between teaching and learning problems. Other limitations are the reliance on heuristics and fixed rules, which hinders generalization of the teaching task.<br />
<br />
=Learning to Teach=<br />
To introduce the problem and framework, without loss of generality, consider the setting of supervised learning.<br />
<br />
In supervised learning, each sample <math>x</math> is from a fixed but unknown distribution <math>P(x)</math>, and the corresponding label <math> y </math> is from a fixed but unknown distribution <math>P(y|x) </math>. The goal is to find a function <math>f_\omega(x)</math> with parameter vector <math>\omega</math> that minimizes the gap between the predicted label and the actual label.<br />
<br />
<br />
<br />
==Problem Definition==<br />
In supervised learning, the goal is to choose a function <math display="inline">f_w(x)</math> with <math display="inline">w</math> as the parameter vector to predict the supervisor's label as good as possible. The goodness of a function <math display="inline">f_w</math> is evaluated by the risk function: <br />
<br />
\begin{align*}R(w) = \int M(y, f_w(x))dP(x,y)\end{align*}<br />
<br />
where <math display="inline">\mathcal{M}(,)</math> is the metric which evaluate the gap between the label and the prediction.<br />
<br />
The student model, denoted &mu;(), takes the set of training data <math> D </math>, the function class <math> Ω </math>, and loss function <math> L </math> as input to output a function, <math> f(ω) </math>, with parameter <math>ω^*</math> which minimizes risk <math>R(ω)</math> as in:<br />
<br />
\begin{align*}<br />
ω^* = arg min_{w \in \Omega} \sum_{x,y \in D} L(y, f_ω(x)) =: \mu (D, L, \Omega)<br />
\end{align*}<br />
<br />
The teaching model, denoted φ, tries to provide <math> D </math>, <math> L </math>, and <math> Ω </math> (or any combination, denoted <math> A </math>) to the student model such that the student model either achieves lower risk R(ω) or progresses as fast as possible.<br />
In contrast to traditional machine learning, which is only concerned with the student model in the<br />
learning to teach framework, the problem in the paper is also concerned with a teacher model, which tries to provide<br />
appropriate inputs to the student model so that it can achieve low risk functional as efficiently<br />
as possible.<br />
<br />
<br />
::'''Training Data''': Outputting a good training set <math> D \in \mathcal{D} </math>, where <math>\mathcal{D}</math> is the Borel set on the input space and label space. This is analogous to human teachers providing students with proper learning materials such as textbooks. <br />
::'''Loss Function''': Designing a good loss function <math> L \in \mathcal{L} </math>, where <math>\mathcal{L}</math> is the set of all possible loss functions. This is analogous to providing useful assessment criteria for students.<br />
::'''Hypothesis Space''': Defining a good function class <math> Ω \in \mathcal{W}</math>, where <math>\mathcal{W}</math> is the set of possible hypothesis spaces, which the student model can select from. This is analogous to human teachers providing appropriate context, eg. middle school students taught math with basic algebra while undergraduate students are taught with calculus. Different Ω leads to different errors and optimization problem (Mohri et al., 2012).<br />
<br />
==Framework==<br />
The training phase consists of the teacher providing the student with the subset <math> A_{train} </math> of <math> A </math> and then taking feedback to improve its own parameters.After the convergence of the training process,<br />
the teacher model can be used to teach either<br />
new student models, or the same student<br />
models in new learning scenarios such as another<br />
subset <math> A_{test} </math>is provided.Such a generalization is feasible as long as the state representations<br />
S are the same across different student<br />
models and different scenarios. The L2T process is outlined in figure below:<br />
<br />
[[File: L2T_process.png | 500px|center]]<br />
<br />
* <math> s_t &isin; S </math> represents information available to the teacher model at time <math> t </math>. <math> s_t </math> is typically constructed from the current student model <math> f_{t−1} </math> and the past teaching history of the teacher model. <math> S </math> represents the set of states.<br />
* <math> a_t &isin; A </math> represents action taken the teacher model at time <math> t </math>, given state <math>s_t</math>. <math> A </math> represents the set of actions, where the action(s) can be any combination of teaching tasks involving the training data, loss function, and hypothesis space. <br />
* <math> φ_θ : S → A </math> is policy used by the teacher model to generate its action <math> φ_θ(s_t) = a_t </math><br />
* Student model takes <math> a_t </math> as input and outputs function <math> f_t </math>, by using the conventional ML techniques.<br />
<br />
Mathematically, taking data teaching as an example where <math>L</math> <math>/Omega</math> as fixed, the objective of teacher in the L2T framework is <br />
<br />
<center> <math>\max\limits_{\theta}{\sum\limits_{t}{r_t}} = \max\limits_{\theta}{\sum\limits_{t}{r(f_t)}} = \max\limits_{\theta}{\sum\limits_{t}{r(\mu(\phi_{\theta}(s_t), L, \Omega))}}</math> </center><br />
<br />
Once the training process converges, the teacher model may be utilized to teach a different subset of <math> A </math> or teach a different student model.<br />
<br />
=Application=<br />
<br />
There are different approaches to training the teacher model, this paper will apply reinforcement learning with <math> φ_θ </math> being the ''policy'' that interacts with <math> S </math>, the ''environment''. The paper applies data teaching to train a deep neural network student, <math> f </math>, for several classification tasks. Thus the student feedback measure will be classification accuracy. Its learning rule will be mini-batch stochastic gradient descent, where batches of data will arrive sequentially in random order. The teacher model is responsible for providing the training data, which in this case means it must determine which instances (subset) of the mini-batch of data will be fed to the student. In order to reach the convergence faster, the reward was set to relate to the speed the student model learns. <br />
<br />
The authors also designed a state feature vector <math> g(s) </math> in order to efficiently represent the current states which include arrived training data and the student model. Within the State Features, there are three categories including Data features, student model features and the combination of both data and learner model. This state feature will be computed when each mini-batch of data arrives.<br />
<br />
Data features contain information for data instance, such as its label category, (for texts) the length of sentence, linguistic features for text segments (Tsvetkov et al., 2016), or (for images) gradients histogram features (Dalal & Triggs, 2005).<br />
<br />
Student model features include the signals reflecting how well current neural network is trained. The authors collect several simple features, such as passed mini-batch number (i.e., iteration), the average historical training loss and historical validation accuracy.<br />
<br />
Some additional features are collected to represent the combination of both data and learner model. By using these features, the authors aim to represent how important the arrived training data is for current leaner. The authors mainly use three parts of such signals in our classification tasks: 1) the predicted probabilities of each class; 2) the loss value on that data, which appears frequently in self-paced learning (Kumar et al., 2010; Jiang et al., 2014a; Sachan & Xing, 2016); 3) the margin value.<br />
<br />
The optimizer for training the teacher model is the maximum expected reward: <br />
<br />
\begin{align} <br />
J(θ) = E_{φ_θ(a|s)}[R(s,a)]<br />
\end{align}<br />
<br />
Which is non-differentiable w.r.t. <math> θ </math>, thus a likelihood ratio policy gradient algorithm is used to optimize <math> J(θ) </math> (Williams, 1992) [4]. The estimation is based on the gradient <math>\nabla_{\theta} = \sum_{t=1}^{T}E_{\phi_{\theta}}(a_t|s_t)[\nabla_{\theta}log(\phi_{\theta}(a_t|s_t))R(s_t, a_t)]</math>, which is empirically estimated as <math>\sum_{t=1}^{T} \nabla_{\theta}log(\phi_{\theta}(a_t|s_t))v_t</math>. <math>v_t</math> is defined as the sampled estimation of reward <math>R(s_t, a_t)</math> from one execution of the policy. Given that the reward is just the terminal reward, we have <math>\nabla_{\theta} = \sum_{t=1}^{T} \nabla_{\theta}log(\phi_{\theta}(a_t|s_t))r_T</math><br />
<br />
==Experiments==<br />
<br />
The L2T framework is tested on the following student models: multi-layer perceptron (MLP), ResNet (CNN), and Long-Short-Term-Memory network (RNN). <br />
<br />
The student tasks are Image classification for MNIST, for CIFAR-10, and sentiment classification for IMDB movie review dataset. <br />
<br />
The strategy will be benchmarked against the following teaching strategies:<br />
<br />
::'''NoTeach''': NoTeach removes the entire Teacher-Student paradigm and reverts back to the classical machine learning paradigm. In the context of data teaching, we consider the architecture fixed, and feed data in a pre-determined way. One would pre-define batch-size and cross-validation procedures as needed.<br />
::'''Self-Paced Learning (SPL)''': Teaching by ''hardness'' of data, defined as the loss. This strategy begins by filtering out data with larger loss value to train the student with "easy" data and gradually increases the hardness. Mathematically speaking, those training data <math>d </math> satisfying loss value <math>l(d) > \eta </math> will be filtered out, where the threshold <math> \eta </math> grows from smaller to larger during the training process. To improve the robustness of SPL, following the widely used trick in common SPL implementation (Jiang et al., 2014b), the authors filter training data using its loss rank in one mini-batch rather than the absolute loss value: they filter data instances with top <math>K </math>largest training loss values within a <math>M</math>-sized mini-batch, where <math>K</math> linearly drops from <math>M − 1 </math>to 0 during training.<br />
<br />
::'''L2T''': The Learning to Teach framework.<br />
::'''RandTeach''': Randomly filter data in each epoch according to the logged ratio of filtered data instances per epoch (as opposed to deliberate and dynamic filtering by L2T).<br />
<br />
For all teaching strategies, they make sure that the base neural network model will not be updated until <math>M </math> un-trained, yet selected data instances are accumulated. That is to guarantee that the convergence speed is only determined by the quality of taught data, not by different model updating frequencies. The model is implemented with Theano and run on one NVIDIA Tesla K40 GPU for each training/testing process.<br />
===Training a New Student===<br />
<br />
In the first set of experiments, the datasets or divided into two folds. The first folder is used to train the teacher; This is done by having the teacher train a student network on that half of the data, with a certain portion being used for computing rewards. After training, the teacher parameters are fixed, and used to train a new student network (with the same structure) on the second half of the dataset. When teaching a new student with the same model architecture, we observe that L2T achieves significantly faster convergence than other strategies across all tasks, especially compared to the NoTeach and RandTeach methods:<br />
<br />
[[File: L2T_speed.png | 1100px|center]]<br />
<br />
===Filtration Number===<br />
<br />
When investigating the details of filtered data instances per epoch, for the two image classification tasks, the L2T teacher filters an increasing amount of data as training goes on. The authors' intuition for the two image classification tasks is that the student model can learn from harder instances of data from the beginning, and thus the teacher can filter redundant data. In contrast, for training while for the natural language task, the student model must first learn from easy data instances.<br />
<br />
[[File: L2T_fig3.png | 1100px|center]]<br />
<br />
===Teaching New Student with Different Model Architecture===<br />
<br />
In this part, first a teacher model is trained by interacting with a student model. Then using the teacher model, another student model<br />
which has a different model architecture is taught.<br />
The results of Applying the teacher trained on ResNet32 to teach other architectures is shown below. The L2T algorithm can be seen to obtain higher accuracies earlier than the SPL, RandTeach, or NoTeach algorithms.<br />
<br />
[[File: L2T_fig4.png | 1100px|center]]<br />
<br />
===Training Time Analysis===<br />
<br />
The learning curves demonstrate the efficiency in accuracy achieved by the L2T over the other strategies. This is especially evident during the earlier training stages.<br />
<br />
[[File: L2T_fig5.png | 600px|center]]<br />
<br />
===Accuracy Improvement===<br />
<br />
When comparing training accuracy on the IMDB sentiment classification task, L2T improves on teaching policy over NoTeach and SPL.<br />
<br />
[[File: L2T_t1.png | 500px|center]]<br />
<br />
Table 1 shows that we boost the convergence speed, while the teacher model improves final accuracy. The student model is the LSTM network trained on IMDB. Prior to teaching the student model, we train the teacher model on half of the training data, and define the terminal reward as the set accuracy after the teacher model trains the student for 15 epochs. Then the teacher model is applied to train the student model on the full dataset till its convergence. The state features are kept the same as those in previous experiments. We can see that L2T achieves better classification accuracy for training LSTM network, surpassing the SPL baseline by more than 0.6 point (with p value < 0.001).<br />
<br />
=Future Work=<br />
<br />
There is some useful future work that can be extended from this work: <br />
<br />
1) Recent advances in multi-agent reinforcement learning could be tried on the Reinforcement Learning problem formulation of this paper. <br />
<br />
2) Some human in the loop architectures like CHAT and HAT (https://www.ijcai.org/proceedings/2017/0422.pdf) should give better results for the same framework. <br />
<br />
3) It would be interesting to try out the framework suggested in this paper (L2T) in Imperfect information and partially observable settings. <br />
<br />
4) As they have focused on data teaching exploring loss function teaching would be interesting.<br />
<br />
=Critique=<br />
<br />
While the conceptual framework of L2T is sound, the paper only experimentally demonstrates efficacy for ''data teaching'' which would seem to be the simplest to implement. The feasibility and effectiveness of teaching the loss function and hypothesis space are not explored in a real-world scenario. Also, this paper does not provide enough mathematical foundation to prove that this model can be generalized to other datasets and other general problems. The method presented here where the teacher model filters data does not seem to provide enough action space for the teacher model. Furthermore, the experimental results for data teaching suggest that the speed of convergence is the main improvement over other teaching strategies whereas the difference in accuracy less remarkable. The paper also assesses accuracy only by comparing L2T with NoTeach and SPL on the IMDB classification task, the improvement (or lack thereof) on the other classification tasks and teaching strategies is omitted. Again, this distinction is not possible to assess in loss function or hypothesis space teaching within the scope of this paper. They could have included larger datasets such as ImageNet and CIFAR100 in their experiments which would have provided some more insight.<br />
<br />
Also, teaching should not be limited to data, loss function and hypothesis space. In a human teacher-student model, the teaching contents are concepts and logical rules, similar to weights of hidden layers in neural networks. How to transfer such knowledge is interesting to investigate.<br />
<br />
The idea of having a generalizable teacher model to enhance student learning is admirable. In fact, the L2T framework is similar to the reinforcement learning actor-critic model, which is known to be effective. In general, one expects an effective teacher model would facilitate transfer learning and can significantly reduce student model training time. However, the T2L framework seems to fall short of that goal. Consider the CIFAR10 training scenario, the L2T model achieve 85% accuracy after 2 million training data, which is only about 3% more accuracy than a no-teacher model. Perhaps in the future, the L2T framework can improve and produce better performance.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Learning_to_Teach&diff=42320Learning to Teach2018-12-07T02:17:57Z<p>Ka2khan: /* Related Work */</p>
<hr />
<div><br />
<br />
=Introduction=<br />
<br />
This paper proposed the "learning to teach" (L2T) framework with two intelligent agents: a student model/agent, corresponding to the learner in traditional machine learning algorithms, and a teacher model/agent, determining the appropriate data, loss function, and hypothesis space to facilitate the learning of the student model.<br />
<br />
In modern human society, the role of teaching is heavily implicated in our education system; the goal is to equip students with the necessary knowledge and skills in an efficient manner. This is the fundamental ''student'' and ''teacher'' framework on which education stands. However, in the field of artificial intelligence (AI) and specifically machine learning, researchers have focused most of their efforts on the ''student'' (ie. designing various optimization algorithms to enhance the learning ability of intelligent agents). The paper argues that a formal study on the role of ‘teaching’ in AI is required. Analogous to teaching in human society, the teaching framework can: select training data that corresponds to the appropriate teaching materials (e.g. textbooks selected for the right difficulty), design loss functions that correspond to targeted examinations, and define the hypothesis space that corresponds to imparting the proper methodologies. Furthermore, an optimization framework (instead of heuristics) should be used to update the teaching skills based on the feedback from students, so as to achieve teacher-student co-evolution.<br />
<br />
Thus, the training phase of L2T would have several episodes of interactions between the teacher and the student model. Based on the state information in each step, the teacher model would update the teaching actions so that the student model could perform better on the Machine Learning problem. The student model would then provide reward signals back to the teacher model. These reward signals are used by the teacher model as part of the Reinforcement Learning process to update its parameters. In this paper policy gradient algorithm is incorporated. This process is end-to-end trainable and the authors are convinced that once converged, the teacher model could be applied to new learning scenarios and even new students, without extra efforts on re-training.<br />
<br />
To demonstrate the practical value of the proposed approach, the '''training data scheduling''' problem is chosen as an example. The authors show that by using the proposed method to adaptively select the most<br />
suitable training data, they can significantly improve the accuracy and convergence speed of various neural networks including multi-layer perceptron (MLP), convolutional neural networks (CNNs)<br />
and recurrent neural networks (RNNs), for different applications including image classification and text understanding.<br />
Further more , the teacher model obtained by the paper from one task can be smoothly transferred to other tasks. As an example, the teacher model trained on MNIST with the MLP learner, one can achieve a satisfactory performance on CIFAR-10 only using roughly half<br />
of the training data to train a ResNet model as the student.<br />
<br />
=Related Work=<br />
The L2T framework connects with two emerging trends in machine learning. The first is the movement from simple to advanced learning. This includes meta-learning (Schmidhuber, 1987; Thrun & Pratt, 2012) which explores automatic learning by transferring learned knowledge from meta tasks [1]. This approach has been applied to few-shot learning scenarios and in designing general optimizers and neural network architectures. (Hochreiter et al., 2001; Andrychowicz et al., 2016; Li & Malik, 2016; Zoph & Le, 2017)<br />
<br />
The second is the teaching, which can be classified into either machine-teaching (Zhu, 2015) [2] or hardness based methods. The former seeks to construct a minimal training set for the student to learn a target model (ie. an oracle). The latter assumes an order of data from easy instances to hard ones, hardness being determined in different ways, is beneficial to the learning process. In curriculum learning (CL) (Bengio et al, 2009; Spitkovsky et al. 2010; Tsvetkov et al, 2016) [3] measures hardness through heuristics based understanding of the data while self-paced learning (SPL) (Kumar et al., 2010; Lee & Grauman, 2011; Jiang et al., 2014; Supancic & Ramanan, 2013) [4] measures hardness by loss on data. Another teaching method called 'pedagogical teaching' with applications in inverse reinforcement learning is quite close, in setup, to the method being proposed in this paper. The similarity can be observed in the way the teacher adjusts its behaviour to facilitate student learning and in the way the teacher communicates with the student. <br />
<br />
The limitations of these works include the lack of formal definition of the teaching problem whereas a learning problem has been formal mathematical definition. This makes it difficult to differentiate between teaching and learning problems. Other limitations are the reliance on heuristics and fixed rules, which hinders generalization of the teaching task.<br />
<br />
=Learning to Teach=<br />
To introduce the problem and framework, without loss of generality, consider the setting of supervised learning.<br />
<br />
In supervised learning, each sample <math>x</math> is from a fixed but unknown distribution <math>P(x)</math>, and the corresponding label <math> y </math> is from a fixed but unknown distribution <math>P(y|x) </math>. The goal is to find a function <math>f_\omega(x)</math> with parameter vector <math>\omega</math> that minimizes the gap between the predicted label and the actual label.<br />
<br />
<br />
<br />
==Problem Definition==<br />
In supervised learning, the goal is to choose a function <math display="inline">f_w(x)</math> with <math display="inline">w</math> as the parameter vector to predict the supervisor's label as good as possible. The goodness of a function <math display="inline">f_w</math> is evaluated by the risk function: <br />
<br />
\begin{align*}R(w) = \int M(y, f_w(x))dP(x,y)\end{align*}<br />
<br />
where <math display="inline">M(,)</math> is the metric which evaluate the gap between the label and the prediction.<br />
<br />
The student model, denoted &mu;(), takes the set of training data <math> D </math>, the function class <math> Ω </math>, and loss function <math> L </math> as input to output a function, <math> f(ω) </math>, with parameter <math>ω^*</math> which minimizes risk <math>R(ω)</math> as in:<br />
<br />
\begin{align*}<br />
ω^* = arg min_{w \in \Omega} \sum_{x,y \in D} L(y, f_ω(x)) =: \mu (D, L, \Omega)<br />
\end{align*}<br />
<br />
The teaching model, denoted φ, tries to provide <math> D </math>, <math> L </math>, and <math> Ω </math> (or any combination, denoted <math> A </math>) to the student model such that the student model either achieves lower risk R(ω) or progresses as fast as possible.<br />
In contrast to traditional machine learning, which is only concerned with the student model in the<br />
learning to teach framework, the problem in the paper is also concerned with a teacher model, which tries to provide<br />
appropriate inputs to the student model so that it can achieve low risk functional as efficiently<br />
as possible.<br />
<br />
<br />
::'''Training Data''': Outputting a good training set <math> D </math>, analogous to human teachers providing students with proper learning materials such as textbooks.<br />
::'''Loss Function''': Designing a good loss function <math> L </math> , analogous to providing useful assessment criteria for students.<br />
::'''Hypothesis Space''': Defining a good function class <math> Ω </math> which the student model can select from. This is analogous to human teachers providing appropriate context, eg. middle school students taught math with basic algebra while undergraduate students are taught with calculus. Different Ω leads to different errors and optimization problem (Mohri et al., 2012).<br />
<br />
==Framework==<br />
The training phase consists of the teacher providing the student with the subset <math> A_{train} </math> of <math> A </math> and then taking feedback to improve its own parameters.After the convergence of the training process,<br />
the teacher model can be used to teach either<br />
new student models, or the same student<br />
models in new learning scenarios such as another<br />
subset <math> A_{test} </math>is provided.Such a generalization is feasible as long as the state representations<br />
S are the same across different student<br />
models and different scenarios. The L2T process is outlined in figure below:<br />
<br />
[[File: L2T_process.png | 500px|center]]<br />
<br />
* <math> s_t &isin; S </math> represents information available to the teacher model at time <math> t </math>. <math> s_t </math> is typically constructed from the current student model <math> f_{t−1} </math> and the past teaching history of the teacher model. <math> S </math> represents the set of states.<br />
* <math> a_t &isin; A </math> represents action taken the teacher model at time <math> t </math>, given state <math>s_t</math>. <math> A </math> represents the set of actions, where the action(s) can be any combination of teaching tasks involving the training data, loss function, and hypothesis space. <br />
* <math> φ_θ : S → A </math> is policy used by the teacher model to generate its action <math> φ_θ(s_t) = a_t </math><br />
* Student model takes <math> a_t </math> as input and outputs function <math> f_t </math>, by using the conventional ML techniques.<br />
<br />
Mathematically, taking data teaching as an example where <math>L</math> <math>/Omega</math> as fixed, the objective of teacher in the L2T framework is <br />
<br />
<center> <math>\max\limits_{\theta}{\sum\limits_{t}{r_t}} = \max\limits_{\theta}{\sum\limits_{t}{r(f_t)}} = \max\limits_{\theta}{\sum\limits_{t}{r(\mu(\phi_{\theta}(s_t), L, \Omega))}}</math> </center><br />
<br />
Once the training process converges, the teacher model may be utilized to teach a different subset of <math> A </math> or teach a different student model.<br />
<br />
=Application=<br />
<br />
There are different approaches to training the teacher model, this paper will apply reinforcement learning with <math> φ_θ </math> being the ''policy'' that interacts with <math> S </math>, the ''environment''. The paper applies data teaching to train a deep neural network student, <math> f </math>, for several classification tasks. Thus the student feedback measure will be classification accuracy. Its learning rule will be mini-batch stochastic gradient descent, where batches of data will arrive sequentially in random order. The teacher model is responsible for providing the training data, which in this case means it must determine which instances (subset) of the mini-batch of data will be fed to the student. In order to reach the convergence faster, the reward was set to relate to the speed the student model learns. <br />
<br />
The authors also designed a state feature vector <math> g(s) </math> in order to efficiently represent the current states which include arrived training data and the student model. Within the State Features, there are three categories including Data features, student model features and the combination of both data and learner model. This state feature will be computed when each mini-batch of data arrives.<br />
<br />
Data features contain information for data instance, such as its label category, (for texts) the length of sentence, linguistic features for text segments (Tsvetkov et al., 2016), or (for images) gradients histogram features (Dalal & Triggs, 2005).<br />
<br />
Student model features include the signals reflecting how well current neural network is trained. The authors collect several simple features, such as passed mini-batch number (i.e., iteration), the average historical training loss and historical validation accuracy.<br />
<br />
Some additional features are collected to represent the combination of both data and learner model. By using these features, the authors aim to represent how important the arrived training data is for current leaner. The authors mainly use three parts of such signals in our classification tasks: 1) the predicted probabilities of each class; 2) the loss value on that data, which appears frequently in self-paced learning (Kumar et al., 2010; Jiang et al., 2014a; Sachan & Xing, 2016); 3) the margin value.<br />
<br />
The optimizer for training the teacher model is the maximum expected reward: <br />
<br />
\begin{align} <br />
J(θ) = E_{φ_θ(a|s)}[R(s,a)]<br />
\end{align}<br />
<br />
Which is non-differentiable w.r.t. <math> θ </math>, thus a likelihood ratio policy gradient algorithm is used to optimize <math> J(θ) </math> (Williams, 1992) [4]. The estimation is based on the gradient <math>\nabla_{\theta} = \sum_{t=1}^{T}E_{\phi_{\theta}}(a_t|s_t)[\nabla_{\theta}log(\phi_{\theta}(a_t|s_t))R(s_t, a_t)]</math>, which is empirically estimated as <math>\sum_{t=1}^{T} \nabla_{\theta}log(\phi_{\theta}(a_t|s_t))v_t</math>. <math>v_t</math> is defined as the sampled estimation of reward <math>R(s_t, a_t)</math> from one execution of the policy. Given that the reward is just the terminal reward, we have <math>\nabla_{\theta} = \sum_{t=1}^{T} \nabla_{\theta}log(\phi_{\theta}(a_t|s_t))r_T</math><br />
<br />
==Experiments==<br />
<br />
The L2T framework is tested on the following student models: multi-layer perceptron (MLP), ResNet (CNN), and Long-Short-Term-Memory network (RNN). <br />
<br />
The student tasks are Image classification for MNIST, for CIFAR-10, and sentiment classification for IMDB movie review dataset. <br />
<br />
The strategy will be benchmarked against the following teaching strategies:<br />
<br />
::'''NoTeach''': NoTeach removes the entire Teacher-Student paradigm and reverts back to the classical machine learning paradigm. In the context of data teaching, we consider the architecture fixed, and feed data in a pre-determined way. One would pre-define batch-size and cross-validation procedures as needed.<br />
::'''Self-Paced Learning (SPL)''': Teaching by ''hardness'' of data, defined as the loss. This strategy begins by filtering out data with larger loss value to train the student with "easy" data and gradually increases the hardness. Mathematically speaking, those training data <math>d </math> satisfying loss value <math>l(d) > \eta </math> will be filtered out, where the threshold <math> \eta </math> grows from smaller to larger during the training process. To improve the robustness of SPL, following the widely used trick in common SPL implementation (Jiang et al., 2014b), the authors filter training data using its loss rank in one mini-batch rather than the absolute loss value: they filter data instances with top <math>K </math>largest training loss values within a <math>M</math>-sized mini-batch, where <math>K</math> linearly drops from <math>M − 1 </math>to 0 during training.<br />
<br />
::'''L2T''': The Learning to Teach framework.<br />
::'''RandTeach''': Randomly filter data in each epoch according to the logged ratio of filtered data instances per epoch (as opposed to deliberate and dynamic filtering by L2T).<br />
<br />
For all teaching strategies, they make sure that the base neural network model will not be updated until <math>M </math> un-trained, yet selected data instances are accumulated. That is to guarantee that the convergence speed is only determined by the quality of taught data, not by different model updating frequencies. The model is implemented with Theano and run on one NVIDIA Tesla K40 GPU for each training/testing process.<br />
===Training a New Student===<br />
<br />
In the first set of experiments, the datasets or divided into two folds. The first folder is used to train the teacher; This is done by having the teacher train a student network on that half of the data, with a certain portion being used for computing rewards. After training, the teacher parameters are fixed, and used to train a new student network (with the same structure) on the second half of the dataset. When teaching a new student with the same model architecture, we observe that L2T achieves significantly faster convergence than other strategies across all tasks, especially compared to the NoTeach and RandTeach methods:<br />
<br />
[[File: L2T_speed.png | 1100px|center]]<br />
<br />
===Filtration Number===<br />
<br />
When investigating the details of filtered data instances per epoch, for the two image classification tasks, the L2T teacher filters an increasing amount of data as training goes on. The authors' intuition for the two image classification tasks is that the student model can learn from harder instances of data from the beginning, and thus the teacher can filter redundant data. In contrast, for training while for the natural language task, the student model must first learn from easy data instances.<br />
<br />
[[File: L2T_fig3.png | 1100px|center]]<br />
<br />
===Teaching New Student with Different Model Architecture===<br />
<br />
In this part, first a teacher model is trained by interacting with a student model. Then using the teacher model, another student model<br />
which has a different model architecture is taught.<br />
The results of Applying the teacher trained on ResNet32 to teach other architectures is shown below. The L2T algorithm can be seen to obtain higher accuracies earlier than the SPL, RandTeach, or NoTeach algorithms.<br />
<br />
[[File: L2T_fig4.png | 1100px|center]]<br />
<br />
===Training Time Analysis===<br />
<br />
The learning curves demonstrate the efficiency in accuracy achieved by the L2T over the other strategies. This is especially evident during the earlier training stages.<br />
<br />
[[File: L2T_fig5.png | 600px|center]]<br />
<br />
===Accuracy Improvement===<br />
<br />
When comparing training accuracy on the IMDB sentiment classification task, L2T improves on teaching policy over NoTeach and SPL.<br />
<br />
[[File: L2T_t1.png | 500px|center]]<br />
<br />
Table 1 shows that we boost the convergence speed, while the teacher model improves final accuracy. The student model is the LSTM network trained on IMDB. Prior to teaching the student model, we train the teacher model on half of the training data, and define the terminal reward as the set accuracy after the teacher model trains the student for 15 epochs. Then the teacher model is applied to train the student model on the full dataset till its convergence. The state features are kept the same as those in previous experiments. We can see that L2T achieves better classification accuracy for training LSTM network, surpassing the SPL baseline by more than 0.6 point (with p value < 0.001).<br />
<br />
=Future Work=<br />
<br />
There is some useful future work that can be extended from this work: <br />
<br />
1) Recent advances in multi-agent reinforcement learning could be tried on the Reinforcement Learning problem formulation of this paper. <br />
<br />
2) Some human in the loop architectures like CHAT and HAT (https://www.ijcai.org/proceedings/2017/0422.pdf) should give better results for the same framework. <br />
<br />
3) It would be interesting to try out the framework suggested in this paper (L2T) in Imperfect information and partially observable settings. <br />
<br />
4) As they have focused on data teaching exploring loss function teaching would be interesting.<br />
<br />
=Critique=<br />
<br />
While the conceptual framework of L2T is sound, the paper only experimentally demonstrates efficacy for ''data teaching'' which would seem to be the simplest to implement. The feasibility and effectiveness of teaching the loss function and hypothesis space are not explored in a real-world scenario. Also, this paper does not provide enough mathematical foundation to prove that this model can be generalized to other datasets and other general problems. The method presented here where the teacher model filters data does not seem to provide enough action space for the teacher model. Furthermore, the experimental results for data teaching suggest that the speed of convergence is the main improvement over other teaching strategies whereas the difference in accuracy less remarkable. The paper also assesses accuracy only by comparing L2T with NoTeach and SPL on the IMDB classification task, the improvement (or lack thereof) on the other classification tasks and teaching strategies is omitted. Again, this distinction is not possible to assess in loss function or hypothesis space teaching within the scope of this paper. They could have included larger datasets such as ImageNet and CIFAR100 in their experiments which would have provided some more insight.<br />
<br />
Also, teaching should not be limited to data, loss function and hypothesis space. In a human teacher-student model, the teaching contents are concepts and logical rules, similar to weights of hidden layers in neural networks. How to transfer such knowledge is interesting to investigate.<br />
<br />
The idea of having a generalizable teacher model to enhance student learning is admirable. In fact, the L2T framework is similar to the reinforcement learning actor-critic model, which is known to be effective. In general, one expects an effective teacher model would facilitate transfer learning and can significantly reduce student model training time. However, the T2L framework seems to fall short of that goal. Consider the CIFAR10 training scenario, the L2T model achieve 85% accuracy after 2 million training data, which is only about 3% more accuracy than a no-teacher model. Perhaps in the future, the L2T framework can improve and produce better performance.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=stat946w18/Wavelet_Pooling_For_Convolutional_Neural_Networks&diff=42318stat946w18/Wavelet Pooling For Convolutional Neural Networks2018-12-07T01:43:16Z<p>Ka2khan: /* Background */</p>
<hr />
<div>=Wavelet Pooling For Convolutional Neural Networks=<br />
<br />
[https://goo.gl/forms/8NucSpF36K6IUZ0V2 Your feedback on presentations]<br />
<br />
<br />
== Introduction, Important Terms and Brief Summary==<br />
<br />
This paper focuses on the following important techniques: <br />
<br />
1) Convolutional Neural Nets (CNN): These are networks with layered structures that conform to the shape of inputs rather than vector-based features and consistently obtain high accuracies in the classification of images and objects. Researchers continue to focus on CNN to improve their performances. <br />
<br />
2) Pooling: Pooling subsamples the results of the convolution layers and gradually reduces spatial dimensions of the data throughout the network. It is done to reduce parameters, increase computational efficiency and regulate overfitting. <br />
<br />
Some of the pooling methods, including max pooling and average pooling, are deterministic. Deterministic pooling methods are efficient and simple, but can hinder the potential for optimal network learning. In contrast, mixed pooling and stochastic pooling use a probabilistic approach, which can address some problems of deterministic methods. The neighborhood approach is used in all the mentioned pooling methods due to its simplicity and efficiency. Nevertheless, the approach can cause edge halos, blurring, and aliasing which need to be minimized. This paper introduces wavelet pooling, which uses a second-level wavelet decomposition to subsample features. The nearest neighbor interpolation is replaced by an organic, subband method that more accurately represents the feature contents with fewer artifacts. The method decomposes features into a second level decomposition and discards first level subbands to reduce feature dimensions. This method is compared to other state-of-the-art pooling methods to demonstrate superior results. Tests are conducted on benchmark classification tests like MNIST, CIFAR10, SHVN and KDEF.<br />
<br />
For further information on wavelets, follow this link to MathWorks' [https://www.mathworks.com/videos/understanding-wavelets-part-1-what-are-wavelets-121279.html Understanding Wavelets] video series.<br />
<br />
== Intuition ==<br />
<br />
Convolutional networks commonly employ convolutional layers to extract features and use pooling methods for spatial dimensionality reduction. In this study, wavelet pooling is introduced as an alternative to traditional neighborhood pooling by providing a more structural feature dimension reduction method. Max pooling is addressed to have over-fitting problems and average pooling is mentioned to smooth out or 'dilute' details in features.<br />
<br />
Pooling is often introduced within networks to ensure local invariance to prevent overfitting due to small transitional shifts within an image. Despite the effectiveness of traditional pooling methods such as max pooling introduce this translational invariance by discarding information using methods analogous to nearest neighbour interpolation. With the hope of providing a more organic way of pooling, the authors leverage all information within cells inputted within a pooling operation with the hope that the resulting dim-reduced features are able to contain information from all high level cells using various dot products.<br />
<br />
== History ==<br />
<br />
A history of different pooling methods have been introduced and referenced in this study:<br />
* Manual subsampling at 1979<br />
* Max pooling at 1992<br />
* Mixed pooling at 2014<br />
* Pooling methods with probabilistic approaches at 2014 and 2015<br />
<br />
== Background ==<br />
Average Pooling and Max Pooling are well-known pooling methods and are popular techniques used in the literature. These pooling methods reduce input data dimensionality by taking the maximum value or the average value of specific areas and condense them into one single value. While these methods are simple and effective, they still have some limitations. The authors identify the following limitations:<br />
<br />
'''Limitations of Max Pooling and Average Pooling'''<br />
<br />
'''Max pooling''': takes the maximum value of a region <math>R_{ij} </math> and selects it to obtain a condensed feature map. It can '''erase the details''' of the image (happens if the main details have less intensity than the insignificant details) and also commonly '''over-fits''' the training data. The max-pooling is defined as:<br />
<br />
\begin{align}<br />
a_{kij} = max_{(p,q)\in R_{ij}}(a_{kpq})<br />
\end{align}<br />
<br />
'''Average pooling''': calculates the average value of a region and selects it to obtain a condensed feature map. Depending on the data, this method can '''dilute pertinent details''' from an image (happens for data with values much lower than the significant details) The avg-pooling is defined as:<br />
<br />
\begin{align}<br />
a_{kij} = \frac{1}{|R_{ij}|}\sum_{(p,q)\in R_{ij}}{{a_{kpq}}}<br />
\end{align}<br />
<br />
Where <math>a_{kij}</math> is the output activation of the <math>k^{th}</math> feature map at <math>(i,j)</math>, <math>a_{kpq}</math> is the input activation at<br />
<math>(p,q)</math> within <math>R_{ij}</math>, and <math>|R_{ij}|</math> is the size of the pooling region. Figure 1 gives a quick visual example of max and average pooling:<br />
<br />
[[File: pooling.png| 700px|center]]<br />
<br />
Figure 2 provides an example of the weaknesses of these two methods using toy images:<br />
<br />
[[File: fig0001.PNG| 700px|center]]<br />
<br />
<br />
'''How the researchers try to '''combat these issues'''?'''<br />
Using '''probabilistic pooling methods''' such as:<br />
<br />
1. '''Mixed pooling''': In general, when facing a new problem in which one would want to use a CNN, it is not intuitively known whether average or max-pooling should be preferred. Notably, both techniques have significant drawbacks. Average pooling forces the network to consider low magnitude (and possibly irrelevant information) in constructing representations, while max pooling can force the network to ignore fundamental differences between neighboring groups of pixels. To counteract this, mixed pooling probabilistically decides which to use during training / testing. It should be noted that, for training, it is only probabilistic in the forward pass. During back-propagation the network defaults to the earlier chosen method. Mixed pooling can be applied in 3 different ways.<br />
<br />
* For all features within a layer<br />
* Mixed between features within a layer<br />
* Mixed between regions for different features within a layer<br />
<br />
Mixed Pooling is defined as:<br />
<br />
\begin{align}<br />
a_{kij} = \lambda \cdot max_{(p,q)\in R_{ij}}(a_{kpq})+(1-\lambda) \cdot \frac{1}{|R_{ij}|}\sum_{(p,q)\in R_{ij}}{{a_{kpq}}}<br />
\end{align}<br />
<br />
Where <math>\lambda</math> is a random value 0 or 1, indicating max or average pooling for a particular region/feature/layer.<br />
<br />
2. '''Stochastic pooling''': improves upon max pooling by randomly sampling from neighborhood regions based on the probability values of each activation. This is defined as:<br />
<br />
\begin{align}<br />
a_{kij} = a_l ~ \text{where } ~ l\sim P(p_1,p_2,...,p_{|R_{ij}|})<br />
\end{align}<br />
<br />
with probability of activations within each region defined as follows:<br />
<br />
\begin{align}<br />
p_{pq} = \dfrac{a_{pq}}{\sum_{(p,q)} \in R_{ij} a_{pq}}<br />
\end{align}<br />
<br />
The figure below describes the process of Stochastic Pooling. The figure on the left shows the activations of a given region, and the corresponding probability is shown in the center. The activations with the highest probability is selected by the pooling method. However, any activation can be selected. In this case, the midrange activation of 13% is selected. <br />
<br />
[[File: stochastic pooling.jpeg| 700px|center]]<br />
<br />
As stochastic pooling is based on probability and is not deterministic, it avoids the shortcomings of max and average pooling and enjoys some of the advantages of max pooling.<br />
<br />
3. "Top-k activation pooling" is the method that picks the top-k activation in every pooling region. This makes sure that the maximum information can pass through subsampling gates. It is to be used with max pooling, but after max pooling, to further improve the representation capability, they pick top-k activation, sum them up, and constrain the summation by a constant. <br />
Details in this paper: https://www.hindawi.com/journals/wcmc/2018/8196906/<br />
<br />
'''Wavelets and Wavelet Transform'''<br />
A wavelet is a representation of some signal. For use in wavelet transforms, they are generally represented as combinations of basis signal functions.<br />
<br />
The wavelet transform involves taking the inner product of a signal (in this case, the image), with these basis functions. This produces a set of coefficients for the signal. These coefficients can then be quantized and coded in order to compress the image.<br />
<br />
One issue of note is that wavelets offer a tradeoff between resolution in frequency, or in time (or presumably, image location). For example, a sine wave will be useful to detect signals with its own frequency, but cannot detect where along the sine wave this alignment of signals is occuring. Thus, basis functions must be chosen with this tradeoff in mind.<br />
<br />
Source: Compressing still and moving images with wavelets<br />
<br />
The following images show the result of applying wavelet transform to an image for denoising:<br />
<br />
[[File: Noise Wavelet.jpg| 700px]] [[File: Denoised Wavelet.jpg| 700px]]<br />
<br />
images were taken from [https://en.wikipedia.org/wiki/Discrete_wavelet_transform#Example_in_Image_Processing here].<br />
<br />
== Proposed Method ==<br />
<br />
The previously highlighted pooling methods use neighborhoods to subsample, almost identical to nearest neighbor interpolation.<br />
<br />
The proposed pooling method uses wavelets (i.e. small waves - generally used in signal processing) to reduce the dimensions of the feature maps. They use wavelet transform to minimize artifacts resulting from neighborhood reduction. They postulate that their approach, which discards the first-order sub-bands, more organically captures the data compression. The authors say that this organic reduction therefore lessens the creation of jagged edges and other artifacts that may impede correct image classification.<br />
<br />
* '''Forward Propagation'''<br />
<br />
The proposed wavelet pooling scheme pools features by performing a 2nd order decomposition in the wavelet domain according to the fast wavelet transform (FWT) which is a more efficient implementation of the two-dimensional discrete wavelet transform (DWT) as follows:<br />
<br />
\begin{align}<br />
W_{\varphi}[j+1,k] = h_{\varphi}[-n]*W_{\varphi}[j,n]|_{n=2k,k\leq0}<br />
\end{align}<br />
<br />
\begin{align}<br />
W_{\psi}[j+1,k] = h_{\psi}[-n]*W_{\psi}[j,n]|_{n=2k,k\leq0}<br />
\end{align}<br />
<br />
where <math>\varphi</math> is the approximation function, and <math>\psi</math> is the detail function, <math>W_{\varphi},W_{\psi}</math> are called approximation and detail coefficients. <math>h_{\varphi[-n]}</math> and <math>h_{\psi[-n]}</math> are the time reversed scaling and wavelet vectors, (n) represents the sample in the vector, while (j) denotes the resolution level<br />
<br />
When using the FWT on images, it is applied twice (once on the rows, then again on the columns). By doing this in combination, the detail sub-bands (LH, HL, HH) at each decomposition level, and approximation sub-band (LL) for the highest decomposition level is obtained.<br />
After performing the 2nd order decomposition, the image features are reconstructed, but only using the 2nd order wavelet sub-bands. This method pools the image features by a factor of 2 using the inverse FWT (IFWT) which is based off the inverse DWT (IDWT).<br />
<br />
\begin{align}<br />
W_{\varphi}[j,k] = h_{\varphi}[-n]*W_{\varphi}[j+1,n]+h_{\psi}[-n]*W_{\psi}[j+1,n]|_{n=\frac{k}{2},k\leq0}<br />
\end{align}<br />
<br />
[[File: wavelet pooling forward.PNG| 700px|center]]<br />
<br />
<br />
* '''Backpropagation'''<br />
<br />
The proposed wavelet pooling algorithm performs backpropagation by reversing the process of its forward propagation. First, the image feature being backpropagated undergoes 1st order wavelet decomposition. After decomposition, the detail coefficient sub-bands up-sample by a factor of 2 to create a new 1st level decomposition. The initial decomposition then becomes the 2nd level decomposition. Finally, this new 2nd order wavelet decomposition reconstructs the image feature for further backpropagation using the IDWT. Figure 5, illustrates the wavelet pooling backpropagation algorithm in details:<br />
<br />
[[File:wavelet pooling backpropagation.PNG| 700px|center]]<br />
<br />
== Results and Discussion ==<br />
<br />
All experiments have been performed using the MatConvNet(Vedaldi & Lenc, 2015) architecture. Stochastic gradient descent has been used for training. For the proposed method, the Haar wavelet has been chosen as the basis wavelet for its property of having even, square sub-bands. All CNN structures except for MNIST use a network loosely based on Zeilers network (Zeiler & Fergus, 2013). The experiments are repeated with Dropout (Srivastava, 2013) and the Local Response Normalization (Krizhevsky, 2009) is replaced with Batch Normalization (Ioffe & Szegedy, 2015) for CIFAR-10 and SHVN (Dropout only) to examine how these regularization techniques change the pooling results. The authors have tested the proposed method on four different datasets as shown in the figure:<br />
<br />
[[File: selection of image datasets.PNG| 700px|center]]<br />
<br />
Different methods based on Max, Average, Mixed, Stochastic and Wavelet have been used at the pooling section of each architecture. Accuracy and Model Energy have been used as the metrics to evaluate the performance of the proposed methods. These have been evaluated and their performances have been compared on different data-sets.<br />
<br />
* MNIST:<br />
<br />
The network architecture is based on the example MNIST structure from MatConvNet, with batch-normalization, inserted. All other parameters are the same. The figure below shows their network structure for the MNIST experiments.<br />
<br />
[[File: CNN MNIST.PNG| 700px|center]]<br />
<br />
The input training data and test data come from the MNIST database of handwritten digits. The full training set of 60,000 images is used, as well as the full testing set of 10,000 images. The table below shows their proposed method outperforms all methods. Given the small number of epochs, max pooling is the only method to start to over-fit the data during training. Mixed and stochastic pooling show a rocky trajectory but do not over-fit. Average and wavelet pooling show a smoother descent in learning and error reduction. The figure below shows the energy of each method per epoch.<br />
<br />
[[File: MNIST pooling method energy.PNG| 700px|center]]<br />
<br />
<br />
The accuracies for both paradigms are shown below:<br />
<br />
<br />
[[File: MNIST perf.PNG| 700px|center]]<br />
<br />
* CIFAR-10:<br />
<br />
The authors perform two sets of experiments with the pooling methods. The first is a regular network structure with no dropout layers. They use this network to observe each pooling method without extra regularization. The second uses dropout and batch normalization and performs over 30 more epochs to observe the effects of these changes. <br />
<br />
[[File: CNN CIFAR.PNG| 700px|center]]<br />
<br />
The input training and test data come from the CIFAR-10 dataset. <br />
The full training set of 50,000 images is used, as well as the full testing set of 10,000 images. For both cases, with no dropout, and with dropout, Tables below show that the proposed method has the second highest accuracy.<br />
<br />
[[File: fig0000.jpg| 700px|center]]<br />
<br />
Max pooling over-fits fairly quickly, while wavelet pooling resists over-fitting. The change in learning rate prevents their method from over-fitting, and it continues to show a slower propensity for learning. Mixed and stochastic pooling maintain a consistent progression of learning, and their validation sets trend at a similar, but better rate than their proposed method. Average pooling shows the smoothest descent in learning and error reduction, especially in the validation set. The energy of each method per epoch is also shown below:<br />
<br />
[[File: CIFAR_pooling_method_energy.PNG| 700px|center]]<br />
<br />
<br />
* SHVN:<br />
<br />
Two sets of experiments are performed with the pooling methods. The first is a regular network structure with no dropout layers. They use this network to observe each pooling method without extra regularization same as what happened in the previous datasets.<br />
The second network uses dropout to observe the effects of this change. The figure below shows their network structure for the SHVN experiments:<br />
<br />
[[File: CNN SHVN.PNG| 700px|center]]<br />
<br />
The input training and test data come from the SHVN dataset. For the case with no dropout, they use 55,000 images from the training set. For the case with dropout, they use the full training set of 73,257 images, a validation set of 30,000 images they extract from the extra training set of 531,131 images, as well as the full testing set of 26,032 images. For both cases, with no dropout, and with dropout, Tables below show their proposed method has the second lowest accuracy.<br />
<br />
[[File: SHVN perf.PNG| 700px|center]]<br />
<br />
Max and wavelet pooling both slightly over-fit the data. Their method follows the path of max pooling but performs slightly better in maintaining some stability. Mixed, stochastic, and average pooling maintain a slow progression of learning, and their validation sets trend at near identical rates. The figure below shows the energy of each method per epoch.<br />
<br />
[[File: SHVN pooling method energy.PNG| 700px|center]]<br />
<br />
* KDEF:<br />
<br />
They run one set of experiments with the pooling methods that includes dropout. The figure below shows their network structure for the KDEF experiments:<br />
<br />
[[File:CNN KDEF.PNG| 700px|center]]<br />
<br />
The input training and test data come from the KDEF dataset. This dataset contains 4,900 images of 35 people displaying seven basic emotions (afraid, angry, disgusted, happy, neutral, sad, and surprised) using facial expressions. They display emotions at five poses (full left and right profiles, half left and right profiles, and straight).<br />
<br />
This dataset contains a few errors that they have fixed (missing or corrupted images, uncropped images, etc.). All of the missing images are at angles of -90, -45, 45, or 90 degrees. They fix the missing and corrupt images by mirroring their counterparts in MATLAB and adding them back to the dataset. They manually crop the images that need to match the dimensions set by the creators (762 x 562).<br />
KDEF does not designate a training or test data set. They shuffle the data and separate 3,900 images as training data, and 1,000 images as test data. They resize the images to 128x128 because of memory and time constraints.<br />
<br />
The dropout layers regulate the network and maintain stability in spite of some pooling methods known to over-fit. The table below shows their proposed method has the second highest accuracy. Max pooling eventually over-fits, while wavelet pooling resists over-fitting. Average and mixed pooling resist over-fitting but are unstable for most of the learning. Stochastic pooling maintains a consistent progression of learning. Wavelet pooling also follows a smoother, consistent progression of learning.<br />
The figure below shows the energy of each method per epoch.<br />
<br />
[[File: KDEF pooling method energy.PNG| 700px|center]]<br />
<br />
The accuracies for both paradigms are shown below:<br />
<br />
[[File: KDEF perf.PNG| 700px|center]]<br />
<br />
<br />
<br />
* Computational Complexity:<br />
Above experiments and implementations on wavelet pooling were more of a proof-of-concept rather than an optimized method. In terms of mathematical operations, the wavelet pooling method is the least computationally efficient compared to all other pooling methods mentioned above. Among all the methods, average pooling is the most efficient methods, max pooling and mix pooling are at a similar level while wavelet pooling is way more expensive to complete the calculation.<br />
<br />
== Conclusion ==<br />
<br />
They prove wavelet pooling has the potential to equal or eclipse some of the traditional methods currently utilized in CNNs. Their proposed method outperforms all others in the MNIST dataset, outperforms all but one in the CIFAR-10 and KDEF datasets, and performs within respectable ranges of the pooling methods that outdo it in the SHVN dataset. The addition of dropout and batch normalization show their proposed methods response to network regularization. Like the non-dropout cases, it outperforms all but one in both the CIFAR-10 & KDEF datasets and performs within respectable ranges of the pooling methods that outdo it in the SHVN dataset.<br />
<br />
The authors' results confirm previous studies proving that no one pooling method is superior, but some perform better than others depending on the dataset and network structure Boureau et al. (2010); Lee et al. (2016). Furthermore, many networks alternate between different pooling methods to maximize the effectiveness of each method. [1]<br />
<br />
Future work and improvements in this area could be to vary the wavelet basis to explore which basis performs best for the pooling. Altering the upsampling and downsampling factors in the decomposition and reconstruction can lead to better image feature reductions outside of the 2x2 scale. Retention of the subbands we discard for the backpropagation could lead to higher accuracies and fewer errors. Improving the method of FTW we use could greatly increase computational efficiency. Finally, analyzing the structural similarity (SSIM) of wavelet pooling versus other methods could further prove the vitality of using the authors' approach. [1]<br />
<br />
== Suggested Future work ==<br />
<br />
Upsampling and downsampling factors in decomposition and reconstruction needs to be changed to achieve more feature reduction.<br />
The subbands that we previously discard should be kept for higher accuracies. To achieve higher computational efficiency, improving the FTW method is needed.<br />
<br />
== Critiques and Suggestions ==<br />
*The functionality of backpropagation process which can be a positive point of the study is not described enough comparing to the existing methods.<br />
* The main study is on wavelet decomposition while the reason of using Haar as mother wavelet and the number of decomposition levels selection has not been described and are just mentioned as a future study! <br />
* At the beginning, the study mentions that the pooling method is not under attention as it should be. In the end, results show that choosing the pooling method depends on the dataset and they mention trial and test as a reasonable approach to choose the pooling method. In my point of view, the authors have not really been focused on providing a pooling method which can help the current conditions to be improved effectively. At least, trying to extract a better pattern for relating results to the dataset structure could be so helpful.<br />
* Average pooling origins which are mentioned as the main pooling algorithm to compare with, is not even referenced in the introduction.<br />
* Combination of the wavelet, Max and Average pooling can be an interesting option to investigate more on this topic; both in a row(Max/Avg after wavelet pooling) and combined like mix pooling option.<br />
* While the current datasets express the performance of the proposed method in an appropriate way, it could be a good idea to evaluate the method using some larger datasets. Maybe it helps to understand whether the size of a dataset can affect the overfitting behavior of max pooling which is mentioned in the paper.<br />
* Adding asymptotic notations to the computational complexity of the proposed algorithm would be meaningful, particularly since the given results are for a single/fixed input size (one image in forward propagation) and consequently are not generalizable. <br />
* They could have considered comparing against Fast Fourier Transform (FFT). Including a non wavelet form seems to be an obvious candidate for comparison<br />
* If they went beyond the 2x2 pooling window this would have further supported their method<br />
* ([[https://openreview.net/forum?id=rkhlb8lCZ]]) The experiments are largely conducted with very small scale datasets. As a result, I am not sure if they are representative enough to show the performance difference between different pooling methods.<br />
* ([[https://openreview.net/forum?id=rkhlb8lCZ]]) No comparison to non-wavelet methods. For example, one obvious comparison would have been to look at using a DCT or FFT transform where the output would discard high-frequency components (this can get very close to the wavelet idea!). Also, this critique might provides us with some interesting research directions since DCT or FFT transforms as pooling are not throughly studied yet.<br />
* Also, convolutional neural network are not only used in image related tasks. Evaluating the efficiency of wavelet pooling in convolutional neural network applied to natural languages or other applicable areas will be interesting. Such experiments shall also show if such approach can be generalized. <br />
<br />
== References ==<br />
<br />
Williams, Travis, and Robert Li. "Wavelet Pooling for Convolutional Neural Networks." (2018).<br />
<br />
Hilton, Michael L., Björn D. Jawerth, and Ayan Sengupta. "Compressing still and moving images with wavelets." Multimedia systems 2.5 (1994): 218-227.<br />
<br />
<br />
== Revisions == <br />
<br />
*Two reviewers really liked the paper and one of them called it in the top 15% papers in the conference which supports the novelty and potential of the idea. One other reviewer, however, believed that this was not good enough to be accepted and the main reason for rejection was the linearity nature of wavelet(which was not convincingly described). <br />
<br />
*The main concern of two of the reviewers has been the size of the datasets that have been used to test the method and the authors have mentioned future works concerning bigger datasets to test the method.<br />
<br />
*The computational cost section had not been included in the paper initially and was added after one of the reviewer's concern. So, the other reviewers have not been curious about this and unfortunately, there is no comment on that from them. However, the description on the non-efficient implementation seemed to be satisfactory to the reviewer which resulted in being accepted. <br />
<br />
[https://openreview.net/forum?id=rkhlb8lCZ Revisions]<br />
<br />
At the end, if you are interested in implementing the method, they are willing to share their code but after making it efficient. So, maybe there will be another paper regarding less computational cost on larger datasets with a publishable code.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=File:Denoised_Wavelet.jpg&diff=42317File:Denoised Wavelet.jpg2018-12-07T01:41:01Z<p>Ka2khan: </p>
<hr />
<div></div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=File:Noise_Wavelet.jpg&diff=42316File:Noise Wavelet.jpg2018-12-07T01:40:20Z<p>Ka2khan: </p>
<hr />
<div></div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=stat946w18/Wavelet_Pooling_For_Convolutional_Neural_Networks&diff=42315stat946w18/Wavelet Pooling For Convolutional Neural Networks2018-12-07T01:24:41Z<p>Ka2khan: /* Background */</p>
<hr />
<div>=Wavelet Pooling For Convolutional Neural Networks=<br />
<br />
[https://goo.gl/forms/8NucSpF36K6IUZ0V2 Your feedback on presentations]<br />
<br />
<br />
== Introduction, Important Terms and Brief Summary==<br />
<br />
This paper focuses on the following important techniques: <br />
<br />
1) Convolutional Neural Nets (CNN): These are networks with layered structures that conform to the shape of inputs rather than vector-based features and consistently obtain high accuracies in the classification of images and objects. Researchers continue to focus on CNN to improve their performances. <br />
<br />
2) Pooling: Pooling subsamples the results of the convolution layers and gradually reduces spatial dimensions of the data throughout the network. It is done to reduce parameters, increase computational efficiency and regulate overfitting. <br />
<br />
Some of the pooling methods, including max pooling and average pooling, are deterministic. Deterministic pooling methods are efficient and simple, but can hinder the potential for optimal network learning. In contrast, mixed pooling and stochastic pooling use a probabilistic approach, which can address some problems of deterministic methods. The neighborhood approach is used in all the mentioned pooling methods due to its simplicity and efficiency. Nevertheless, the approach can cause edge halos, blurring, and aliasing which need to be minimized. This paper introduces wavelet pooling, which uses a second-level wavelet decomposition to subsample features. The nearest neighbor interpolation is replaced by an organic, subband method that more accurately represents the feature contents with fewer artifacts. The method decomposes features into a second level decomposition and discards first level subbands to reduce feature dimensions. This method is compared to other state-of-the-art pooling methods to demonstrate superior results. Tests are conducted on benchmark classification tests like MNIST, CIFAR10, SHVN and KDEF.<br />
<br />
For further information on wavelets, follow this link to MathWorks' [https://www.mathworks.com/videos/understanding-wavelets-part-1-what-are-wavelets-121279.html Understanding Wavelets] video series.<br />
<br />
== Intuition ==<br />
<br />
Convolutional networks commonly employ convolutional layers to extract features and use pooling methods for spatial dimensionality reduction. In this study, wavelet pooling is introduced as an alternative to traditional neighborhood pooling by providing a more structural feature dimension reduction method. Max pooling is addressed to have over-fitting problems and average pooling is mentioned to smooth out or 'dilute' details in features.<br />
<br />
Pooling is often introduced within networks to ensure local invariance to prevent overfitting due to small transitional shifts within an image. Despite the effectiveness of traditional pooling methods such as max pooling introduce this translational invariance by discarding information using methods analogous to nearest neighbour interpolation. With the hope of providing a more organic way of pooling, the authors leverage all information within cells inputted within a pooling operation with the hope that the resulting dim-reduced features are able to contain information from all high level cells using various dot products.<br />
<br />
== History ==<br />
<br />
A history of different pooling methods have been introduced and referenced in this study:<br />
* Manual subsampling at 1979<br />
* Max pooling at 1992<br />
* Mixed pooling at 2014<br />
* Pooling methods with probabilistic approaches at 2014 and 2015<br />
<br />
== Background ==<br />
Average Pooling and Max Pooling are well-known pooling methods and are popular techniques used in the literature. These pooling methods reduce input data dimensionality by taking the maximum value or the average value of specific areas and condense them into one single value. While these methods are simple and effective, they still have some limitations. The authors identify the following limitations:<br />
<br />
'''Limitations of Max Pooling and Average Pooling'''<br />
<br />
'''Max pooling''': takes the maximum value of a region <math>R_{ij} </math> and selects it to obtain a condensed feature map. It can '''erase the details''' of the image (happens if the main details have less intensity than the insignificant details) and also commonly '''over-fits''' the training data. The max-pooling is defined as:<br />
<br />
\begin{align}<br />
a_{kij} = max_{(p,q)\in R_{ij}}(a_{kpq})<br />
\end{align}<br />
<br />
'''Average pooling''': calculates the average value of a region and selects it to obtain a condensed feature map. Depending on the data, this method can '''dilute pertinent details''' from an image (happens for data with values much lower than the significant details) The avg-pooling is defined as:<br />
<br />
\begin{align}<br />
a_{kij} = \frac{1}{|R_{ij}|}\sum_{(p,q)\in R_{ij}}{{a_{kpq}}}<br />
\end{align}<br />
<br />
Where <math>a_{kij}</math> is the output activation of the <math>k^{th}</math> feature map at <math>(i,j)</math>, <math>a_{kpq}</math> is the input activation at<br />
<math>(p,q)</math> within <math>R_{ij}</math>, and <math>|R_{ij}|</math> is the size of the pooling region. Figure 1 gives a quick visual example of max and average pooling:<br />
<br />
[[File: pooling.png| 700px|center]]<br />
<br />
Figure 2 provides an example of the weaknesses of these two methods using toy images:<br />
<br />
[[File: fig0001.PNG| 700px|center]]<br />
<br />
<br />
'''How the researchers try to '''combat these issues'''?'''<br />
Using '''probabilistic pooling methods''' such as:<br />
<br />
1. '''Mixed pooling''': In general, when facing a new problem in which one would want to use a CNN, it is unintuitive to whether average or max-pooling is preferred. Notably, both techniques have significant drawbacks. Average pooling forces the network to consider low magnitude (and possibly irrelevant information) in constructing representations, while max pooling can force the network to ignore fundamental differences between neighbouring groups of pixels. To counteract this, mixed pooling probabilistically decides which to use during training / testing. It should be noted that, for training, it is only probabilistic in the forward pass. During back-propagation the network defaults to the earlier chosen method. Mixed pooling can be applied in 3 different ways.<br />
<br />
* For all features within a layer<br />
* Mixed between features within a layer<br />
* Mixed between regions for different features within a layer<br />
<br />
Mixed Pooling is defined as:<br />
<br />
\begin{align}<br />
a_{kij} = \lambda \cdot max_{(p,q)\in R_{ij}}(a_{kpq})+(1-\lambda) \cdot \frac{1}{|R_{ij}|}\sum_{(p,q)\in R_{ij}}{{a_{kpq}}}<br />
\end{align}<br />
<br />
Where <math>\lambda</math> is a random value 0 or 1, indicating max or average pooling.<br />
<br />
2. '''Stochastic pooling''': improves upon max pooling by randomly sampling from neighborhood regions based on the probability values of each activation. This is defined as:<br />
<br />
\begin{align}<br />
a_{kij} = a_l ~ \text{where } ~ l\sim P(p_1,p_2,...,p_{|R_{ij}|})<br />
\end{align}<br />
<br />
with probability of activations within each region defined as follows:<br />
<br />
\begin{align}<br />
p_{pq} = \dfrac{a_{pq}}{\sum_{(p,q)} \in R_{ij} a_{pq}}<br />
\end{align}<br />
<br />
The figure below describes the process of Stochastic Pooling. The figure on the left shows the activations of a given region, and the corresponding probability is shown in the center. The activations with the highest probability is selected by the pooling method. However, any activation can be selected. In this case, the midrange activation of 13% is selected. <br />
<br />
[[File: stochastic pooling.jpeg| 700px|center]]<br />
<br />
As stochastic pooling is based on probability and is not deterministic, it avoids the shortcomings of max and average pooling and enjoys some of the advantages of max pooling.<br />
<br />
3. "Top-k activation pooling" is the method that picks the top-k activation in every pooling region. This makes sure that the maximum information can pass through subsampling gates. It is to be used with max pooling, but after max pooling, to further improve the representation capability, they pick top-k activation, sum them up, and constrain the summation by a constant. <br />
Details in this paper: https://www.hindawi.com/journals/wcmc/2018/8196906/<br />
<br />
'''Wavelets and Wavelet Transform'''<br />
A wavelet is a representation of some signal. For use in wavelet transforms, they are generally represented as combinations of basis signal functions.<br />
<br />
The wavelet transform involves taking the inner product of a signal (in this case, the image), with these basis functions. This produces a set of coefficients for the signal. These coefficients can then be quantized and coded in order to compress the image.<br />
<br />
One issue of note is that wavelets offer a tradeoff between resolution in frequency, or in time (or presumably, image location). For example, a sine wave will be useful to detect signals with its own frequency, but cannot detect where along the sine wave this alignment of signals is occuring. Thus, basis functions must be chosen with this tradeoff in mind.<br />
<br />
Source: Compressing still and moving images with wavelets<br />
<br />
== Proposed Method ==<br />
<br />
The previously highlighted pooling methods use neighborhoods to subsample, almost identical to nearest neighbor interpolation.<br />
<br />
The proposed pooling method uses wavelets (i.e. small waves - generally used in signal processing) to reduce the dimensions of the feature maps. They use wavelet transform to minimize artifacts resulting from neighborhood reduction. They postulate that their approach, which discards the first-order sub-bands, more organically captures the data compression. The authors say that this organic reduction therefore lessens the creation of jagged edges and other artifacts that may impede correct image classification.<br />
<br />
* '''Forward Propagation'''<br />
<br />
The proposed wavelet pooling scheme pools features by performing a 2nd order decomposition in the wavelet domain according to the fast wavelet transform (FWT) which is a more efficient implementation of the two-dimensional discrete wavelet transform (DWT) as follows:<br />
<br />
\begin{align}<br />
W_{\varphi}[j+1,k] = h_{\varphi}[-n]*W_{\varphi}[j,n]|_{n=2k,k\leq0}<br />
\end{align}<br />
<br />
\begin{align}<br />
W_{\psi}[j+1,k] = h_{\psi}[-n]*W_{\psi}[j,n]|_{n=2k,k\leq0}<br />
\end{align}<br />
<br />
where <math>\varphi</math> is the approximation function, and <math>\psi</math> is the detail function, <math>W_{\varphi},W_{\psi}</math> are called approximation and detail coefficients. <math>h_{\varphi[-n]}</math> and <math>h_{\psi[-n]}</math> are the time reversed scaling and wavelet vectors, (n) represents the sample in the vector, while (j) denotes the resolution level<br />
<br />
When using the FWT on images, it is applied twice (once on the rows, then again on the columns). By doing this in combination, the detail sub-bands (LH, HL, HH) at each decomposition level, and approximation sub-band (LL) for the highest decomposition level is obtained.<br />
After performing the 2nd order decomposition, the image features are reconstructed, but only using the 2nd order wavelet sub-bands. This method pools the image features by a factor of 2 using the inverse FWT (IFWT) which is based off the inverse DWT (IDWT).<br />
<br />
\begin{align}<br />
W_{\varphi}[j,k] = h_{\varphi}[-n]*W_{\varphi}[j+1,n]+h_{\psi}[-n]*W_{\psi}[j+1,n]|_{n=\frac{k}{2},k\leq0}<br />
\end{align}<br />
<br />
[[File: wavelet pooling forward.PNG| 700px|center]]<br />
<br />
<br />
* '''Backpropagation'''<br />
<br />
The proposed wavelet pooling algorithm performs backpropagation by reversing the process of its forward propagation. First, the image feature being backpropagated undergoes 1st order wavelet decomposition. After decomposition, the detail coefficient sub-bands up-sample by a factor of 2 to create a new 1st level decomposition. The initial decomposition then becomes the 2nd level decomposition. Finally, this new 2nd order wavelet decomposition reconstructs the image feature for further backpropagation using the IDWT. Figure 5, illustrates the wavelet pooling backpropagation algorithm in details:<br />
<br />
[[File:wavelet pooling backpropagation.PNG| 700px|center]]<br />
<br />
== Results and Discussion ==<br />
<br />
All experiments have been performed using the MatConvNet(Vedaldi & Lenc, 2015) architecture. Stochastic gradient descent has been used for training. For the proposed method, the Haar wavelet has been chosen as the basis wavelet for its property of having even, square sub-bands. All CNN structures except for MNIST use a network loosely based on Zeilers network (Zeiler & Fergus, 2013). The experiments are repeated with Dropout (Srivastava, 2013) and the Local Response Normalization (Krizhevsky, 2009) is replaced with Batch Normalization (Ioffe & Szegedy, 2015) for CIFAR-10 and SHVN (Dropout only) to examine how these regularization techniques change the pooling results. The authors have tested the proposed method on four different datasets as shown in the figure:<br />
<br />
[[File: selection of image datasets.PNG| 700px|center]]<br />
<br />
Different methods based on Max, Average, Mixed, Stochastic and Wavelet have been used at the pooling section of each architecture. Accuracy and Model Energy have been used as the metrics to evaluate the performance of the proposed methods. These have been evaluated and their performances have been compared on different data-sets.<br />
<br />
* MNIST:<br />
<br />
The network architecture is based on the example MNIST structure from MatConvNet, with batch-normalization, inserted. All other parameters are the same. The figure below shows their network structure for the MNIST experiments.<br />
<br />
[[File: CNN MNIST.PNG| 700px|center]]<br />
<br />
The input training data and test data come from the MNIST database of handwritten digits. The full training set of 60,000 images is used, as well as the full testing set of 10,000 images. The table below shows their proposed method outperforms all methods. Given the small number of epochs, max pooling is the only method to start to over-fit the data during training. Mixed and stochastic pooling show a rocky trajectory but do not over-fit. Average and wavelet pooling show a smoother descent in learning and error reduction. The figure below shows the energy of each method per epoch.<br />
<br />
[[File: MNIST pooling method energy.PNG| 700px|center]]<br />
<br />
<br />
The accuracies for both paradigms are shown below:<br />
<br />
<br />
[[File: MNIST perf.PNG| 700px|center]]<br />
<br />
* CIFAR-10:<br />
<br />
The authors perform two sets of experiments with the pooling methods. The first is a regular network structure with no dropout layers. They use this network to observe each pooling method without extra regularization. The second uses dropout and batch normalization and performs over 30 more epochs to observe the effects of these changes. <br />
<br />
[[File: CNN CIFAR.PNG| 700px|center]]<br />
<br />
The input training and test data come from the CIFAR-10 dataset. <br />
The full training set of 50,000 images is used, as well as the full testing set of 10,000 images. For both cases, with no dropout, and with dropout, Tables below show that the proposed method has the second highest accuracy.<br />
<br />
[[File: fig0000.jpg| 700px|center]]<br />
<br />
Max pooling over-fits fairly quickly, while wavelet pooling resists over-fitting. The change in learning rate prevents their method from over-fitting, and it continues to show a slower propensity for learning. Mixed and stochastic pooling maintain a consistent progression of learning, and their validation sets trend at a similar, but better rate than their proposed method. Average pooling shows the smoothest descent in learning and error reduction, especially in the validation set. The energy of each method per epoch is also shown below:<br />
<br />
[[File: CIFAR_pooling_method_energy.PNG| 700px|center]]<br />
<br />
<br />
* SHVN:<br />
<br />
Two sets of experiments are performed with the pooling methods. The first is a regular network structure with no dropout layers. They use this network to observe each pooling method without extra regularization same as what happened in the previous datasets.<br />
The second network uses dropout to observe the effects of this change. The figure below shows their network structure for the SHVN experiments:<br />
<br />
[[File: CNN SHVN.PNG| 700px|center]]<br />
<br />
The input training and test data come from the SHVN dataset. For the case with no dropout, they use 55,000 images from the training set. For the case with dropout, they use the full training set of 73,257 images, a validation set of 30,000 images they extract from the extra training set of 531,131 images, as well as the full testing set of 26,032 images. For both cases, with no dropout, and with dropout, Tables below show their proposed method has the second lowest accuracy.<br />
<br />
[[File: SHVN perf.PNG| 700px|center]]<br />
<br />
Max and wavelet pooling both slightly over-fit the data. Their method follows the path of max pooling but performs slightly better in maintaining some stability. Mixed, stochastic, and average pooling maintain a slow progression of learning, and their validation sets trend at near identical rates. The figure below shows the energy of each method per epoch.<br />
<br />
[[File: SHVN pooling method energy.PNG| 700px|center]]<br />
<br />
* KDEF:<br />
<br />
They run one set of experiments with the pooling methods that includes dropout. The figure below shows their network structure for the KDEF experiments:<br />
<br />
[[File:CNN KDEF.PNG| 700px|center]]<br />
<br />
The input training and test data come from the KDEF dataset. This dataset contains 4,900 images of 35 people displaying seven basic emotions (afraid, angry, disgusted, happy, neutral, sad, and surprised) using facial expressions. They display emotions at five poses (full left and right profiles, half left and right profiles, and straight).<br />
<br />
This dataset contains a few errors that they have fixed (missing or corrupted images, uncropped images, etc.). All of the missing images are at angles of -90, -45, 45, or 90 degrees. They fix the missing and corrupt images by mirroring their counterparts in MATLAB and adding them back to the dataset. They manually crop the images that need to match the dimensions set by the creators (762 x 562).<br />
KDEF does not designate a training or test data set. They shuffle the data and separate 3,900 images as training data, and 1,000 images as test data. They resize the images to 128x128 because of memory and time constraints.<br />
<br />
The dropout layers regulate the network and maintain stability in spite of some pooling methods known to over-fit. The table below shows their proposed method has the second highest accuracy. Max pooling eventually over-fits, while wavelet pooling resists over-fitting. Average and mixed pooling resist over-fitting but are unstable for most of the learning. Stochastic pooling maintains a consistent progression of learning. Wavelet pooling also follows a smoother, consistent progression of learning.<br />
The figure below shows the energy of each method per epoch.<br />
<br />
[[File: KDEF pooling method energy.PNG| 700px|center]]<br />
<br />
The accuracies for both paradigms are shown below:<br />
<br />
[[File: KDEF perf.PNG| 700px|center]]<br />
<br />
<br />
<br />
* Computational Complexity:<br />
Above experiments and implementations on wavelet pooling were more of a proof-of-concept rather than an optimized method. In terms of mathematical operations, the wavelet pooling method is the least computationally efficient compared to all other pooling methods mentioned above. Among all the methods, average pooling is the most efficient methods, max pooling and mix pooling are at a similar level while wavelet pooling is way more expensive to complete the calculation.<br />
<br />
== Conclusion ==<br />
<br />
They prove wavelet pooling has the potential to equal or eclipse some of the traditional methods currently utilized in CNNs. Their proposed method outperforms all others in the MNIST dataset, outperforms all but one in the CIFAR-10 and KDEF datasets, and performs within respectable ranges of the pooling methods that outdo it in the SHVN dataset. The addition of dropout and batch normalization show their proposed methods response to network regularization. Like the non-dropout cases, it outperforms all but one in both the CIFAR-10 & KDEF datasets and performs within respectable ranges of the pooling methods that outdo it in the SHVN dataset.<br />
<br />
The authors' results confirm previous studies proving that no one pooling method is superior, but some perform better than others depending on the dataset and network structure Boureau et al. (2010); Lee et al. (2016). Furthermore, many networks alternate between different pooling methods to maximize the effectiveness of each method. [1]<br />
<br />
Future work and improvements in this area could be to vary the wavelet basis to explore which basis performs best for the pooling. Altering the upsampling and downsampling factors in the decomposition and reconstruction can lead to better image feature reductions outside of the 2x2 scale. Retention of the subbands we discard for the backpropagation could lead to higher accuracies and fewer errors. Improving the method of FTW we use could greatly increase computational efficiency. Finally, analyzing the structural similarity (SSIM) of wavelet pooling versus other methods could further prove the vitality of using the authors' approach. [1]<br />
<br />
== Suggested Future work ==<br />
<br />
Upsampling and downsampling factors in decomposition and reconstruction needs to be changed to achieve more feature reduction.<br />
The subbands that we previously discard should be kept for higher accuracies. To achieve higher computational efficiency, improving the FTW method is needed.<br />
<br />
== Critiques and Suggestions ==<br />
*The functionality of backpropagation process which can be a positive point of the study is not described enough comparing to the existing methods.<br />
* The main study is on wavelet decomposition while the reason of using Haar as mother wavelet and the number of decomposition levels selection has not been described and are just mentioned as a future study! <br />
* At the beginning, the study mentions that the pooling method is not under attention as it should be. In the end, results show that choosing the pooling method depends on the dataset and they mention trial and test as a reasonable approach to choose the pooling method. In my point of view, the authors have not really been focused on providing a pooling method which can help the current conditions to be improved effectively. At least, trying to extract a better pattern for relating results to the dataset structure could be so helpful.<br />
* Average pooling origins which are mentioned as the main pooling algorithm to compare with, is not even referenced in the introduction.<br />
* Combination of the wavelet, Max and Average pooling can be an interesting option to investigate more on this topic; both in a row(Max/Avg after wavelet pooling) and combined like mix pooling option.<br />
* While the current datasets express the performance of the proposed method in an appropriate way, it could be a good idea to evaluate the method using some larger datasets. Maybe it helps to understand whether the size of a dataset can affect the overfitting behavior of max pooling which is mentioned in the paper.<br />
* Adding asymptotic notations to the computational complexity of the proposed algorithm would be meaningful, particularly since the given results are for a single/fixed input size (one image in forward propagation) and consequently are not generalizable. <br />
* They could have considered comparing against Fast Fourier Transform (FFT). Including a non wavelet form seems to be an obvious candidate for comparison<br />
* If they went beyond the 2x2 pooling window this would have further supported their method<br />
* ([[https://openreview.net/forum?id=rkhlb8lCZ]]) The experiments are largely conducted with very small scale datasets. As a result, I am not sure if they are representative enough to show the performance difference between different pooling methods.<br />
* ([[https://openreview.net/forum?id=rkhlb8lCZ]]) No comparison to non-wavelet methods. For example, one obvious comparison would have been to look at using a DCT or FFT transform where the output would discard high-frequency components (this can get very close to the wavelet idea!). Also, this critique might provides us with some interesting research directions since DCT or FFT transforms as pooling are not throughly studied yet.<br />
* Also, convolutional neural network are not only used in image related tasks. Evaluating the efficiency of wavelet pooling in convolutional neural network applied to natural languages or other applicable areas will be interesting. Such experiments shall also show if such approach can be generalized. <br />
<br />
== References ==<br />
<br />
Williams, Travis, and Robert Li. "Wavelet Pooling for Convolutional Neural Networks." (2018).<br />
<br />
Hilton, Michael L., Björn D. Jawerth, and Ayan Sengupta. "Compressing still and moving images with wavelets." Multimedia systems 2.5 (1994): 218-227.<br />
<br />
<br />
== Revisions == <br />
<br />
*Two reviewers really liked the paper and one of them called it in the top 15% papers in the conference which supports the novelty and potential of the idea. One other reviewer, however, believed that this was not good enough to be accepted and the main reason for rejection was the linearity nature of wavelet(which was not convincingly described). <br />
<br />
*The main concern of two of the reviewers has been the size of the datasets that have been used to test the method and the authors have mentioned future works concerning bigger datasets to test the method.<br />
<br />
*The computational cost section had not been included in the paper initially and was added after one of the reviewer's concern. So, the other reviewers have not been curious about this and unfortunately, there is no comment on that from them. However, the description on the non-efficient implementation seemed to be satisfactory to the reviewer which resulted in being accepted. <br />
<br />
[https://openreview.net/forum?id=rkhlb8lCZ Revisions]<br />
<br />
At the end, if you are interested in implementing the method, they are willing to share their code but after making it efficient. So, maybe there will be another paper regarding less computational cost on larger datasets with a publishable code.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=File:pooling.png&diff=42314File:pooling.png2018-12-07T01:22:45Z<p>Ka2khan: </p>
<hr />
<div></div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Fairness_Without_Demographics_in_Repeated_Loss_Minimization&diff=42312Fairness Without Demographics in Repeated Loss Minimization2018-12-07T01:17:32Z<p>Ka2khan: /* Representation Disparity */</p>
<hr />
<div>This page contains the summary of the paper "[http://proceedings.mlr.press/v80/hashimoto18a.html Fairness Without Demographics in Repeated Loss Minimization]" by Hashimoto, T. B., Srivastava, M., Namkoong, H., & Liang, P. which was published at the International Conference of Machine Learning (ICML) in 2018. <br />
<br />
=Introduction=<br />
<br />
Usually, machine learning models are minimized in their average loss to achieve high overall accuracy. While this works well for the majority, minority groups that use the system suffer high error rates because they contribute fewer data to the model. For example, non-native speakers may contribute less to the speech recognizer machine learning model. This phenomenon is known as '''''representation disparity''''' and has been observed in many models that, for instance, recognize faces, or identify the language. This disparity further widens for minority users who suffer higher error rates, as they will lower usage of the system in the future. As a result, minority groups provide even less data for future optimization of the model. When unbalanced group risk gets worse over time this is referred to as '''''disparity amplification'''''. <br />
<br />
[[File:fairness_example.JPG|700px|center]]<br />
<br />
In this paper, Hashimoto et al. provide a strategy for controlling the worst case risk amongst all groups. They first show that standard '''''empirical risk minimization (ERM)''''' does not control the loss of minority groups, and thus causes representation disparity . This representation disparity is further amplified over time (even if the model is fair in the beginning). Second, the researchers try to mitigate this unfairness by proposing the use of '''''distributionally robust optimization (DRO)'''''. Indeed, Hashimoto et al. are able to show that DRO can bound the loss for minority groups at every step of time, and is fair for models that ERM turns unfair by applying it to Amazon Mechanical Turk task.<br />
<br />
===Note on Fairness===<br />
<br />
Hashimoto et al. follow the ''difference principle'' to achieve and measure fairness. It is defined as the maximization of the welfare of the worst-off group, rather than the whole group (cf. utilitarianism).<br />
<br />
===Related Works===<br />
The recent advancements on the topic of fairness in Machine learning can be classified into the following approaches:<br />
<br />
1. Rawls Difference principle (Rawls, 2001, p155) - Defines that maximizing the welfare of the worst-off group is fair and stable over time, which increases the chance that minorities will consent to status-quo. The current work builds on this as it sees predictive accuracy as a resource to be allocated.<br />
<br />
2. Labels of minorities present in the data:<br />
* Chouldechova, 2017: Use of race (a protected label) in recidivism protection. This study evaluated the likelihood for a criminal defendant to reoffend at a later time, which assisted with criminal justice decision-making. However, a risk assessment instrument called COMPAS was studied and discovered to be biased against black defendants. As the consequences for misclassification can be dire, fairness regarding using race as a label was studied.<br />
* Barocas & Selbst, 2016: Guaranteeing fairness for a protected label through constraints such as equalized odds, disparate impact, and calibration.<br />
In the case specific to this paper, this information is not present.<br />
<br />
3. Fairness when minority grouping are not present explicitly<br />
* Dwork et al., 2012 used Individual notions of fairness using fixed similarity function whereas Kearns et al., 2018; Hebert-Johnson et al., 2017 used subgroups of a set of protected labels.<br />
* Rawlsian Fairness for Machine Learning, Matthew Joseph, Michael Kearns, Jamie Morgenstern, Seth Neel †Aaron Roth November 1, 2016 <br />
* Kearns et al. (2018); Hebert-Johnson et al. (2017) consider subgroups of a set of protected features.<br />
Again for the specific case in this paper, this is not possible.<br />
<br />
4. Online settings<br />
* Joseph et al., 2016; Jabbari et al., 2017 looked at fairness in bandit learning using algorithms compatible with Rawls’ principle on equality of opportunity.<br />
* Liu et al. (2018) analyzed fairness temporally in the context of constraint-based fairness criteria. It showed that fairness is not ensured over time when static fairness constraints are enforced.<br />
<br />
=Representation Disparity=<br />
<br />
If a user makes a query <math display="inline">Z \sim P</math>, a model <math display="inline">\theta \in \Theta</math> makes a prediction, and the user experiences loss <math display="inline">\ell (\theta; Z)</math>. <br />
<br />
The expected loss of a model <math display="inline">\theta</math> is denoted as the risk <math display="inline">\mathcal{R}(\theta) = \mathbb{E}_{Z \sim P} [\ell (\theta; Z)] </math>. <br />
<br />
If input queries are made by users from <math display="inline">K</math> latent groups, then the distribution over all queries can be re-written as <math display="inline">Z \sim P := \sum_{k \in [K]} \alpha_kP_k</math>, where <math display="inline">\alpha_k</math> is the population portion of group <math display="inline">k</math> and <math display="inline">P_k</math> is its individual distribution, and we assume these two variables are unknown.<br />
<br />
The risk associated with group <math>k</math> can be written as, <math>\mathcal{R}_k(\theta) := \mathbb{E}_{P_k} [\ell (\theta; Z)]</math>.<br />
<br />
The worst-case risk over all groups can then be defined as,<br />
\begin{align}<br />
\mathcal{R}_{max}(\theta) := \underset{k \in [K]}{\text{max}}\ \mathcal{R}_k(\theta).<br />
\end{align}<br />
<br />
Minimizing this function is equivalent to minimizing the risk for the worst-off group. <br />
<br />
There is high representation disparity if the expected loss of the model <math display="inline">\mathcal{R}(\theta)</math> is low, but the worst-case risk <math display="inline">\mathcal{R}_{max}(\theta)</math> is high. A model with high representation disparity performs well on average (i.e. has low overall loss), but fails to represent some groups <math display="inline">k</math> (i.e. the risk for the worst-off group is high).<br />
<br />
Often, groups are latent and <math display="inline">k, P_k</math> are unknown and the worst-case risks are inaccessible. The technique proposed by Hashimoto et al does not require direct access to these.<br />
<br />
=Disparity Amplification=<br />
<br />
Representation disparity can amplify as time passes and loss is minimized. Over <math display="inline">t = 1, 2, ..., T</math> minimization rounds, the group proportions <math display="inline">\alpha_k^{(t)}</math> are not constant, but vary depending on past losses. <br />
<br />
At each round the expected number of users <math display="inline">\lambda_k^{(t+1)}</math> from group <math display="inline">k</math> is determined by <br />
\begin{align}<br />
\lambda_k^{(t+1)} := \lambda_k^{(t)} \nu(\mathcal{R}_k(\theta^{(t)})) + b_k<br />
\end{align}<br />
<br />
where <math display="inline">\lambda_k^{(t)} \nu(\mathcal{R}_k(\theta^{(t)}))</math> describes the fraction of retained users from the previous optimization, <math>\nu(x)</math> is a function that decreases as <math>x</math> increases, and <math display="inline">b_k</math> is the number of new users of group <math display="inline">k</math>. <br />
<br />
Furthermore, the group proportions <math display="inline">\alpha_k^{(t)}</math>, dependent on past losses is defined as:<br />
\begin{align}<br />
\alpha_k^{(t+1)} := \dfrac{\lambda_k^{(t+1)}}{\sum_{k'\in[K]} \lambda_{k'}^{(t+1)}}<br />
\end{align}<br />
<br />
To put simply, the number of expected users of a group depends on the number of new users of that group and the fraction of users that continue to use the system from the previous optimization step. If fewer users from minority groups return to the model (i.e. the model has a low retention rate of minority group users), Hashimoto et al. argue that the representation disparity amplifies. The decrease in user retention for the minority group is exacerbated over time since once a group shrinks sufficiently, it receives higher losses relative to others, leading to even fewer samples from the group.<br />
<br />
==Empirical Risk Minimization (ERM)==<br />
<br />
Without the knowledge of population proportions <math display="inline">\alpha_k^{(t)}</math>, the new user rate <math display="inline">b_k</math>, and the retention function <math display="inline">\nu</math> it is hard to control the worst-case risk over all time periods <math display="inline">\mathcal{R}_{max}^T</math>. That is why it is the standard approach to fit a sequence of models <math display="inline">\theta^{(t)}</math> by empirically approximating them. Using ERM, for instance, the optimal model is approached by minimizing the loss of the model:<br />
<br />
\begin{align}<br />
\theta^{(t)} = arg min_{\theta \in \Theta} \sum_i \ell(\theta; Z_i^{(t)})<br />
\end{align}<br />
<br />
However, ERM fails to prevent disparity amplification. By minimizing the expected loss of the model, minority groups experience higher loss (because the loss of the majority group is minimized), and do not return to use the system. In doing so, the population proportions <math display="inline">\alpha_k^{(t)}</math> shift, and certain minority groups contribute even less to the system. This is mirrored in the expected user count <math display="inline">\lambda^{(t)}</math> at each optimization point. In their paper Hashimoto et al. show that, if using ERM, <math display="inline">\lambda^{(t)}</math> is unstable because it loses its fair fixed point (i.e. the population fraction where risk minimization maintains the same population fraction over time). Therefore, ERM fails to control minority risk over time and is considered unfair.<br />
<br />
=Distributionally Robust Optimization (DRO)=<br />
<br />
To overcome the unfairness of ERM, Hashimoto et al. developed a distributionally robust optimization (DRO). At this point the goal is still to minimize the worst-case group risk over a single time-step <math display="inline">\mathcal{R}_{max} (\theta^{(t)}) </math> (time steps are omitted in this section's formulas). As previously mentioned, this is difficult to do because neither the population proportions <math display="inline">\alpha_k </math> nor group distributions <math display="inline">P_k </math> are known, which means the data was sampled from different unknown groups. Therefore, in order to improve the performance across different groups, Hashimoto et al. developed an optimization technique that is robust "against '''''all''''' directions around the data generating distribution". This refers to the notion that DRO is robust to any group distribution <math display="inline">P_k </math> whose loss other optimization techniques such as ERM might try to optimize. To create this distributionally robustness, the optimizations risk function <math display="inline">\mathcal{R}_{dro} </math> has to "up-weigh" data <math display="inline">Z</math> that cause high loss <math display="inline">\ell(\theta, Z)</math>. In other words, the risk function has to over-represent mixture components (i.e. group distributions <math display="inline">P_k </math>) in relation to their original mixture weights (i.e. the population proportions <math display="inline">\alpha_k </math>) for groups that suffer high loss. <br />
<br />
To do this Hashimoto et al. considered the worst-case loss (i.e. the highest risk) over all perturbations <math display="inline">P_k </math> around <math display="inline">P</math> within a certain limit (because obviously not every outlier should be up-weighed). This limit is described by the <math display="inline">\chi^2</math>-divergence (i.e. the distance, roughly speaking) between probability distributions. For two distributions <math display="inline">P</math> and <math display="inline">Q</math> the divergence is defined as <math display="inline">D_{\chi^2} (P || Q):= \int (\frac{dP}{dQ} - 1)^2</math>. If <math display="inline">P</math> is not absolutely continuous w.r.t <math display="inline">Q</math>, then <math display="inline">D_{\chi^2} (P || Q):= \infty</math>. With the help of the <math display="inline">\chi^2</math>-divergence, Hashimoto et al. defined the chi-squared ball <math display="inline">\mathcal{B}(P,r)</math> around the probability distribution P. This ball is defined so that <math display="inline">\mathcal{B}(P,r) := \{Q \ll P : D_{\chi^2} (Q || P) \leq r \}</math>. With the help of this ball the worst-case loss (i.e. the highest risk) over all perturbations <math display="inline">P_k </math> that lie inside the ball (i.e. within reasonable range) around the probability distribution <math display="inline">P</math> can be considered. This loss is given by<br />
<br />
\begin{align}<br />
\mathcal{R}_{dro}(\theta, r) := \underset{Q \in \mathcal{B}(P,r)}{sup} \mathbb{E}_Q [\ell(\theta;Z)]<br />
\end{align}<br />
<br />
which for <math display="inline">P:= \sum_{k \in [K]} \alpha_k P_k</math> for all models <math display="inline">\theta \in \Theta</math> where <math display="inline">r_k := (1/a_k -1)^2</math> bounds the risk <math display="inline">\mathcal{R}_k(\theta) \leq \mathcal{R}_{dro} (\theta; r_k)</math> for each group with risk <math display="inline">\mathcal{R}_k(\theta)</math>. Furthermore, if the lower bound on the group proportions <math display="inline">\alpha_{min} \leq min_{k \in [K]} \alpha_k</math> is specified, and the radius is defined as <math display="inline">r_{max} := (1/\alpha_{min} -1)^2</math>, the worst-case risk <math display="inline">\mathcal{R}_{max} (\theta) </math> can be controlled by <math display="inline">\mathcal{R}_{dro} (\theta; r_{max}) </math> by forming an upper bound that can be minimized.<br />
<br />
==Optimization of DRO==<br />
<br />
To minimize <math display="inline">\mathcal{R}_{dro}(\theta, r) := \underset{Q \in \mathcal{B}(P,r)}{sup} \mathbb{E}_Q [\ell(\theta;Z)]</math> Hashimoto et al. look at the dual of this maximization problem (i.e. every maximization problem can be transformed into a minimization problem and vice-versa). This dual is given by the minimization problem<br />
<br />
\begin{align}<br />
\mathcal{R}_{dro}(\theta, r) = \underset{\eta \in \mathbb{R}}{inf} \left\{ F(\theta; \eta):= C\left(\mathbb{E}_P \left[ [\ell(\theta;Z) - \eta]_+^2 \right] \right)^\frac{1}{2} + \eta \right\}<br />
\end{align}<br />
<br />
with <math display="inline">C = (2(1/a_{min} - 1)^2 + 1)^{1/2}</math>. <math display="inline">\eta</math> describes the dual variable (i.e. the variable that appears in creating the dual). Since <math display="inline">F(\theta; \eta)</math> involves an expectation <math display="inline">\mathbb{E}_P</math> over the data generating distribution <math display="inline">P</math>, <math display="inline">F(\theta; \eta)</math> can be directly minimized. For convex losses <math display="inline">\ell(\theta;Z)</math>, <math display="inline">F(\theta; \eta)</math> is convex, and can be minimized by performing a binary search over <math display="inline">\eta</math>. In their paper, Hashimoto et al. further show that optimizing <math display="inline">\mathcal{R}_{dro}(\theta, r_{max})</math> at each time step controls the ''future'' worst-case risk <math display="inline">\mathcal{R}_{max} (\theta) </math>, and therefore retention rates. That means if the initial group proportions satisfy <math display="inline">\alpha_k^{(0)} \geq a_{min}</math>, and <math display="inline">\mathcal{R}_{dro}(\theta, r_{max})</math> is optimized for every time step (and therefore <math display="inline">\mathcal{R}_{max} (\theta) </math> is minimized), <math display="inline">\mathcal{R}_{max}^T (\theta) </math> over all time steps is controlled. In other words, optimizing <math display="inline">\mathcal{R}_{dro}(\theta, r_{max})</math> every time step is enough to avoid disparity amplification.<br />
<br />
<br />
Pros of DRO: In many cases, the expected value is a good measure of performance<br />
Cons of DRO: One has to know the exact distribution of the underlying distribution to perform the stochastic optimization. Deviant from the assumed distribution may result in sub-optimal solutions. The paper makes strong assumptions on <math>\mathcal{P}</math> with respect to group allocations, and thus requires a high amount of data to optimize; when assumptions are violated, the algorithm fails to perform as intended.<br />
<br />
=Experiments=<br />
<br />
The paper demonstrate the effectiveness of DRO and human evaluation of a text autocomplete system on Amazon Mechanical Turk. In both cases, DRO controls the worst-case risk over time steps and improves minority retention.<br />
Below Figure gives Inferred dynamics from a Mechanical Turk based evaluation of autocomplete systems.DRO increases minority (a) user<br />
satisfaction and (b) retention, leading to a corresponding increase in (c) user count. Error bars indicates bootstrap quartiles.<br />
[[File:fig4999.png|thumb|center|600px|]].<br />
<br />
Below figure shows how Disparity amplification in corrected by DRO. Error bars indicate quartiles over 10 replicates.<br />
[[File:fig5999.png|thumb|center|400px|]].<br />
<br />
<br />
Below figure shows Classifier accuracy as a function of group imbalance.Dotted lines show accuracy on majority group.<br />
[[File:fig6999.png|thumb|center|400px|]].<br />
<br />
It is a surprising result that the minority group has higher satisfaction and retention under DRO. Analysis of long-form comments from Turkers attribute this phenomenon to to users valuing<br />
the model’s ability to complete slang more highly than completion of common words, and indicates a slight mismatch between the authors' training loss and human satisfaction with an autocomplete system.<br />
<br />
=Critiques=<br />
<br />
This paper works on representational disparity which is a critical problem to contribute to. The methods are well developed and the paper reads coherently. However, the authors have several limiting assumptions that are not very intuitive or scientifically suggestive. The first assumption is that the <math display="inline">\eta</math> function denoting the fraction of users retained is differentiable and strictly decreasing function. This assumption does not seem practical. The second assumption is that the learned parameters are having a Poisson distribution. There is no explanation of such an assumption and reasons hinted at are hand-wavy at best. Though the authors are building a case against the Empirical risk minimization method, this method is exactly solvable when the data is linearly separable. The DRO method is computationally more complex than ERM and is not entirely clear if it will always have an advantage for a different class of problems.<br />
<br />
Note: The first assumption about <math>\eta</math> can be weakened by introducing discrete yet smooth enough function for computational proposes only. Such function will be enough to mimic for differentiability.<br />
<br />
=Other Sources=<br />
# [https://blog.acolyer.org/2018/08/17/fairness-without-demographics-in-repeated-loss-minimization/] is a easy-to-read paper description.<br />
# [https://vimeo.com/295743125] a video of the authors explaining the paper in ICML 2018<br />
<br />
=References=<br />
Rawls, J. Justice as fairness: a restatement. Harvard University Press, 2001.<br />
<br />
Barocas, S. and Selbst, A. D. Big data’s disparate impact. 104 California Law Review, 3:671–732, 2016.<br />
<br />
Chouldechova, A. A study of bias in recidivism prediction instruments. Big Data, pp. 153–163, 2017<br />
<br />
Dwork, C., Hardt, M., Pitassi, T., Reingold, O., and Zemel, R. Fairness through awareness. In Innovations in Theoretical Computer Science (ITCS), pp. 214–226, 2012.<br />
<br />
Kearns, M., Neel, S., Roth, A., and Wu, Z. S. Preventing fairness gerrymandering: Auditing and learning for subgroup fairness. arXiv preprint arXiv:1711.05144, 2018.<br />
<br />
Hebert-Johnson, ´ U., Kim, M. P., Reingold, O., and Roth-blum, G. N. Calibration for the (computationally identifiable) masses. arXiv preprint arXiv:1711.08513, 2017.<br />
<br />
Joseph, M., Kearns, M., Morgenstern, J., Neel, S., and Roth, A. Rawlsian fairness for machine learning. In FATML, 2016.<br />
<br />
Jabbari, S., Joseph, M., Kearns, M., Morgenstern, J., and Roth, A. Fairness in reinforcement learning. In International Conference on Machine Learning (ICML), pp. 1617–1626, 2017.<br />
<br />
Liu, L. T., Dean, S., Rolf, E., Simchowitz, M., and Hardt, M. Delayed impact of fair machine learning. arXiv preprint arXiv:1803.04383, 2018.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=DON%27T_DECAY_THE_LEARNING_RATE_,_INCREASE_THE_BATCH_SIZE&diff=42082DON'T DECAY THE LEARNING RATE , INCREASE THE BATCH SIZE2018-11-30T16:38:12Z<p>Ka2khan: /* INTRODUCTION */</p>
<hr />
<div>Summary of the ICLR 2018 paper: '''Don't Decay the learning Rate, Increase the Batch Size ''' <br />
<br />
Link: [https://arxiv.org/pdf/1711.00489.pdf]<br />
<br />
Summarized by: Afify, Ahmed [ID: 20700841]<br />
<br />
==INTUITION==<br />
Nowadays, it is a common practice not to have a singular steady learning rate for the learning phase of neural network models. Instead, we use adaptive learning rates with the standard gradient descent method. The intuition behind this is that when we are far away from the minima, it is beneficial for us to take large steps towards the minima, as it would require a lesser number of steps to converge, but as we approach the minima, our step size should decrease, otherwise we may just keep oscillating around the minima. In practice, this is generally achieved by methods like SGD with momentum, Nesterov momentum, and Adam. However, the core claim of this paper is that the same effect can be achieved by increasing the batch size during the gradient descent process while keeping the learning rate constant throughout. In addition, the paper argues that such an approach also reduces the parameter updates required to reach the minima, thus leading to greater parallelism and shorter training times.<br />
<br />
== INTRODUCTION ==<br />
Stochastic gradient descent (SGD) is the most widely used optimization technique for training deep learning models. The reason for this is that the minima found using this process generalizes well (Zhang et al., 2016; Wilson et al., 2017), but the optimization process is slow and time consuming as each parameter update corresponds to a small step towards the gooal. According to (Goyal et al., 2017; Hoffer et al., 2017; You et al., 2017a), this has motivated researchers to try to speed up this optimization process by taking bigger steps, and hence reduce the number of parameter updates in training a model. This can be achieved by using large batch training, which can be divided across many machines. <br />
<br />
However, increasing the batch size leads to decreasing the test set accuracy (Keskar et al., 2016; Goyal et al., 2017). Smith and Le (2017) believed that SGD has a scale of random fluctuations <math> g = \epsilon (\frac{N}{B}-1) </math>, where <math> \epsilon </math> is the learning rate, N number of training samples, and B batch size. They concluded that there is an optimal batch size proportional to the learning rate when <math> B \ll N </math>, and optimum fluctuation scale <math>g</math> at constant learning rate which maximizes test set accuracy. This was observed empirically by Goyal et al., 2017 and used to train a ResNet-50 in under an hour with 76.3% validation accuracy on ImageNet dataset.<br />
<br />
In this paper, the authors' main goal is to provide evidence that increasing the batch size is quantitatively equivalent to decreasing the learning rate. They show that this approach achieves almost equivalent model performance on the test set with the same number of training epochs but with remarkably fewer number of parameter updates. The starategy of increasing the batch size during training is in effect decreasing the scale of random fluctuations. Moreover, an additional reduction in the number of parameter updates can be attained by increasing the learning rate and scaling <math> B \propto \epsilon </math> or even more reduction by increasing the momentum coefficient and scaling <math> B \propto \frac{1}{1-m} </math> although the latter decreases the test accuracy. This has been demonstrated by several experiments on the ImageNet and CIFAR-10 datasets using ResNet-50 and Inception-ResNet-V2 architectures respectively.<br />
<br />
== STOCHASTIC GRADIENT DESCENT AND CONVEX OPTIMIZATION ==<br />
As mentioned in the previous section, the drawback of SGD when compared to full-batch training is the noise that it introduces that hinders optimization. According to (Robbins & Monro, 1951), there are two equations that govern how to reach the minimum of a convex function: (<math> \epsilon_i </math> denotes the learning rate at the <math> i^{th} </math> gradient update)<br />
<br />
<math> \sum_{i=1}^{\infty} \epsilon_i = \infty </math>. This equation guarantees that we will reach the minimum. <br />
<br />
<math> \sum_{i=1}^{\infty} \epsilon^2_i < \infty </math>. This equation, which is valid only for a fixed batch size, guarantees that learning rate decays fast enough allowing us to reach the minimum rather than bouncing due to noise.<br />
<br />
These equations indicate that the learning rate must decay during training, and second equation is only available when the batch size is constant. To change the batch size, Smith and Le (2017) proposed to interpret SGD as integrating this stochastic differential equation <math> \frac{dw}{dt} = -\frac{dC}{dw} + \eta(t) </math>, where <math>C</math> represents cost function, <math>w</math> represents the parameters, and <math>\eta</math> represents the Gaussian random noise. Furthermore, they proved that noise scale <math>g</math> controls the magnitude of random fluctuations in the training dynamics by this formula: <math> g = \epsilon (\frac{N}{B}-1) </math>, where <math> \epsilon </math> is the learning rate, N is the training set size and <math>B</math> is the batch size. As we usually have <math> B \ll N </math>, we can define <math> g \approx \epsilon \frac{N}{B} </math>. This explains why when the learning rate decreases, noise <math>g</math> decreases, enabling us to converge to the minimum of the cost function. However, increasing the batch size has the same effect and makes <math>g</math> decays with constant learning rate. In this work, the batch size is increased until <math> B \approx \frac{N}{10} </math>, then the conventional way of decaying the learning rate is followed.<br />
<br />
== SIMULATED ANNEALING AND THE GENERALIZATION GAP ==<br />
'''Simulated Annealing:''' decaying learning rates are empirically successful. To understand this, they note that introducing random fluctuations<br />
whose scale falls during training is also a well established technique in non-convex optimization; simulated annealing. The initial noisy optimization phase allows to explore a larger fraction of the parameter space without becoming trapped in local minima. Once a promising region of parameter space is located, the noise is reduced to fine-tune the parameters.<br />
<br />
For more info: Simulated annealing (SA) is a probabilistic technique for approximating the global optimum of a given function. Specifically, it is a metaheuristic to approximate global optimization in a large search space for an optimization problem. It is often used when the search space is discrete (e.g., all tours that visit a given set of cities). For problems where finding an approximate global optimum is more important than finding a precise local optimum in a fixed amount of time, simulated annealing may be preferable to alternatives such as gradient descent. [https://en.wikipedia.org/wiki/Simulated_annealing [Reference]]<br />
<br />
'''Generalization Gap:''' Small batch data generalizes better to the test set than large batch data.<br />
<br />
Smith and Le (2017) found that there is an optimal batch size which corresponds to optimal noise scale g <math> (g \approx \epsilon \frac{N}{B}) </math> and concluded that <math> B_{opt} \propto \epsilon N </math> that corresponds to maximum test set accuracy. This means that gradient noise is helpful as it makes SGD escape sharp minima, which does not generalize well. <br />
<br />
Simulated Annealing is a famous technique in non-convex optimization. Starting with noise in the training process helps us to discover a wide range of parameters then once we are near the optimum value, noise is reduced to fine tune our final parameters. However, more and more researches like to use the sharper decay schedules like cosine decay or step-function drops. In physical sciences, slowly annealing (or decaying) the temperature (which is the noise scale in this situation) helps to converge to the global minimum, which is sharp. But decaying the temperature in discrete steps can make the system stuck in a local minimum, which lead to higher cost and lower curvature. The authors think that deep learning has the same intuition.<br />
.<br />
<br />
== THE EFFECTIVE LEARNING RATE AND THE ACCUMULATION VARIABLE ==<br />
'''The Effective Learning Rate''' : <math> \epsilon_{eff} = \frac{\epsilon}{1-m} </math><br />
<br />
Smith and Le (2017) included momentum to the equation of the vanilla SGD noise scale that was defined above to be: <math> g = \frac{\epsilon}{1-m}(\frac{N}{B}-1)\approx \frac{\epsilon N}{B(1-m)} </math>, which is the same as the previous equation when m goes to 0. They found that increasing the learning rate and momentum coefficient and scaling <math> B \propto \frac{\epsilon }{1-m} </math> reduces the number of parameter updates, but the test accuracy decreases when the momentum coefficient is increased. <br />
<br />
To understand the reasons behind this, we need to analyze momentum update equations below:<br />
<br />
<center><math><br />
\Delta A = -(1-m)A + \frac{d\widehat{C}}{dw} <br />
</math><br />
<br />
<math><br />
\Delta w = -A\epsilon<br />
</math><br />
</center><br />
<br />
We can see that the Accumulation variable A, which is initially set to 0, then increases exponentially to reach its steady state value during <math> \frac{B}{N(1-m)} </math> training epochs while <math> \Delta w </math> is suppressed that can reduce the rate of convergence. Moreover, at high momentum, we have three challenges:<br />
<br />
1- Additional epochs are needed to catch up with the accumulation.<br />
<br />
2- Accumulation needs more time <math> \frac{B}{N(1-m)} </math> to forget old gradients. <br />
<br />
3- After this time, however, the accumulation cannot adapt to changes in the loss landscape.<br />
<br />
4- In the early stage, large batch size will lead to the instabilities.<br />
<br />
== EXPERIMENTS ==<br />
=== SIMULATED ANNEALING IN A WIDE RESNET ===<br />
<br />
'''Dataset:''' CIFAR-10 (50,000 training images)<br />
<br />
'''Network Architecture:''' “16-4” wide ResNet<br />
<br />
'''Training Schedules used as in the below figure:''' . These demonstrate the equivalence between decreasing the learning rate and increasing the batch size.<br />
<br />
- Decaying learning rate: learning rate decays by a factor of 5 at a sequence of “steps”, and the batch size is constant<br />
<br />
- Increasing batch size: learning rate is constant, and the batch size is increased by a factor of 5 at every step.<br />
<br />
- Hybrid: At the beginning, the learning rate is constant and batch size is increased by a factor of 5. Then, the learning rate decays by a factor of 5 at each subsequent step, and the batch size is constant. This is the schedule that will be used if there is a hardware limit affecting a maximum batch size limit.<br />
<br />
If the learning rate itself must decay during training, then these schedules should show different learning curves (as a function of the number of training epochs) and reach different final test set accuracies. Meanwhile if it is the noise scale which should decay, all three schedules should be indistinguishable.<br />
[[File:Paper_40_Fig_1.png | 800px|center]]<br />
<br />
As shown in the below figure: in the left figure (2a), we can observe that for the training set, the three learning curves are exactly the same while in figure 2b, increasing the batch size has a huge advantage of reducing the number of parameter updates.<br />
This concludes that noise scale is the one that needs to be decayed and not the learning rate itself<br />
[[File:Paper_40_Fig_2.png | 800px|center]] <br />
<br />
To make sure that these results are the same for the test set as well, in figure 3, we can see that the three learning curves are exactly the same for SGD with momentum, and Nesterov momentum<br />
[[File:Paper_40_Fig_3.png | 800px|center]]<br />
<br />
To check for other optimizers as well. the below figure shows the same experiment as in figure 3, which is the three learning curves for test set, but for vanilla SGD and Adam, and showing <br />
[[File:Paper_40_Fig_4.png | 800px|center]]<br />
<br />
'''Conclusion:''' Decreasing the learning rate and increasing the batch size during training are equivalent<br />
<br />
=== INCREASING THE EFFECTIVE LEARNING RATE===<br />
<br />
Here, the focus is on minimizing the number of parameter updates required to train a model. As shown above, the first step is to replace decaying learning rates by increasing batch sizes. Now, the authors show here that we can also increase the effective learning rate <math>\epsilon_{eff} = \epsilon/(1 − m) </math> at the start of training, while scaling the initial batch size <math>B \propto \epsilon_{eff} </math> . All experiments are conducted using SGD with momentum. There are 50000 images in the CIFAR-10 training set, and since the scaling rules only hold when <math>B << N </math> , we decided to set a maximum batch size <math>B_{max} </math>= 5120 .<br />
<br />
'''Dataset:''' CIFAR-10 (50,000 training images)<br />
<br />
'''Network Architecture:''' “16-4” wide ResNet<br />
<br />
'''Training Parameters:''' Optimization Algorithm: SGD with momentum / Maximum batch size = 5120<br />
<br />
'''Training Schedules:''' <br />
<br />
The authors consider four training schedules, all of which decay the noise scale by a factor of five in a series of three steps with the same number of epochs.<br />
<br />
Original training schedule: initial learning rate of 0.1 which decays by a factor of 5 at each step, a momentum coefficient of 0.9, and a batch size of 128. Follows the implementation of Zagoruyko & Komodakis (2016).<br />
<br />
Increasing batch size: learning rate of 0.1, momentum coefficient of 0.9, initial batch size of 128 that increases by a factor of 5 at each step. <br />
<br />
Increased initial learning rate: initial learning rate of 0.5, initial batch size of 640 that increase during training.<br />
<br />
Increased momentum coefficient: increased initial learning rate of 0.5, initial batch size of 3200 that increase during training, and an increased momentum coefficient of 0.98.<br />
<br />
The results of all training schedules, which are presented in the below figure, are documented in the following table:<br />
<br />
[[File:Paper_40_Table_1.png | 800px|center]]<br />
<br />
[[File:Paper_40_Fig_5.png | 800px|center]]<br />
<br />
<br />
<br />
'''Conclusion:''' Increasing the effective learning rate and scaling the batch size results in further reduction in the number of parameter updates<br />
<br />
=== TRAINING IMAGENET IN 2500 PARAMETER UPDATES===<br />
<br />
'''A) Experiment Goal:''' Control Batch Size<br />
<br />
'''Dataset:''' ImageNet (1.28 million training images)<br />
<br />
The paper modified the setup of Goyal et al. (2017), and used the following configuration:<br />
<br />
'''Network Architecture:''' Inception-ResNet-V2 <br />
<br />
'''Training Parameters:''' <br />
<br />
90 epochs / noise decayed at epoch 30, 60, and 80 by a factor of 10 / Initial ghost batch size = 32 / Learning rate = 3 / momentum coefficient = 0.9 / Initial batch size = 8192<br />
<br />
Two training schedules were used:<br />
<br />
“Decaying learning rate”, where batch size is fixed and the learning rate is decayed<br />
<br />
“Increasing batch size”, where batch size is increased to 81920 then the learning rate is decayed at two steps.<br />
<br />
[[File:Paper_40_Table_2.png | 800px|center]]<br />
<br />
[[File:Paper_40_Fig_6.png | 800px|center]]<br />
<br />
'''Conclusion:''' Increasing the batch size resulted in reducing the number of parameter updates from 14,000 to 6,000.<br />
<br />
'''B) Experiment Goal:''' Control Batch Size and Momentum Coefficient<br />
<br />
'''Training Parameters:''' Ghost batch size = 64 / noise decayed at epoch 30, 60, and 80 by a factor of 10. <br />
<br />
The below table shows the number of parameter updates and accuracy for different set of training parameters:<br />
<br />
[[File:Paper_40_Table_3.png | 800px|center]]<br />
<br />
[[File:Paper_40_Fig_7.png | 800px|center]]<br />
<br />
'''Conclusion:''' Increasing the momentum reduces the number of parameter updates, but leads to a drop in the test accuracy.<br />
<br />
=== TRAINING IMAGENET IN 30 MINUTES===<br />
<br />
'''Dataset:''' ImageNet (Already introduced in the previous section)<br />
<br />
'''Network Architecture:''' ResNet-50<br />
<br />
The paper replicated the setup of Goyal et al. (2017) while modifying the number of TPU devices, batch size, learning rate, and then calculating the time to complete 90 epochs, and measuring the accuracy, and performed the following experiments below:<br />
<br />
[[File:Paper_40_Table_4.png | 800px|center]]<br />
<br />
'''Conclusion:''' Model training times can be reduced by increasing the batch size during training.<br />
<br />
== RELATED WORK ==<br />
Main related work mentioned in the paper is as follows:<br />
<br />
- Smith & Le (2017) interpreted Stochastic gradient descent as stochastic differential equation; the paper built on this idea to include decaying learning rate.<br />
<br />
- Mandt et al. (2017) analyzed how to modify SGD for the task of Bayesian posterior sampling.<br />
<br />
- Keskar et al. (2016) focused on the analysis of noise once the training is started.<br />
<br />
- Moreover, the proportional relationship between batch size and learning rate was first discovered by Goyal et al. (2017) and successfully trained ResNet-50 on ImageNet in one hour after discovering the proportionality relationship between batch size and learning rate.<br />
<br />
- Furthermore, You et al. (2017a) presented Layer-wise Adaptive Rate Scaling (LARS), which is applying different learning rates to train ImageNet in 14 minutes and 74.9% accuracy. <br />
<br />
- Wilson et al. (2017) argued that adaptive optimization methods tend to generalize less well than SGD and SGD with momentum (although<br />
they did not include K-FAC in their study), while the authors' work reduces the gap in convergence speed.<br />
<br />
- Finally, another strategy called Asynchronous-SGD that allowed (Recht et al., 2011; Dean et al., 2012) to use multiple GPUs even with small batch sizes.<br />
<br />
== CONCLUSIONS ==<br />
Increasing the batch size during training has the same benefits of decaying the learning rate in addition to reducing the number of parameter updates, which corresponds to faster training time. Experiments were performed on different image datasets and various optimizers with different training schedules to prove this result. The paper proposed to increase the learning rate and momentum parameter <math>m</math>, while scaling <math> B \propto \frac{\epsilon}{1-m} </math>, which achieves fewer parameter updates, but slightly less test set accuracy as mentioned in detail in the experiments’ section. In summary, on ImageNet dataset, Inception-ResNet-V2 achieved 77% validation accuracy in under 2500 parameter updates, and ResNet-50 achieved 76.1% validation set accuracy on TPU in less than 30 minutes. One of the great finding of this paper is that all the methods use the hyper-parameters directly from previous works in the literature, and no additional hyper-parameter tuning was performed.<br />
<br />
== CRITIQUE ==<br />
'''Pros:'''<br />
<br />
- The paper showed empirically that increasing batch size and decaying learning rate are equivalent.<br />
<br />
- Several experiments were performed on different optimizers such as SGD and Adam.<br />
<br />
- Had several comparisons with previous experimental setups.<br />
<br />
'''Cons:'''<br />
<br />
<br />
- All datasets used are image datasets. Other experiments should have been done on datasets from different domains to ensure generalization. <br />
<br />
- The number of parameter updates was used as a comparison criterion, but wall-clock times could have provided additional measurable judgment although they depend on the hardware used.<br />
<br />
- Special hardware is needed for large batch training, which is not always feasible. As batch-size increases, we generally need more RAM to train the same model. However, if learning rate is decreased, the RAM use remains constant. As a result, learning rate decay will allow us to train bigger models.<br />
<br />
- In section 5.2 (Increasing the Effective Learning rate), the authors did not test a range of learning rate values and used only (0.1 and 0.5). Additional results from varying the initial learning rate values from 0.1 to 3.2 are provided in the appendix, which indicates that the test accuracy begins to fall for initial learning rates greater than ~0.4. The appended results do not show validation set accuracy curves like in Figure 6, however. It would be beneficial to see if they were similar to the original 0.1 and 0.5 initial learning rate baselines.<br />
<br />
- Although the main idea of the paper is interesting, its results does not seem to be too surprising in comparison with other recent papers in the subject.<br />
<br />
- The paper could benefit from using some other models to demonstrate its claim and generalize its idea by adding some comparisons with other models as well as other recent methods to increase batch size.<br />
<br />
- The paper presents interesting ideas. However, it lacks of mathematical and theoretical analysis beyond the idea. Since the experiment is primary on image dataset and it does not provide sufficient theories, the paper itself presents limited applicability to other types. <br />
<br />
- Also, in experimental setting, only single training runs from one random initialization is used. It would be better to take the best of many runs or to show confidence intervals.<br />
<br />
- It is proposed that we should compare learning rate decay with batch-size increase under the setting that total budget / number of training samples is fixed.<br />
<br />
== REFERENCES ==<br />
# Takuya Akiba, Shuji Suzuki, and Keisuke Fukuda. Extremely large minibatch sgd: Training resnet-50 on imagenet in 15 minutes. arXiv preprint arXiv:1711.04325, 2017.<br />
#Lukas Balles, Javier Romero, and Philipp Hennig. Coupling adaptive batch sizes with learning rates.arXiv preprint arXiv:1612.05086, 2016.<br />
#L´eon Bottou, Frank E Curtis, and Jorge Nocedal. Optimization methods for large-scale machine learning.arXiv preprint arXiv:1606.04838, 2016.<br />
#Richard H Byrd, Gillian M Chin, Jorge Nocedal, and Yuchen Wu. Sample size selection in optimization methods for machine learning. Mathematical programming, 134(1):127–155, 2012.<br />
#Pratik Chaudhari, Anna Choromanska, Stefano Soatto, and Yann LeCun. Entropy-SGD: Biasing gradient descent into wide valleys. arXiv preprint arXiv:1611.01838, 2016.<br />
#Soham De, Abhay Yadav, David Jacobs, and Tom Goldstein. Automated inference with adaptive batches. In Artificial Intelligence and Statistics, pp. 1504–1513, 2017.<br />
#Jeffrey Dean, Greg Corrado, Rajat Monga, Kai Chen, Matthieu Devin, Mark Mao, Andrew Senior, Paul Tucker, Ke Yang, Quoc V Le, et al. Large scale distributed deep networks. In Advances in neural information processing systems, pp. 1223–1231, 2012.<br />
#Michael P Friedlander and Mark Schmidt. Hybrid deterministic-stochastic methods for data fitting.SIAM Journal on Scientific Computing, 34(3):A1380–A1405, 2012.<br />
#Priya Goyal, Piotr Doll´ar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch SGD: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017.<br />
#Sepp Hochreiter and J¨urgen Schmidhuber. Flat minima. Neural Computation, 9(1):1–42, 1997.<br />
#Elad Hoffer, Itay Hubara, and Daniel Soudry. Train longer, generalize better: closing the generalization gap in large batch training of neural networks. arXiv preprint arXiv:1705.08741, 2017.<br />
#Norman P Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder Bajwa, Sarah Bates, Suresh Bhatia, Nan Boden, Al Borchers, et al. In-datacenter performance analysis of a tensor processing unit. In Proceedings of the 44th Annual International Symposium on Computer Architecture, pp. 1–12. ACM, 2017.<br />
#Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. arXiv preprint arXiv:1609.04836, 2016.<br />
#Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.<br />
#Alex Krizhevsky. One weird trick for parallelizing convolutional neural networks. arXiv preprint arXiv:1404.5997, 2014.<br />
#Qianxiao Li, Cheng Tai, and E Weinan. Stochastic modified equations and adaptive stochastic gradient algorithms. arXiv preprint arXiv:1511.06251, 2017.<br />
#Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with restarts. arXiv preprint arXiv:1608.03983, 2016.<br />
#Stephan Mandt, Matthew D Hoffman, and DavidMBlei. Stochastic gradient descent as approximate bayesian inference. arXiv preprint arXiv:1704.04289, 2017.<br />
#James Martens and Roger Grosse. Optimizing neural networks with kronecker-factored approximate curvature. In International Conference on Machine Learning, pp. 2408–2417, 2015.<br />
#Yurii Nesterov. A method of solving a convex programming problem with convergence rate o (1/k2). In Soviet Mathematics Doklady, volume 27, pp. 372–376, 1983.<br />
#Lutz Prechelt. Early stopping-but when? Neural Networks: Tricks of the trade, pp. 553–553, 1998.<br />
#Benjamin Recht, Christopher Re, Stephen Wright, and Feng Niu. Hogwild: A lock-free approach to parallelizing stochastic gradient descent. In Advances in neural information processing systems, pp. 693–701, 2011.<br />
#Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pp. 400–407, 1951.<br />
#Samuel L. Smith and Quoc V. Le. A bayesian perspective on generalization and stochastic gradient descent. arXiv preprint arXiv:1710.06451, 2017.<br />
#Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A Alemi. Inception-v4, Inception-ResNet and the impact of residual connections on learning. In AAAI, pp. 4278–4284, 2017.<br />
#Max Welling and Yee W Teh. Bayesian learning via stochastic gradient langevin dynamics. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pp. 681–688, 2011.<br />
#Ashia C Wilson, Rebecca Roelofs, Mitchell Stern, Nathan Srebro, and Benjamin Recht. The marginal value of adaptive gradient methods in machine learning. arXiv preprint arXiv:1705.08292, 2017.<br />
#Yang You, Igor Gitman, and Boris Ginsburg. Scaling SGD batch size to 32k for imagenet training. arXiv preprint arXiv:1708.03888, 2017a.<br />
#Yang You, Zhao Zhang, C Hsieh, James Demmel, and Kurt Keutzer. Imagenet training in minutes. CoRR, abs/1709.05011, 2017b.<br />
#Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016.<br />
#Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. arXiv preprint arXiv:1611.03530, 2016.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=DON%27T_DECAY_THE_LEARNING_RATE_,_INCREASE_THE_BATCH_SIZE&diff=42079DON'T DECAY THE LEARNING RATE , INCREASE THE BATCH SIZE2018-11-30T16:31:31Z<p>Ka2khan: /* INTRODUCTION */</p>
<hr />
<div>Summary of the ICLR 2018 paper: '''Don't Decay the learning Rate, Increase the Batch Size ''' <br />
<br />
Link: [https://arxiv.org/pdf/1711.00489.pdf]<br />
<br />
Summarized by: Afify, Ahmed [ID: 20700841]<br />
<br />
==INTUITION==<br />
Nowadays, it is a common practice not to have a singular steady learning rate for the learning phase of neural network models. Instead, we use adaptive learning rates with the standard gradient descent method. The intuition behind this is that when we are far away from the minima, it is beneficial for us to take large steps towards the minima, as it would require a lesser number of steps to converge, but as we approach the minima, our step size should decrease, otherwise we may just keep oscillating around the minima. In practice, this is generally achieved by methods like SGD with momentum, Nesterov momentum, and Adam. However, the core claim of this paper is that the same effect can be achieved by increasing the batch size during the gradient descent process while keeping the learning rate constant throughout. In addition, the paper argues that such an approach also reduces the parameter updates required to reach the minima, thus leading to greater parallelism and shorter training times.<br />
<br />
== INTRODUCTION ==<br />
Stochastic gradient descent (SGD) is the most widely used optimization technique for training deep learning models. The reason for this is that the minima found using this process generalizes well (Zhang et al., 2016; Wilson et al., 2017), but the optimization process is slow and time consuming as each parameter update corresponds to a small step towards the gooal. According to (Goyal et al., 2017; Hoffer et al., 2017; You et al., 2017a), this has motivated researchers to try to speed up this optimization process by taking bigger steps, and hence reduce the number of parameter updates in training a model. This can be achieved by using large batch training, which can be divided across many machines. <br />
<br />
However, increasing the batch size leads to decreasing the test set accuracy (Keskar et al., 2016; Goyal et al., 2017). Smith and Le (2017) believed that SGD has a scale of random fluctuations <math> g = \epsilon (\frac{N}{B}-1) </math>, where <math> \epsilon </math> is the learning rate, N number of training samples, and B batch size. They concluded that there is an optimal batch size proportional to the learning rate when <math> B \ll N </math>, and optimum fluctuation scale <math>g</math> at constant learning rate which maximizes test set accuracy. This was observed empirically by Goyal et al., 2017 and used to train a ResNet-50 in under an hour with 76.3% validation accuracy on ImageNet dataset.<br />
<br />
In this paper, the authors' main goal is to provide evidence that increasing the batch size is quantitatively equivalent to decreasing the learning rate. They show that this approach achieves almost equivalent model performance on the test set with the same number of training epochs in, but with remarkably less number of parameter updates. The starategy of increasing the batch size during training is in effect decreasing the scale of random fluctuations. Moreover, an additional reduction in the number of parameter updates can be attained by increasing the learning rate and scaling <math> B \propto \epsilon </math> or even more reduction by increasing the momentum coefficient and scaling <math> B \propto \frac{1}{1-m} </math> although the latter decreases the test accuracy. This has been demonstrated by several experiments on the ImageNet and CIFAR-10 datasets using ResNet-50 and Inception-ResNet-V2 architectures respectively.<br />
<br />
== STOCHASTIC GRADIENT DESCENT AND CONVEX OPTIMIZATION ==<br />
As mentioned in the previous section, the drawback of SGD when compared to full-batch training is the noise that it introduces that hinders optimization. According to (Robbins & Monro, 1951), there are two equations that govern how to reach the minimum of a convex function: (<math> \epsilon_i </math> denotes the learning rate at the <math> i^{th} </math> gradient update)<br />
<br />
<math> \sum_{i=1}^{\infty} \epsilon_i = \infty </math>. This equation guarantees that we will reach the minimum. <br />
<br />
<math> \sum_{i=1}^{\infty} \epsilon^2_i < \infty </math>. This equation, which is valid only for a fixed batch size, guarantees that learning rate decays fast enough allowing us to reach the minimum rather than bouncing due to noise.<br />
<br />
These equations indicate that the learning rate must decay during training, and second equation is only available when the batch size is constant. To change the batch size, Smith and Le (2017) proposed to interpret SGD as integrating this stochastic differential equation <math> \frac{dw}{dt} = -\frac{dC}{dw} + \eta(t) </math>, where <math>C</math> represents cost function, <math>w</math> represents the parameters, and <math>\eta</math> represents the Gaussian random noise. Furthermore, they proved that noise scale <math>g</math> controls the magnitude of random fluctuations in the training dynamics by this formula: <math> g = \epsilon (\frac{N}{B}-1) </math>, where <math> \epsilon </math> is the learning rate, N is the training set size and <math>B</math> is the batch size. As we usually have <math> B \ll N </math>, we can define <math> g \approx \epsilon \frac{N}{B} </math>. This explains why when the learning rate decreases, noise <math>g</math> decreases, enabling us to converge to the minimum of the cost function. However, increasing the batch size has the same effect and makes <math>g</math> decays with constant learning rate. In this work, the batch size is increased until <math> B \approx \frac{N}{10} </math>, then the conventional way of decaying the learning rate is followed.<br />
<br />
== SIMULATED ANNEALING AND THE GENERALIZATION GAP ==<br />
'''Simulated Annealing:''' decaying learning rates are empirically successful. To understand this, they note that introducing random fluctuations<br />
whose scale falls during training is also a well established technique in non-convex optimization; simulated annealing. The initial noisy optimization phase allows to explore a larger fraction of the parameter space without becoming trapped in local minima. Once a promising region of parameter space is located, the noise is reduced to fine-tune the parameters.<br />
<br />
For more info: Simulated annealing (SA) is a probabilistic technique for approximating the global optimum of a given function. Specifically, it is a metaheuristic to approximate global optimization in a large search space for an optimization problem. It is often used when the search space is discrete (e.g., all tours that visit a given set of cities). For problems where finding an approximate global optimum is more important than finding a precise local optimum in a fixed amount of time, simulated annealing may be preferable to alternatives such as gradient descent. [https://en.wikipedia.org/wiki/Simulated_annealing [Reference]]<br />
<br />
'''Generalization Gap:''' Small batch data generalizes better to the test set than large batch data.<br />
<br />
Smith and Le (2017) found that there is an optimal batch size which corresponds to optimal noise scale g <math> (g \approx \epsilon \frac{N}{B}) </math> and concluded that <math> B_{opt} \propto \epsilon N </math> that corresponds to maximum test set accuracy. This means that gradient noise is helpful as it makes SGD escape sharp minima, which does not generalize well. <br />
<br />
Simulated Annealing is a famous technique in non-convex optimization. Starting with noise in the training process helps us to discover a wide range of parameters then once we are near the optimum value, noise is reduced to fine tune our final parameters. However, more and more researches like to use the sharper decay schedules like cosine decay or step-function drops. In physical sciences, slowly annealing (or decaying) the temperature (which is the noise scale in this situation) helps to converge to the global minimum, which is sharp. But decaying the temperature in discrete steps can make the system stuck in a local minimum, which lead to higher cost and lower curvature. The authors think that deep learning has the same intuition.<br />
.<br />
<br />
== THE EFFECTIVE LEARNING RATE AND THE ACCUMULATION VARIABLE ==<br />
'''The Effective Learning Rate''' : <math> \epsilon_{eff} = \frac{\epsilon}{1-m} </math><br />
<br />
Smith and Le (2017) included momentum to the equation of the vanilla SGD noise scale that was defined above to be: <math> g = \frac{\epsilon}{1-m}(\frac{N}{B}-1)\approx \frac{\epsilon N}{B(1-m)} </math>, which is the same as the previous equation when m goes to 0. They found that increasing the learning rate and momentum coefficient and scaling <math> B \propto \frac{\epsilon }{1-m} </math> reduces the number of parameter updates, but the test accuracy decreases when the momentum coefficient is increased. <br />
<br />
To understand the reasons behind this, we need to analyze momentum update equations below:<br />
<br />
<center><math><br />
\Delta A = -(1-m)A + \frac{d\widehat{C}}{dw} <br />
</math><br />
<br />
<math><br />
\Delta w = -A\epsilon<br />
</math><br />
</center><br />
<br />
We can see that the Accumulation variable A, which is initially set to 0, then increases exponentially to reach its steady state value during <math> \frac{B}{N(1-m)} </math> training epochs while <math> \Delta w </math> is suppressed that can reduce the rate of convergence. Moreover, at high momentum, we have three challenges:<br />
<br />
1- Additional epochs are needed to catch up with the accumulation.<br />
<br />
2- Accumulation needs more time <math> \frac{B}{N(1-m)} </math> to forget old gradients. <br />
<br />
3- After this time, however, the accumulation cannot adapt to changes in the loss landscape.<br />
<br />
4- In the early stage, large batch size will lead to the instabilities.<br />
<br />
== EXPERIMENTS ==<br />
=== SIMULATED ANNEALING IN A WIDE RESNET ===<br />
<br />
'''Dataset:''' CIFAR-10 (50,000 training images)<br />
<br />
'''Network Architecture:''' “16-4” wide ResNet<br />
<br />
'''Training Schedules used as in the below figure:''' . These demonstrate the equivalence between decreasing the learning rate and increasing the batch size.<br />
<br />
- Decaying learning rate: learning rate decays by a factor of 5 at a sequence of “steps”, and the batch size is constant<br />
<br />
- Increasing batch size: learning rate is constant, and the batch size is increased by a factor of 5 at every step.<br />
<br />
- Hybrid: At the beginning, the learning rate is constant and batch size is increased by a factor of 5. Then, the learning rate decays by a factor of 5 at each subsequent step, and the batch size is constant. This is the schedule that will be used if there is a hardware limit affecting a maximum batch size limit.<br />
<br />
If the learning rate itself must decay during training, then these schedules should show different learning curves (as a function of the number of training epochs) and reach different final test set accuracies. Meanwhile if it is the noise scale which should decay, all three schedules should be indistinguishable.<br />
[[File:Paper_40_Fig_1.png | 800px|center]]<br />
<br />
As shown in the below figure: in the left figure (2a), we can observe that for the training set, the three learning curves are exactly the same while in figure 2b, increasing the batch size has a huge advantage of reducing the number of parameter updates.<br />
This concludes that noise scale is the one that needs to be decayed and not the learning rate itself<br />
[[File:Paper_40_Fig_2.png | 800px|center]] <br />
<br />
To make sure that these results are the same for the test set as well, in figure 3, we can see that the three learning curves are exactly the same for SGD with momentum, and Nesterov momentum<br />
[[File:Paper_40_Fig_3.png | 800px|center]]<br />
<br />
To check for other optimizers as well. the below figure shows the same experiment as in figure 3, which is the three learning curves for test set, but for vanilla SGD and Adam, and showing <br />
[[File:Paper_40_Fig_4.png | 800px|center]]<br />
<br />
'''Conclusion:''' Decreasing the learning rate and increasing the batch size during training are equivalent<br />
<br />
=== INCREASING THE EFFECTIVE LEARNING RATE===<br />
<br />
Here, the focus is on minimizing the number of parameter updates required to train a model. As shown above, the first step is to replace decaying learning rates by increasing batch sizes. Now, the authors show here that we can also increase the effective learning rate <math>\epsilon_{eff} = \epsilon/(1 − m) </math> at the start of training, while scaling the initial batch size <math>B \propto \epsilon_{eff} </math> . All experiments are conducted using SGD with momentum. There are 50000 images in the CIFAR-10 training set, and since the scaling rules only hold when <math>B << N </math> , we decided to set a maximum batch size <math>B_{max} </math>= 5120 .<br />
<br />
'''Dataset:''' CIFAR-10 (50,000 training images)<br />
<br />
'''Network Architecture:''' “16-4” wide ResNet<br />
<br />
'''Training Parameters:''' Optimization Algorithm: SGD with momentum / Maximum batch size = 5120<br />
<br />
'''Training Schedules:''' <br />
<br />
The authors consider four training schedules, all of which decay the noise scale by a factor of five in a series of three steps with the same number of epochs.<br />
<br />
Original training schedule: initial learning rate of 0.1 which decays by a factor of 5 at each step, a momentum coefficient of 0.9, and a batch size of 128. Follows the implementation of Zagoruyko & Komodakis (2016).<br />
<br />
Increasing batch size: learning rate of 0.1, momentum coefficient of 0.9, initial batch size of 128 that increases by a factor of 5 at each step. <br />
<br />
Increased initial learning rate: initial learning rate of 0.5, initial batch size of 640 that increase during training.<br />
<br />
Increased momentum coefficient: increased initial learning rate of 0.5, initial batch size of 3200 that increase during training, and an increased momentum coefficient of 0.98.<br />
<br />
The results of all training schedules, which are presented in the below figure, are documented in the following table:<br />
<br />
[[File:Paper_40_Table_1.png | 800px|center]]<br />
<br />
[[File:Paper_40_Fig_5.png | 800px|center]]<br />
<br />
<br />
<br />
'''Conclusion:''' Increasing the effective learning rate and scaling the batch size results in further reduction in the number of parameter updates<br />
<br />
=== TRAINING IMAGENET IN 2500 PARAMETER UPDATES===<br />
<br />
'''A) Experiment Goal:''' Control Batch Size<br />
<br />
'''Dataset:''' ImageNet (1.28 million training images)<br />
<br />
The paper modified the setup of Goyal et al. (2017), and used the following configuration:<br />
<br />
'''Network Architecture:''' Inception-ResNet-V2 <br />
<br />
'''Training Parameters:''' <br />
<br />
90 epochs / noise decayed at epoch 30, 60, and 80 by a factor of 10 / Initial ghost batch size = 32 / Learning rate = 3 / momentum coefficient = 0.9 / Initial batch size = 8192<br />
<br />
Two training schedules were used:<br />
<br />
“Decaying learning rate”, where batch size is fixed and the learning rate is decayed<br />
<br />
“Increasing batch size”, where batch size is increased to 81920 then the learning rate is decayed at two steps.<br />
<br />
[[File:Paper_40_Table_2.png | 800px|center]]<br />
<br />
[[File:Paper_40_Fig_6.png | 800px|center]]<br />
<br />
'''Conclusion:''' Increasing the batch size resulted in reducing the number of parameter updates from 14,000 to 6,000.<br />
<br />
'''B) Experiment Goal:''' Control Batch Size and Momentum Coefficient<br />
<br />
'''Training Parameters:''' Ghost batch size = 64 / noise decayed at epoch 30, 60, and 80 by a factor of 10. <br />
<br />
The below table shows the number of parameter updates and accuracy for different set of training parameters:<br />
<br />
[[File:Paper_40_Table_3.png | 800px|center]]<br />
<br />
[[File:Paper_40_Fig_7.png | 800px|center]]<br />
<br />
'''Conclusion:''' Increasing the momentum reduces the number of parameter updates, but leads to a drop in the test accuracy.<br />
<br />
=== TRAINING IMAGENET IN 30 MINUTES===<br />
<br />
'''Dataset:''' ImageNet (Already introduced in the previous section)<br />
<br />
'''Network Architecture:''' ResNet-50<br />
<br />
The paper replicated the setup of Goyal et al. (2017) while modifying the number of TPU devices, batch size, learning rate, and then calculating the time to complete 90 epochs, and measuring the accuracy, and performed the following experiments below:<br />
<br />
[[File:Paper_40_Table_4.png | 800px|center]]<br />
<br />
'''Conclusion:''' Model training times can be reduced by increasing the batch size during training.<br />
<br />
== RELATED WORK ==<br />
Main related work mentioned in the paper is as follows:<br />
<br />
- Smith & Le (2017) interpreted Stochastic gradient descent as stochastic differential equation; the paper built on this idea to include decaying learning rate.<br />
<br />
- Mandt et al. (2017) analyzed how to modify SGD for the task of Bayesian posterior sampling.<br />
<br />
- Keskar et al. (2016) focused on the analysis of noise once the training is started.<br />
<br />
- Moreover, the proportional relationship between batch size and learning rate was first discovered by Goyal et al. (2017) and successfully trained ResNet-50 on ImageNet in one hour after discovering the proportionality relationship between batch size and learning rate.<br />
<br />
- Furthermore, You et al. (2017a) presented Layer-wise Adaptive Rate Scaling (LARS), which is applying different learning rates to train ImageNet in 14 minutes and 74.9% accuracy. <br />
<br />
- Wilson et al. (2017) argued that adaptive optimization methods tend to generalize less well than SGD and SGD with momentum (although<br />
they did not include K-FAC in their study), while the authors' work reduces the gap in convergence speed.<br />
<br />
- Finally, another strategy called Asynchronous-SGD that allowed (Recht et al., 2011; Dean et al., 2012) to use multiple GPUs even with small batch sizes.<br />
<br />
== CONCLUSIONS ==<br />
Increasing the batch size during training has the same benefits of decaying the learning rate in addition to reducing the number of parameter updates, which corresponds to faster training time. Experiments were performed on different image datasets and various optimizers with different training schedules to prove this result. The paper proposed to increase the learning rate and momentum parameter <math>m</math>, while scaling <math> B \propto \frac{\epsilon}{1-m} </math>, which achieves fewer parameter updates, but slightly less test set accuracy as mentioned in detail in the experiments’ section. In summary, on ImageNet dataset, Inception-ResNet-V2 achieved 77% validation accuracy in under 2500 parameter updates, and ResNet-50 achieved 76.1% validation set accuracy on TPU in less than 30 minutes. One of the great finding of this paper is that all the methods use the hyper-parameters directly from previous works in the literature, and no additional hyper-parameter tuning was performed.<br />
<br />
== CRITIQUE ==<br />
'''Pros:'''<br />
<br />
- The paper showed empirically that increasing batch size and decaying learning rate are equivalent.<br />
<br />
- Several experiments were performed on different optimizers such as SGD and Adam.<br />
<br />
- Had several comparisons with previous experimental setups.<br />
<br />
'''Cons:'''<br />
<br />
<br />
- All datasets used are image datasets. Other experiments should have been done on datasets from different domains to ensure generalization. <br />
<br />
- The number of parameter updates was used as a comparison criterion, but wall-clock times could have provided additional measurable judgment although they depend on the hardware used.<br />
<br />
- Special hardware is needed for large batch training, which is not always feasible. As batch-size increases, we generally need more RAM to train the same model. However, if learning rate is decreased, the RAM use remains constant. As a result, learning rate decay will allow us to train bigger models.<br />
<br />
- In section 5.2 (Increasing the Effective Learning rate), the authors did not test a range of learning rate values and used only (0.1 and 0.5). Additional results from varying the initial learning rate values from 0.1 to 3.2 are provided in the appendix, which indicates that the test accuracy begins to fall for initial learning rates greater than ~0.4. The appended results do not show validation set accuracy curves like in Figure 6, however. It would be beneficial to see if they were similar to the original 0.1 and 0.5 initial learning rate baselines.<br />
<br />
- Although the main idea of the paper is interesting, its results does not seem to be too surprising in comparison with other recent papers in the subject.<br />
<br />
- The paper could benefit from using some other models to demonstrate its claim and generalize its idea by adding some comparisons with other models as well as other recent methods to increase batch size.<br />
<br />
- The paper presents interesting ideas. However, it lacks of mathematical and theoretical analysis beyond the idea. Since the experiment is primary on image dataset and it does not provide sufficient theories, the paper itself presents limited applicability to other types. <br />
<br />
- Also, in experimental setting, only single training runs from one random initialization is used. It would be better to take the best of many runs or to show confidence intervals.<br />
<br />
- It is proposed that we should compare learning rate decay with batch-size increase under the setting that total budget / number of training samples is fixed.<br />
<br />
== REFERENCES ==<br />
# Takuya Akiba, Shuji Suzuki, and Keisuke Fukuda. Extremely large minibatch sgd: Training resnet-50 on imagenet in 15 minutes. arXiv preprint arXiv:1711.04325, 2017.<br />
#Lukas Balles, Javier Romero, and Philipp Hennig. Coupling adaptive batch sizes with learning rates.arXiv preprint arXiv:1612.05086, 2016.<br />
#L´eon Bottou, Frank E Curtis, and Jorge Nocedal. Optimization methods for large-scale machine learning.arXiv preprint arXiv:1606.04838, 2016.<br />
#Richard H Byrd, Gillian M Chin, Jorge Nocedal, and Yuchen Wu. Sample size selection in optimization methods for machine learning. Mathematical programming, 134(1):127–155, 2012.<br />
#Pratik Chaudhari, Anna Choromanska, Stefano Soatto, and Yann LeCun. Entropy-SGD: Biasing gradient descent into wide valleys. arXiv preprint arXiv:1611.01838, 2016.<br />
#Soham De, Abhay Yadav, David Jacobs, and Tom Goldstein. Automated inference with adaptive batches. In Artificial Intelligence and Statistics, pp. 1504–1513, 2017.<br />
#Jeffrey Dean, Greg Corrado, Rajat Monga, Kai Chen, Matthieu Devin, Mark Mao, Andrew Senior, Paul Tucker, Ke Yang, Quoc V Le, et al. Large scale distributed deep networks. In Advances in neural information processing systems, pp. 1223–1231, 2012.<br />
#Michael P Friedlander and Mark Schmidt. Hybrid deterministic-stochastic methods for data fitting.SIAM Journal on Scientific Computing, 34(3):A1380–A1405, 2012.<br />
#Priya Goyal, Piotr Doll´ar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch SGD: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017.<br />
#Sepp Hochreiter and J¨urgen Schmidhuber. Flat minima. Neural Computation, 9(1):1–42, 1997.<br />
#Elad Hoffer, Itay Hubara, and Daniel Soudry. Train longer, generalize better: closing the generalization gap in large batch training of neural networks. arXiv preprint arXiv:1705.08741, 2017.<br />
#Norman P Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder Bajwa, Sarah Bates, Suresh Bhatia, Nan Boden, Al Borchers, et al. In-datacenter performance analysis of a tensor processing unit. In Proceedings of the 44th Annual International Symposium on Computer Architecture, pp. 1–12. ACM, 2017.<br />
#Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. arXiv preprint arXiv:1609.04836, 2016.<br />
#Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.<br />
#Alex Krizhevsky. One weird trick for parallelizing convolutional neural networks. arXiv preprint arXiv:1404.5997, 2014.<br />
#Qianxiao Li, Cheng Tai, and E Weinan. Stochastic modified equations and adaptive stochastic gradient algorithms. arXiv preprint arXiv:1511.06251, 2017.<br />
#Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with restarts. arXiv preprint arXiv:1608.03983, 2016.<br />
#Stephan Mandt, Matthew D Hoffman, and DavidMBlei. Stochastic gradient descent as approximate bayesian inference. arXiv preprint arXiv:1704.04289, 2017.<br />
#James Martens and Roger Grosse. Optimizing neural networks with kronecker-factored approximate curvature. In International Conference on Machine Learning, pp. 2408–2417, 2015.<br />
#Yurii Nesterov. A method of solving a convex programming problem with convergence rate o (1/k2). In Soviet Mathematics Doklady, volume 27, pp. 372–376, 1983.<br />
#Lutz Prechelt. Early stopping-but when? Neural Networks: Tricks of the trade, pp. 553–553, 1998.<br />
#Benjamin Recht, Christopher Re, Stephen Wright, and Feng Niu. Hogwild: A lock-free approach to parallelizing stochastic gradient descent. In Advances in neural information processing systems, pp. 693–701, 2011.<br />
#Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pp. 400–407, 1951.<br />
#Samuel L. Smith and Quoc V. Le. A bayesian perspective on generalization and stochastic gradient descent. arXiv preprint arXiv:1710.06451, 2017.<br />
#Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A Alemi. Inception-v4, Inception-ResNet and the impact of residual connections on learning. In AAAI, pp. 4278–4284, 2017.<br />
#Max Welling and Yee W Teh. Bayesian learning via stochastic gradient langevin dynamics. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pp. 681–688, 2011.<br />
#Ashia C Wilson, Rebecca Roelofs, Mitchell Stern, Nathan Srebro, and Benjamin Recht. The marginal value of adaptive gradient methods in machine learning. arXiv preprint arXiv:1705.08292, 2017.<br />
#Yang You, Igor Gitman, and Boris Ginsburg. Scaling SGD batch size to 32k for imagenet training. arXiv preprint arXiv:1708.03888, 2017a.<br />
#Yang You, Zhao Zhang, C Hsieh, James Demmel, and Kurt Keutzer. Imagenet training in minutes. CoRR, abs/1709.05011, 2017b.<br />
#Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016.<br />
#Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. arXiv preprint arXiv:1611.03530, 2016.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=DON%27T_DECAY_THE_LEARNING_RATE_,_INCREASE_THE_BATCH_SIZE&diff=42075DON'T DECAY THE LEARNING RATE , INCREASE THE BATCH SIZE2018-11-30T16:24:48Z<p>Ka2khan: /* INTRODUCTION */</p>
<hr />
<div>Summary of the ICLR 2018 paper: '''Don't Decay the learning Rate, Increase the Batch Size ''' <br />
<br />
Link: [https://arxiv.org/pdf/1711.00489.pdf]<br />
<br />
Summarized by: Afify, Ahmed [ID: 20700841]<br />
<br />
==INTUITION==<br />
Nowadays, it is a common practice not to have a singular steady learning rate for the learning phase of neural network models. Instead, we use adaptive learning rates with the standard gradient descent method. The intuition behind this is that when we are far away from the minima, it is beneficial for us to take large steps towards the minima, as it would require a lesser number of steps to converge, but as we approach the minima, our step size should decrease, otherwise we may just keep oscillating around the minima. In practice, this is generally achieved by methods like SGD with momentum, Nesterov momentum, and Adam. However, the core claim of this paper is that the same effect can be achieved by increasing the batch size during the gradient descent process while keeping the learning rate constant throughout. In addition, the paper argues that such an approach also reduces the parameter updates required to reach the minima, thus leading to greater parallelism and shorter training times.<br />
<br />
== INTRODUCTION ==<br />
Stochastic gradient descent (SGD) is the most widely used optimization technique for training deep learning models. The reason for this is that the minima found using this process generalizes well (Zhang et al., 2016; Wilson et al., 2017), but the optimization process is slow and time consuming as each parameter update corresponds to a small step towards the gooal. According to (Goyal et al., 2017; Hoffer et al., 2017; You et al., 2017a), this has motivated researchers to try to speed up this optimization process by taking bigger steps, and hence reduce the number of parameter updates in training a model. This can be achieved by using large batch training, which can be divided across many machines. <br />
<br />
However, increasing the batch size leads to decreasing the test set accuracy (Keskar et al., 2016; Goyal et al., 2017). Smith and Le (2017) believed that SGD has a scale of random fluctuations <math> g = \epsilon (\frac{N}{B}-1) </math>, where <math> \epsilon </math> is the learning rate, N number of training samples, and B batch size. They concluded that there is an optimal batch size proportional to the learning rate when <math> B \ll N </math>, and optimum fluctuation scale <math>g</math> at constant learning rate which maximizes test set accuracy. This was observed empirically by Goyal et al., 2017 and used to train a ResNet-50 in under an hour with 76.3% validation accuracy on ImageNet dataset.<br />
<br />
In this paper, the authors' main goal is to provide evidence that increasing the batch size is quantitatively equivalent to decreasing the learning rate with the same number of training epochs in decreasing the scale of random fluctuations, but with remarkably less number of parameter updates. Moreover, an additional reduction in the number of parameter updates can be attained by increasing the learning rate and scaling <math> B \propto \epsilon </math> or even more reduction by increasing the momentum coefficient and scaling <math> B \propto \frac{1}{1-m} </math> although the latter decreases the test accuracy. This has been demonstrated by several experiments on the ImageNet and CIFAR-10 datasets using ResNet-50 and Inception-ResNet-V2 architectures respectively.<br />
<br />
== STOCHASTIC GRADIENT DESCENT AND CONVEX OPTIMIZATION ==<br />
As mentioned in the previous section, the drawback of SGD when compared to full-batch training is the noise that it introduces that hinders optimization. According to (Robbins & Monro, 1951), there are two equations that govern how to reach the minimum of a convex function: (<math> \epsilon_i </math> denotes the learning rate at the <math> i^{th} </math> gradient update)<br />
<br />
<math> \sum_{i=1}^{\infty} \epsilon_i = \infty </math>. This equation guarantees that we will reach the minimum. <br />
<br />
<math> \sum_{i=1}^{\infty} \epsilon^2_i < \infty </math>. This equation, which is valid only for a fixed batch size, guarantees that learning rate decays fast enough allowing us to reach the minimum rather than bouncing due to noise.<br />
<br />
These equations indicate that the learning rate must decay during training, and second equation is only available when the batch size is constant. To change the batch size, Smith and Le (2017) proposed to interpret SGD as integrating this stochastic differential equation <math> \frac{dw}{dt} = -\frac{dC}{dw} + \eta(t) </math>, where <math>C</math> represents cost function, <math>w</math> represents the parameters, and <math>\eta</math> represents the Gaussian random noise. Furthermore, they proved that noise scale <math>g</math> controls the magnitude of random fluctuations in the training dynamics by this formula: <math> g = \epsilon (\frac{N}{B}-1) </math>, where <math> \epsilon </math> is the learning rate, N is the training set size and <math>B</math> is the batch size. As we usually have <math> B \ll N </math>, we can define <math> g \approx \epsilon \frac{N}{B} </math>. This explains why when the learning rate decreases, noise <math>g</math> decreases, enabling us to converge to the minimum of the cost function. However, increasing the batch size has the same effect and makes <math>g</math> decays with constant learning rate. In this work, the batch size is increased until <math> B \approx \frac{N}{10} </math>, then the conventional way of decaying the learning rate is followed.<br />
<br />
== SIMULATED ANNEALING AND THE GENERALIZATION GAP ==<br />
'''Simulated Annealing:''' decaying learning rates are empirically successful. To understand this, they note that introducing random fluctuations<br />
whose scale falls during training is also a well established technique in non-convex optimization; simulated annealing. The initial noisy optimization phase allows to explore a larger fraction of the parameter space without becoming trapped in local minima. Once a promising region of parameter space is located, the noise is reduced to fine-tune the parameters.<br />
<br />
For more info: Simulated annealing (SA) is a probabilistic technique for approximating the global optimum of a given function. Specifically, it is a metaheuristic to approximate global optimization in a large search space for an optimization problem. It is often used when the search space is discrete (e.g., all tours that visit a given set of cities). For problems where finding an approximate global optimum is more important than finding a precise local optimum in a fixed amount of time, simulated annealing may be preferable to alternatives such as gradient descent. [https://en.wikipedia.org/wiki/Simulated_annealing [Reference]]<br />
<br />
'''Generalization Gap:''' Small batch data generalizes better to the test set than large batch data.<br />
<br />
Smith and Le (2017) found that there is an optimal batch size which corresponds to optimal noise scale g <math> (g \approx \epsilon \frac{N}{B}) </math> and concluded that <math> B_{opt} \propto \epsilon N </math> that corresponds to maximum test set accuracy. This means that gradient noise is helpful as it makes SGD escape sharp minima, which does not generalize well. <br />
<br />
Simulated Annealing is a famous technique in non-convex optimization. Starting with noise in the training process helps us to discover a wide range of parameters then once we are near the optimum value, noise is reduced to fine tune our final parameters. However, more and more researches like to use the sharper decay schedules like cosine decay or step-function drops. In physical sciences, slowly annealing (or decaying) the temperature (which is the noise scale in this situation) helps to converge to the global minimum, which is sharp. But decaying the temperature in discrete steps can make the system stuck in a local minimum, which lead to higher cost and lower curvature. The authors think that deep learning has the same intuition.<br />
.<br />
<br />
== THE EFFECTIVE LEARNING RATE AND THE ACCUMULATION VARIABLE ==<br />
'''The Effective Learning Rate''' : <math> \epsilon_{eff} = \frac{\epsilon}{1-m} </math><br />
<br />
Smith and Le (2017) included momentum to the equation of the vanilla SGD noise scale that was defined above to be: <math> g = \frac{\epsilon}{1-m}(\frac{N}{B}-1)\approx \frac{\epsilon N}{B(1-m)} </math>, which is the same as the previous equation when m goes to 0. They found that increasing the learning rate and momentum coefficient and scaling <math> B \propto \frac{\epsilon }{1-m} </math> reduces the number of parameter updates, but the test accuracy decreases when the momentum coefficient is increased. <br />
<br />
To understand the reasons behind this, we need to analyze momentum update equations below:<br />
<br />
<center><math><br />
\Delta A = -(1-m)A + \frac{d\widehat{C}}{dw} <br />
</math><br />
<br />
<math><br />
\Delta w = -A\epsilon<br />
</math><br />
</center><br />
<br />
We can see that the Accumulation variable A, which is initially set to 0, then increases exponentially to reach its steady state value during <math> \frac{B}{N(1-m)} </math> training epochs while <math> \Delta w </math> is suppressed that can reduce the rate of convergence. Moreover, at high momentum, we have three challenges:<br />
<br />
1- Additional epochs are needed to catch up with the accumulation.<br />
<br />
2- Accumulation needs more time <math> \frac{B}{N(1-m)} </math> to forget old gradients. <br />
<br />
3- After this time, however, the accumulation cannot adapt to changes in the loss landscape.<br />
<br />
4- In the early stage, large batch size will lead to the instabilities.<br />
<br />
== EXPERIMENTS ==<br />
=== SIMULATED ANNEALING IN A WIDE RESNET ===<br />
<br />
'''Dataset:''' CIFAR-10 (50,000 training images)<br />
<br />
'''Network Architecture:''' “16-4” wide ResNet<br />
<br />
'''Training Schedules used as in the below figure:''' . These demonstrate the equivalence between decreasing the learning rate and increasing the batch size.<br />
<br />
- Decaying learning rate: learning rate decays by a factor of 5 at a sequence of “steps”, and the batch size is constant<br />
<br />
- Increasing batch size: learning rate is constant, and the batch size is increased by a factor of 5 at every step.<br />
<br />
- Hybrid: At the beginning, the learning rate is constant and batch size is increased by a factor of 5. Then, the learning rate decays by a factor of 5 at each subsequent step, and the batch size is constant. This is the schedule that will be used if there is a hardware limit affecting a maximum batch size limit.<br />
<br />
If the learning rate itself must decay during training, then these schedules should show different learning curves (as a function of the number of training epochs) and reach different final test set accuracies. Meanwhile if it is the noise scale which should decay, all three schedules should be indistinguishable.<br />
[[File:Paper_40_Fig_1.png | 800px|center]]<br />
<br />
As shown in the below figure: in the left figure (2a), we can observe that for the training set, the three learning curves are exactly the same while in figure 2b, increasing the batch size has a huge advantage of reducing the number of parameter updates.<br />
This concludes that noise scale is the one that needs to be decayed and not the learning rate itself<br />
[[File:Paper_40_Fig_2.png | 800px|center]] <br />
<br />
To make sure that these results are the same for the test set as well, in figure 3, we can see that the three learning curves are exactly the same for SGD with momentum, and Nesterov momentum<br />
[[File:Paper_40_Fig_3.png | 800px|center]]<br />
<br />
To check for other optimizers as well. the below figure shows the same experiment as in figure 3, which is the three learning curves for test set, but for vanilla SGD and Adam, and showing <br />
[[File:Paper_40_Fig_4.png | 800px|center]]<br />
<br />
'''Conclusion:''' Decreasing the learning rate and increasing the batch size during training are equivalent<br />
<br />
=== INCREASING THE EFFECTIVE LEARNING RATE===<br />
<br />
Here, the focus is on minimizing the number of parameter updates required to train a model. As shown above, the first step is to replace decaying learning rates by increasing batch sizes. Now, the authors show here that we can also increase the effective learning rate <math>\epsilon_{eff} = \epsilon/(1 − m) </math> at the start of training, while scaling the initial batch size <math>B \propto \epsilon_{eff} </math> . All experiments are conducted using SGD with momentum. There are 50000 images in the CIFAR-10 training set, and since the scaling rules only hold when <math>B << N </math> , we decided to set a maximum batch size <math>B_{max} </math>= 5120 .<br />
<br />
'''Dataset:''' CIFAR-10 (50,000 training images)<br />
<br />
'''Network Architecture:''' “16-4” wide ResNet<br />
<br />
'''Training Parameters:''' Optimization Algorithm: SGD with momentum / Maximum batch size = 5120<br />
<br />
'''Training Schedules:''' <br />
<br />
The authors consider four training schedules, all of which decay the noise scale by a factor of five in a series of three steps with the same number of epochs.<br />
<br />
Original training schedule: initial learning rate of 0.1 which decays by a factor of 5 at each step, a momentum coefficient of 0.9, and a batch size of 128. Follows the implementation of Zagoruyko & Komodakis (2016).<br />
<br />
Increasing batch size: learning rate of 0.1, momentum coefficient of 0.9, initial batch size of 128 that increases by a factor of 5 at each step. <br />
<br />
Increased initial learning rate: initial learning rate of 0.5, initial batch size of 640 that increase during training.<br />
<br />
Increased momentum coefficient: increased initial learning rate of 0.5, initial batch size of 3200 that increase during training, and an increased momentum coefficient of 0.98.<br />
<br />
The results of all training schedules, which are presented in the below figure, are documented in the following table:<br />
<br />
[[File:Paper_40_Table_1.png | 800px|center]]<br />
<br />
[[File:Paper_40_Fig_5.png | 800px|center]]<br />
<br />
<br />
<br />
'''Conclusion:''' Increasing the effective learning rate and scaling the batch size results in further reduction in the number of parameter updates<br />
<br />
=== TRAINING IMAGENET IN 2500 PARAMETER UPDATES===<br />
<br />
'''A) Experiment Goal:''' Control Batch Size<br />
<br />
'''Dataset:''' ImageNet (1.28 million training images)<br />
<br />
The paper modified the setup of Goyal et al. (2017), and used the following configuration:<br />
<br />
'''Network Architecture:''' Inception-ResNet-V2 <br />
<br />
'''Training Parameters:''' <br />
<br />
90 epochs / noise decayed at epoch 30, 60, and 80 by a factor of 10 / Initial ghost batch size = 32 / Learning rate = 3 / momentum coefficient = 0.9 / Initial batch size = 8192<br />
<br />
Two training schedules were used:<br />
<br />
“Decaying learning rate”, where batch size is fixed and the learning rate is decayed<br />
<br />
“Increasing batch size”, where batch size is increased to 81920 then the learning rate is decayed at two steps.<br />
<br />
[[File:Paper_40_Table_2.png | 800px|center]]<br />
<br />
[[File:Paper_40_Fig_6.png | 800px|center]]<br />
<br />
'''Conclusion:''' Increasing the batch size resulted in reducing the number of parameter updates from 14,000 to 6,000.<br />
<br />
'''B) Experiment Goal:''' Control Batch Size and Momentum Coefficient<br />
<br />
'''Training Parameters:''' Ghost batch size = 64 / noise decayed at epoch 30, 60, and 80 by a factor of 10. <br />
<br />
The below table shows the number of parameter updates and accuracy for different set of training parameters:<br />
<br />
[[File:Paper_40_Table_3.png | 800px|center]]<br />
<br />
[[File:Paper_40_Fig_7.png | 800px|center]]<br />
<br />
'''Conclusion:''' Increasing the momentum reduces the number of parameter updates, but leads to a drop in the test accuracy.<br />
<br />
=== TRAINING IMAGENET IN 30 MINUTES===<br />
<br />
'''Dataset:''' ImageNet (Already introduced in the previous section)<br />
<br />
'''Network Architecture:''' ResNet-50<br />
<br />
The paper replicated the setup of Goyal et al. (2017) while modifying the number of TPU devices, batch size, learning rate, and then calculating the time to complete 90 epochs, and measuring the accuracy, and performed the following experiments below:<br />
<br />
[[File:Paper_40_Table_4.png | 800px|center]]<br />
<br />
'''Conclusion:''' Model training times can be reduced by increasing the batch size during training.<br />
<br />
== RELATED WORK ==<br />
Main related work mentioned in the paper is as follows:<br />
<br />
- Smith & Le (2017) interpreted Stochastic gradient descent as stochastic differential equation; the paper built on this idea to include decaying learning rate.<br />
<br />
- Mandt et al. (2017) analyzed how to modify SGD for the task of Bayesian posterior sampling.<br />
<br />
- Keskar et al. (2016) focused on the analysis of noise once the training is started.<br />
<br />
- Moreover, the proportional relationship between batch size and learning rate was first discovered by Goyal et al. (2017) and successfully trained ResNet-50 on ImageNet in one hour after discovering the proportionality relationship between batch size and learning rate.<br />
<br />
- Furthermore, You et al. (2017a) presented Layer-wise Adaptive Rate Scaling (LARS), which is applying different learning rates to train ImageNet in 14 minutes and 74.9% accuracy. <br />
<br />
- Wilson et al. (2017) argued that adaptive optimization methods tend to generalize less well than SGD and SGD with momentum (although<br />
they did not include K-FAC in their study), while the authors' work reduces the gap in convergence speed.<br />
<br />
- Finally, another strategy called Asynchronous-SGD that allowed (Recht et al., 2011; Dean et al., 2012) to use multiple GPUs even with small batch sizes.<br />
<br />
== CONCLUSIONS ==<br />
Increasing the batch size during training has the same benefits of decaying the learning rate in addition to reducing the number of parameter updates, which corresponds to faster training time. Experiments were performed on different image datasets and various optimizers with different training schedules to prove this result. The paper proposed to increase the learning rate and momentum parameter <math>m</math>, while scaling <math> B \propto \frac{\epsilon}{1-m} </math>, which achieves fewer parameter updates, but slightly less test set accuracy as mentioned in detail in the experiments’ section. In summary, on ImageNet dataset, Inception-ResNet-V2 achieved 77% validation accuracy in under 2500 parameter updates, and ResNet-50 achieved 76.1% validation set accuracy on TPU in less than 30 minutes. One of the great finding of this paper is that all the methods use the hyper-parameters directly from previous works in the literature, and no additional hyper-parameter tuning was performed.<br />
<br />
== CRITIQUE ==<br />
'''Pros:'''<br />
<br />
- The paper showed empirically that increasing batch size and decaying learning rate are equivalent.<br />
<br />
- Several experiments were performed on different optimizers such as SGD and Adam.<br />
<br />
- Had several comparisons with previous experimental setups.<br />
<br />
'''Cons:'''<br />
<br />
<br />
- All datasets used are image datasets. Other experiments should have been done on datasets from different domains to ensure generalization. <br />
<br />
- The number of parameter updates was used as a comparison criterion, but wall-clock times could have provided additional measurable judgment although they depend on the hardware used.<br />
<br />
- Special hardware is needed for large batch training, which is not always feasible. As batch-size increases, we generally need more RAM to train the same model. However, if learning rate is decreased, the RAM use remains constant. As a result, learning rate decay will allow us to train bigger models.<br />
<br />
- In section 5.2 (Increasing the Effective Learning rate), the authors did not test a range of learning rate values and used only (0.1 and 0.5). Additional results from varying the initial learning rate values from 0.1 to 3.2 are provided in the appendix, which indicates that the test accuracy begins to fall for initial learning rates greater than ~0.4. The appended results do not show validation set accuracy curves like in Figure 6, however. It would be beneficial to see if they were similar to the original 0.1 and 0.5 initial learning rate baselines.<br />
<br />
- Although the main idea of the paper is interesting, its results does not seem to be too surprising in comparison with other recent papers in the subject.<br />
<br />
- The paper could benefit from using some other models to demonstrate its claim and generalize its idea by adding some comparisons with other models as well as other recent methods to increase batch size.<br />
<br />
- The paper presents interesting ideas. However, it lacks of mathematical and theoretical analysis beyond the idea. Since the experiment is primary on image dataset and it does not provide sufficient theories, the paper itself presents limited applicability to other types. <br />
<br />
- Also, in experimental setting, only single training runs from one random initialization is used. It would be better to take the best of many runs or to show confidence intervals.<br />
<br />
- It is proposed that we should compare learning rate decay with batch-size increase under the setting that total budget / number of training samples is fixed.<br />
<br />
== REFERENCES ==<br />
# Takuya Akiba, Shuji Suzuki, and Keisuke Fukuda. Extremely large minibatch sgd: Training resnet-50 on imagenet in 15 minutes. arXiv preprint arXiv:1711.04325, 2017.<br />
#Lukas Balles, Javier Romero, and Philipp Hennig. Coupling adaptive batch sizes with learning rates.arXiv preprint arXiv:1612.05086, 2016.<br />
#L´eon Bottou, Frank E Curtis, and Jorge Nocedal. Optimization methods for large-scale machine learning.arXiv preprint arXiv:1606.04838, 2016.<br />
#Richard H Byrd, Gillian M Chin, Jorge Nocedal, and Yuchen Wu. Sample size selection in optimization methods for machine learning. Mathematical programming, 134(1):127–155, 2012.<br />
#Pratik Chaudhari, Anna Choromanska, Stefano Soatto, and Yann LeCun. Entropy-SGD: Biasing gradient descent into wide valleys. arXiv preprint arXiv:1611.01838, 2016.<br />
#Soham De, Abhay Yadav, David Jacobs, and Tom Goldstein. Automated inference with adaptive batches. In Artificial Intelligence and Statistics, pp. 1504–1513, 2017.<br />
#Jeffrey Dean, Greg Corrado, Rajat Monga, Kai Chen, Matthieu Devin, Mark Mao, Andrew Senior, Paul Tucker, Ke Yang, Quoc V Le, et al. Large scale distributed deep networks. In Advances in neural information processing systems, pp. 1223–1231, 2012.<br />
#Michael P Friedlander and Mark Schmidt. Hybrid deterministic-stochastic methods for data fitting.SIAM Journal on Scientific Computing, 34(3):A1380–A1405, 2012.<br />
#Priya Goyal, Piotr Doll´ar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch SGD: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017.<br />
#Sepp Hochreiter and J¨urgen Schmidhuber. Flat minima. Neural Computation, 9(1):1–42, 1997.<br />
#Elad Hoffer, Itay Hubara, and Daniel Soudry. Train longer, generalize better: closing the generalization gap in large batch training of neural networks. arXiv preprint arXiv:1705.08741, 2017.<br />
#Norman P Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder Bajwa, Sarah Bates, Suresh Bhatia, Nan Boden, Al Borchers, et al. In-datacenter performance analysis of a tensor processing unit. In Proceedings of the 44th Annual International Symposium on Computer Architecture, pp. 1–12. ACM, 2017.<br />
#Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. arXiv preprint arXiv:1609.04836, 2016.<br />
#Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.<br />
#Alex Krizhevsky. One weird trick for parallelizing convolutional neural networks. arXiv preprint arXiv:1404.5997, 2014.<br />
#Qianxiao Li, Cheng Tai, and E Weinan. Stochastic modified equations and adaptive stochastic gradient algorithms. arXiv preprint arXiv:1511.06251, 2017.<br />
#Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with restarts. arXiv preprint arXiv:1608.03983, 2016.<br />
#Stephan Mandt, Matthew D Hoffman, and DavidMBlei. Stochastic gradient descent as approximate bayesian inference. arXiv preprint arXiv:1704.04289, 2017.<br />
#James Martens and Roger Grosse. Optimizing neural networks with kronecker-factored approximate curvature. In International Conference on Machine Learning, pp. 2408–2417, 2015.<br />
#Yurii Nesterov. A method of solving a convex programming problem with convergence rate o (1/k2). In Soviet Mathematics Doklady, volume 27, pp. 372–376, 1983.<br />
#Lutz Prechelt. Early stopping-but when? Neural Networks: Tricks of the trade, pp. 553–553, 1998.<br />
#Benjamin Recht, Christopher Re, Stephen Wright, and Feng Niu. Hogwild: A lock-free approach to parallelizing stochastic gradient descent. In Advances in neural information processing systems, pp. 693–701, 2011.<br />
#Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pp. 400–407, 1951.<br />
#Samuel L. Smith and Quoc V. Le. A bayesian perspective on generalization and stochastic gradient descent. arXiv preprint arXiv:1710.06451, 2017.<br />
#Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A Alemi. Inception-v4, Inception-ResNet and the impact of residual connections on learning. In AAAI, pp. 4278–4284, 2017.<br />
#Max Welling and Yee W Teh. Bayesian learning via stochastic gradient langevin dynamics. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pp. 681–688, 2011.<br />
#Ashia C Wilson, Rebecca Roelofs, Mitchell Stern, Nathan Srebro, and Benjamin Recht. The marginal value of adaptive gradient methods in machine learning. arXiv preprint arXiv:1705.08292, 2017.<br />
#Yang You, Igor Gitman, and Boris Ginsburg. Scaling SGD batch size to 32k for imagenet training. arXiv preprint arXiv:1708.03888, 2017a.<br />
#Yang You, Zhao Zhang, C Hsieh, James Demmel, and Kurt Keutzer. Imagenet training in minutes. CoRR, abs/1709.05011, 2017b.<br />
#Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016.<br />
#Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. arXiv preprint arXiv:1611.03530, 2016.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=DON%27T_DECAY_THE_LEARNING_RATE_,_INCREASE_THE_BATCH_SIZE&diff=42073DON'T DECAY THE LEARNING RATE , INCREASE THE BATCH SIZE2018-11-30T16:23:27Z<p>Ka2khan: /* INTRODUCTION */</p>
<hr />
<div>Summary of the ICLR 2018 paper: '''Don't Decay the learning Rate, Increase the Batch Size ''' <br />
<br />
Link: [https://arxiv.org/pdf/1711.00489.pdf]<br />
<br />
Summarized by: Afify, Ahmed [ID: 20700841]<br />
<br />
==INTUITION==<br />
Nowadays, it is a common practice not to have a singular steady learning rate for the learning phase of neural network models. Instead, we use adaptive learning rates with the standard gradient descent method. The intuition behind this is that when we are far away from the minima, it is beneficial for us to take large steps towards the minima, as it would require a lesser number of steps to converge, but as we approach the minima, our step size should decrease, otherwise we may just keep oscillating around the minima. In practice, this is generally achieved by methods like SGD with momentum, Nesterov momentum, and Adam. However, the core claim of this paper is that the same effect can be achieved by increasing the batch size during the gradient descent process while keeping the learning rate constant throughout. In addition, the paper argues that such an approach also reduces the parameter updates required to reach the minima, thus leading to greater parallelism and shorter training times.<br />
<br />
== INTRODUCTION ==<br />
Stochastic gradient descent (SGD) is the most widely used optimization technique for training deep learning models. The reason for this is that the minima found using this process generalizes well (Zhang et al., 2016; Wilson et al., 2017), but the optimization process is slow and time consuming as each parameter update corresponds to a small step towards the gooal. According to (Goyal et al., 2017; Hoffer et al., 2017; You et al., 2017a), this has motivated researchers to try to speed up this optimization process by taking bigger steps, and hence reduce the number of parameter updates in training a model. This can be achieved by using large batch training, which can be divided across many machines. <br />
<br />
However, increasing the batch size leads to decreasing the test set accuracy (Keskar et al., 2016; Goyal et al., 2017). Smith and Le (2017) believed that SGD has a scale of random fluctuations <math> g = \epsilon (\frac{N}{B}-1) </math>, where <math> \epsilon </math> is the learning rate, N number of training samples, and B batch size. They concluded that there is an optimal batch size proportional to the learning rate when <math> B \ll N </math>, and optimum fluctuation scale <math>g</math> at constant learning rate which maximizes test set accuracy. This observation was empirically by Goyal et al., 2017 and used to train a ResNet-50 in under an hour with 76.3% validation accuracy on ImageNet dataset.<br />
<br />
In this paper, the authors' main goal is to provide evidence that increasing the batch size is quantitatively equivalent to decreasing the learning rate with the same number of training epochs in decreasing the scale of random fluctuations, but with remarkably less number of parameter updates. Moreover, an additional reduction in the number of parameter updates can be attained by increasing the learning rate and scaling <math> B \propto \epsilon </math> or even more reduction by increasing the momentum coefficient and scaling <math> B \propto \frac{1}{1-m} </math> although the latter decreases the test accuracy. This has been demonstrated by several experiments on the ImageNet and CIFAR-10 datasets using ResNet-50 and Inception-ResNet-V2 architectures respectively.<br />
<br />
== STOCHASTIC GRADIENT DESCENT AND CONVEX OPTIMIZATION ==<br />
As mentioned in the previous section, the drawback of SGD when compared to full-batch training is the noise that it introduces that hinders optimization. According to (Robbins & Monro, 1951), there are two equations that govern how to reach the minimum of a convex function: (<math> \epsilon_i </math> denotes the learning rate at the <math> i^{th} </math> gradient update)<br />
<br />
<math> \sum_{i=1}^{\infty} \epsilon_i = \infty </math>. This equation guarantees that we will reach the minimum. <br />
<br />
<math> \sum_{i=1}^{\infty} \epsilon^2_i < \infty </math>. This equation, which is valid only for a fixed batch size, guarantees that learning rate decays fast enough allowing us to reach the minimum rather than bouncing due to noise.<br />
<br />
These equations indicate that the learning rate must decay during training, and second equation is only available when the batch size is constant. To change the batch size, Smith and Le (2017) proposed to interpret SGD as integrating this stochastic differential equation <math> \frac{dw}{dt} = -\frac{dC}{dw} + \eta(t) </math>, where <math>C</math> represents cost function, <math>w</math> represents the parameters, and <math>\eta</math> represents the Gaussian random noise. Furthermore, they proved that noise scale <math>g</math> controls the magnitude of random fluctuations in the training dynamics by this formula: <math> g = \epsilon (\frac{N}{B}-1) </math>, where <math> \epsilon </math> is the learning rate, N is the training set size and <math>B</math> is the batch size. As we usually have <math> B \ll N </math>, we can define <math> g \approx \epsilon \frac{N}{B} </math>. This explains why when the learning rate decreases, noise <math>g</math> decreases, enabling us to converge to the minimum of the cost function. However, increasing the batch size has the same effect and makes <math>g</math> decays with constant learning rate. In this work, the batch size is increased until <math> B \approx \frac{N}{10} </math>, then the conventional way of decaying the learning rate is followed.<br />
<br />
== SIMULATED ANNEALING AND THE GENERALIZATION GAP ==<br />
'''Simulated Annealing:''' decaying learning rates are empirically successful. To understand this, they note that introducing random fluctuations<br />
whose scale falls during training is also a well established technique in non-convex optimization; simulated annealing. The initial noisy optimization phase allows to explore a larger fraction of the parameter space without becoming trapped in local minima. Once a promising region of parameter space is located, the noise is reduced to fine-tune the parameters.<br />
<br />
For more info: Simulated annealing (SA) is a probabilistic technique for approximating the global optimum of a given function. Specifically, it is a metaheuristic to approximate global optimization in a large search space for an optimization problem. It is often used when the search space is discrete (e.g., all tours that visit a given set of cities). For problems where finding an approximate global optimum is more important than finding a precise local optimum in a fixed amount of time, simulated annealing may be preferable to alternatives such as gradient descent. [https://en.wikipedia.org/wiki/Simulated_annealing [Reference]]<br />
<br />
'''Generalization Gap:''' Small batch data generalizes better to the test set than large batch data.<br />
<br />
Smith and Le (2017) found that there is an optimal batch size which corresponds to optimal noise scale g <math> (g \approx \epsilon \frac{N}{B}) </math> and concluded that <math> B_{opt} \propto \epsilon N </math> that corresponds to maximum test set accuracy. This means that gradient noise is helpful as it makes SGD escape sharp minima, which does not generalize well. <br />
<br />
Simulated Annealing is a famous technique in non-convex optimization. Starting with noise in the training process helps us to discover a wide range of parameters then once we are near the optimum value, noise is reduced to fine tune our final parameters. However, more and more researches like to use the sharper decay schedules like cosine decay or step-function drops. In physical sciences, slowly annealing (or decaying) the temperature (which is the noise scale in this situation) helps to converge to the global minimum, which is sharp. But decaying the temperature in discrete steps can make the system stuck in a local minimum, which lead to higher cost and lower curvature. The authors think that deep learning has the same intuition.<br />
.<br />
<br />
== THE EFFECTIVE LEARNING RATE AND THE ACCUMULATION VARIABLE ==<br />
'''The Effective Learning Rate''' : <math> \epsilon_{eff} = \frac{\epsilon}{1-m} </math><br />
<br />
Smith and Le (2017) included momentum to the equation of the vanilla SGD noise scale that was defined above to be: <math> g = \frac{\epsilon}{1-m}(\frac{N}{B}-1)\approx \frac{\epsilon N}{B(1-m)} </math>, which is the same as the previous equation when m goes to 0. They found that increasing the learning rate and momentum coefficient and scaling <math> B \propto \frac{\epsilon }{1-m} </math> reduces the number of parameter updates, but the test accuracy decreases when the momentum coefficient is increased. <br />
<br />
To understand the reasons behind this, we need to analyze momentum update equations below:<br />
<br />
<center><math><br />
\Delta A = -(1-m)A + \frac{d\widehat{C}}{dw} <br />
</math><br />
<br />
<math><br />
\Delta w = -A\epsilon<br />
</math><br />
</center><br />
<br />
We can see that the Accumulation variable A, which is initially set to 0, then increases exponentially to reach its steady state value during <math> \frac{B}{N(1-m)} </math> training epochs while <math> \Delta w </math> is suppressed that can reduce the rate of convergence. Moreover, at high momentum, we have three challenges:<br />
<br />
1- Additional epochs are needed to catch up with the accumulation.<br />
<br />
2- Accumulation needs more time <math> \frac{B}{N(1-m)} </math> to forget old gradients. <br />
<br />
3- After this time, however, the accumulation cannot adapt to changes in the loss landscape.<br />
<br />
4- In the early stage, large batch size will lead to the instabilities.<br />
<br />
== EXPERIMENTS ==<br />
=== SIMULATED ANNEALING IN A WIDE RESNET ===<br />
<br />
'''Dataset:''' CIFAR-10 (50,000 training images)<br />
<br />
'''Network Architecture:''' “16-4” wide ResNet<br />
<br />
'''Training Schedules used as in the below figure:''' . These demonstrate the equivalence between decreasing the learning rate and increasing the batch size.<br />
<br />
- Decaying learning rate: learning rate decays by a factor of 5 at a sequence of “steps”, and the batch size is constant<br />
<br />
- Increasing batch size: learning rate is constant, and the batch size is increased by a factor of 5 at every step.<br />
<br />
- Hybrid: At the beginning, the learning rate is constant and batch size is increased by a factor of 5. Then, the learning rate decays by a factor of 5 at each subsequent step, and the batch size is constant. This is the schedule that will be used if there is a hardware limit affecting a maximum batch size limit.<br />
<br />
If the learning rate itself must decay during training, then these schedules should show different learning curves (as a function of the number of training epochs) and reach different final test set accuracies. Meanwhile if it is the noise scale which should decay, all three schedules should be indistinguishable.<br />
[[File:Paper_40_Fig_1.png | 800px|center]]<br />
<br />
As shown in the below figure: in the left figure (2a), we can observe that for the training set, the three learning curves are exactly the same while in figure 2b, increasing the batch size has a huge advantage of reducing the number of parameter updates.<br />
This concludes that noise scale is the one that needs to be decayed and not the learning rate itself<br />
[[File:Paper_40_Fig_2.png | 800px|center]] <br />
<br />
To make sure that these results are the same for the test set as well, in figure 3, we can see that the three learning curves are exactly the same for SGD with momentum, and Nesterov momentum<br />
[[File:Paper_40_Fig_3.png | 800px|center]]<br />
<br />
To check for other optimizers as well. the below figure shows the same experiment as in figure 3, which is the three learning curves for test set, but for vanilla SGD and Adam, and showing <br />
[[File:Paper_40_Fig_4.png | 800px|center]]<br />
<br />
'''Conclusion:''' Decreasing the learning rate and increasing the batch size during training are equivalent<br />
<br />
=== INCREASING THE EFFECTIVE LEARNING RATE===<br />
<br />
Here, the focus is on minimizing the number of parameter updates required to train a model. As shown above, the first step is to replace decaying learning rates by increasing batch sizes. Now, the authors show here that we can also increase the effective learning rate <math>\epsilon_{eff} = \epsilon/(1 − m) </math> at the start of training, while scaling the initial batch size <math>B \propto \epsilon_{eff} </math> . All experiments are conducted using SGD with momentum. There are 50000 images in the CIFAR-10 training set, and since the scaling rules only hold when <math>B << N </math> , we decided to set a maximum batch size <math>B_{max} </math>= 5120 .<br />
<br />
'''Dataset:''' CIFAR-10 (50,000 training images)<br />
<br />
'''Network Architecture:''' “16-4” wide ResNet<br />
<br />
'''Training Parameters:''' Optimization Algorithm: SGD with momentum / Maximum batch size = 5120<br />
<br />
'''Training Schedules:''' <br />
<br />
The authors consider four training schedules, all of which decay the noise scale by a factor of five in a series of three steps with the same number of epochs.<br />
<br />
Original training schedule: initial learning rate of 0.1 which decays by a factor of 5 at each step, a momentum coefficient of 0.9, and a batch size of 128. Follows the implementation of Zagoruyko & Komodakis (2016).<br />
<br />
Increasing batch size: learning rate of 0.1, momentum coefficient of 0.9, initial batch size of 128 that increases by a factor of 5 at each step. <br />
<br />
Increased initial learning rate: initial learning rate of 0.5, initial batch size of 640 that increase during training.<br />
<br />
Increased momentum coefficient: increased initial learning rate of 0.5, initial batch size of 3200 that increase during training, and an increased momentum coefficient of 0.98.<br />
<br />
The results of all training schedules, which are presented in the below figure, are documented in the following table:<br />
<br />
[[File:Paper_40_Table_1.png | 800px|center]]<br />
<br />
[[File:Paper_40_Fig_5.png | 800px|center]]<br />
<br />
<br />
<br />
'''Conclusion:''' Increasing the effective learning rate and scaling the batch size results in further reduction in the number of parameter updates<br />
<br />
=== TRAINING IMAGENET IN 2500 PARAMETER UPDATES===<br />
<br />
'''A) Experiment Goal:''' Control Batch Size<br />
<br />
'''Dataset:''' ImageNet (1.28 million training images)<br />
<br />
The paper modified the setup of Goyal et al. (2017), and used the following configuration:<br />
<br />
'''Network Architecture:''' Inception-ResNet-V2 <br />
<br />
'''Training Parameters:''' <br />
<br />
90 epochs / noise decayed at epoch 30, 60, and 80 by a factor of 10 / Initial ghost batch size = 32 / Learning rate = 3 / momentum coefficient = 0.9 / Initial batch size = 8192<br />
<br />
Two training schedules were used:<br />
<br />
“Decaying learning rate”, where batch size is fixed and the learning rate is decayed<br />
<br />
“Increasing batch size”, where batch size is increased to 81920 then the learning rate is decayed at two steps.<br />
<br />
[[File:Paper_40_Table_2.png | 800px|center]]<br />
<br />
[[File:Paper_40_Fig_6.png | 800px|center]]<br />
<br />
'''Conclusion:''' Increasing the batch size resulted in reducing the number of parameter updates from 14,000 to 6,000.<br />
<br />
'''B) Experiment Goal:''' Control Batch Size and Momentum Coefficient<br />
<br />
'''Training Parameters:''' Ghost batch size = 64 / noise decayed at epoch 30, 60, and 80 by a factor of 10. <br />
<br />
The below table shows the number of parameter updates and accuracy for different set of training parameters:<br />
<br />
[[File:Paper_40_Table_3.png | 800px|center]]<br />
<br />
[[File:Paper_40_Fig_7.png | 800px|center]]<br />
<br />
'''Conclusion:''' Increasing the momentum reduces the number of parameter updates, but leads to a drop in the test accuracy.<br />
<br />
=== TRAINING IMAGENET IN 30 MINUTES===<br />
<br />
'''Dataset:''' ImageNet (Already introduced in the previous section)<br />
<br />
'''Network Architecture:''' ResNet-50<br />
<br />
The paper replicated the setup of Goyal et al. (2017) while modifying the number of TPU devices, batch size, learning rate, and then calculating the time to complete 90 epochs, and measuring the accuracy, and performed the following experiments below:<br />
<br />
[[File:Paper_40_Table_4.png | 800px|center]]<br />
<br />
'''Conclusion:''' Model training times can be reduced by increasing the batch size during training.<br />
<br />
== RELATED WORK ==<br />
Main related work mentioned in the paper is as follows:<br />
<br />
- Smith & Le (2017) interpreted Stochastic gradient descent as stochastic differential equation; the paper built on this idea to include decaying learning rate.<br />
<br />
- Mandt et al. (2017) analyzed how to modify SGD for the task of Bayesian posterior sampling.<br />
<br />
- Keskar et al. (2016) focused on the analysis of noise once the training is started.<br />
<br />
- Moreover, the proportional relationship between batch size and learning rate was first discovered by Goyal et al. (2017) and successfully trained ResNet-50 on ImageNet in one hour after discovering the proportionality relationship between batch size and learning rate.<br />
<br />
- Furthermore, You et al. (2017a) presented Layer-wise Adaptive Rate Scaling (LARS), which is applying different learning rates to train ImageNet in 14 minutes and 74.9% accuracy. <br />
<br />
- Wilson et al. (2017) argued that adaptive optimization methods tend to generalize less well than SGD and SGD with momentum (although<br />
they did not include K-FAC in their study), while the authors' work reduces the gap in convergence speed.<br />
<br />
- Finally, another strategy called Asynchronous-SGD that allowed (Recht et al., 2011; Dean et al., 2012) to use multiple GPUs even with small batch sizes.<br />
<br />
== CONCLUSIONS ==<br />
Increasing the batch size during training has the same benefits of decaying the learning rate in addition to reducing the number of parameter updates, which corresponds to faster training time. Experiments were performed on different image datasets and various optimizers with different training schedules to prove this result. The paper proposed to increase the learning rate and momentum parameter <math>m</math>, while scaling <math> B \propto \frac{\epsilon}{1-m} </math>, which achieves fewer parameter updates, but slightly less test set accuracy as mentioned in detail in the experiments’ section. In summary, on ImageNet dataset, Inception-ResNet-V2 achieved 77% validation accuracy in under 2500 parameter updates, and ResNet-50 achieved 76.1% validation set accuracy on TPU in less than 30 minutes. One of the great finding of this paper is that all the methods use the hyper-parameters directly from previous works in the literature, and no additional hyper-parameter tuning was performed.<br />
<br />
== CRITIQUE ==<br />
'''Pros:'''<br />
<br />
- The paper showed empirically that increasing batch size and decaying learning rate are equivalent.<br />
<br />
- Several experiments were performed on different optimizers such as SGD and Adam.<br />
<br />
- Had several comparisons with previous experimental setups.<br />
<br />
'''Cons:'''<br />
<br />
<br />
- All datasets used are image datasets. Other experiments should have been done on datasets from different domains to ensure generalization. <br />
<br />
- The number of parameter updates was used as a comparison criterion, but wall-clock times could have provided additional measurable judgment although they depend on the hardware used.<br />
<br />
- Special hardware is needed for large batch training, which is not always feasible. As batch-size increases, we generally need more RAM to train the same model. However, if learning rate is decreased, the RAM use remains constant. As a result, learning rate decay will allow us to train bigger models.<br />
<br />
- In section 5.2 (Increasing the Effective Learning rate), the authors did not test a range of learning rate values and used only (0.1 and 0.5). Additional results from varying the initial learning rate values from 0.1 to 3.2 are provided in the appendix, which indicates that the test accuracy begins to fall for initial learning rates greater than ~0.4. The appended results do not show validation set accuracy curves like in Figure 6, however. It would be beneficial to see if they were similar to the original 0.1 and 0.5 initial learning rate baselines.<br />
<br />
- Although the main idea of the paper is interesting, its results does not seem to be too surprising in comparison with other recent papers in the subject.<br />
<br />
- The paper could benefit from using some other models to demonstrate its claim and generalize its idea by adding some comparisons with other models as well as other recent methods to increase batch size.<br />
<br />
- The paper presents interesting ideas. However, it lacks of mathematical and theoretical analysis beyond the idea. Since the experiment is primary on image dataset and it does not provide sufficient theories, the paper itself presents limited applicability to other types. <br />
<br />
- Also, in experimental setting, only single training runs from one random initialization is used. It would be better to take the best of many runs or to show confidence intervals.<br />
<br />
- It is proposed that we should compare learning rate decay with batch-size increase under the setting that total budget / number of training samples is fixed.<br />
<br />
== REFERENCES ==<br />
# Takuya Akiba, Shuji Suzuki, and Keisuke Fukuda. Extremely large minibatch sgd: Training resnet-50 on imagenet in 15 minutes. arXiv preprint arXiv:1711.04325, 2017.<br />
#Lukas Balles, Javier Romero, and Philipp Hennig. Coupling adaptive batch sizes with learning rates.arXiv preprint arXiv:1612.05086, 2016.<br />
#L´eon Bottou, Frank E Curtis, and Jorge Nocedal. Optimization methods for large-scale machine learning.arXiv preprint arXiv:1606.04838, 2016.<br />
#Richard H Byrd, Gillian M Chin, Jorge Nocedal, and Yuchen Wu. Sample size selection in optimization methods for machine learning. Mathematical programming, 134(1):127–155, 2012.<br />
#Pratik Chaudhari, Anna Choromanska, Stefano Soatto, and Yann LeCun. Entropy-SGD: Biasing gradient descent into wide valleys. arXiv preprint arXiv:1611.01838, 2016.<br />
#Soham De, Abhay Yadav, David Jacobs, and Tom Goldstein. Automated inference with adaptive batches. In Artificial Intelligence and Statistics, pp. 1504–1513, 2017.<br />
#Jeffrey Dean, Greg Corrado, Rajat Monga, Kai Chen, Matthieu Devin, Mark Mao, Andrew Senior, Paul Tucker, Ke Yang, Quoc V Le, et al. Large scale distributed deep networks. In Advances in neural information processing systems, pp. 1223–1231, 2012.<br />
#Michael P Friedlander and Mark Schmidt. Hybrid deterministic-stochastic methods for data fitting.SIAM Journal on Scientific Computing, 34(3):A1380–A1405, 2012.<br />
#Priya Goyal, Piotr Doll´ar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch SGD: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017.<br />
#Sepp Hochreiter and J¨urgen Schmidhuber. Flat minima. Neural Computation, 9(1):1–42, 1997.<br />
#Elad Hoffer, Itay Hubara, and Daniel Soudry. Train longer, generalize better: closing the generalization gap in large batch training of neural networks. arXiv preprint arXiv:1705.08741, 2017.<br />
#Norman P Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder Bajwa, Sarah Bates, Suresh Bhatia, Nan Boden, Al Borchers, et al. In-datacenter performance analysis of a tensor processing unit. In Proceedings of the 44th Annual International Symposium on Computer Architecture, pp. 1–12. ACM, 2017.<br />
#Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. arXiv preprint arXiv:1609.04836, 2016.<br />
#Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.<br />
#Alex Krizhevsky. One weird trick for parallelizing convolutional neural networks. arXiv preprint arXiv:1404.5997, 2014.<br />
#Qianxiao Li, Cheng Tai, and E Weinan. Stochastic modified equations and adaptive stochastic gradient algorithms. arXiv preprint arXiv:1511.06251, 2017.<br />
#Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with restarts. arXiv preprint arXiv:1608.03983, 2016.<br />
#Stephan Mandt, Matthew D Hoffman, and DavidMBlei. Stochastic gradient descent as approximate bayesian inference. arXiv preprint arXiv:1704.04289, 2017.<br />
#James Martens and Roger Grosse. Optimizing neural networks with kronecker-factored approximate curvature. In International Conference on Machine Learning, pp. 2408–2417, 2015.<br />
#Yurii Nesterov. A method of solving a convex programming problem with convergence rate o (1/k2). In Soviet Mathematics Doklady, volume 27, pp. 372–376, 1983.<br />
#Lutz Prechelt. Early stopping-but when? Neural Networks: Tricks of the trade, pp. 553–553, 1998.<br />
#Benjamin Recht, Christopher Re, Stephen Wright, and Feng Niu. Hogwild: A lock-free approach to parallelizing stochastic gradient descent. In Advances in neural information processing systems, pp. 693–701, 2011.<br />
#Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pp. 400–407, 1951.<br />
#Samuel L. Smith and Quoc V. Le. A bayesian perspective on generalization and stochastic gradient descent. arXiv preprint arXiv:1710.06451, 2017.<br />
#Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A Alemi. Inception-v4, Inception-ResNet and the impact of residual connections on learning. In AAAI, pp. 4278–4284, 2017.<br />
#Max Welling and Yee W Teh. Bayesian learning via stochastic gradient langevin dynamics. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pp. 681–688, 2011.<br />
#Ashia C Wilson, Rebecca Roelofs, Mitchell Stern, Nathan Srebro, and Benjamin Recht. The marginal value of adaptive gradient methods in machine learning. arXiv preprint arXiv:1705.08292, 2017.<br />
#Yang You, Igor Gitman, and Boris Ginsburg. Scaling SGD batch size to 32k for imagenet training. arXiv preprint arXiv:1708.03888, 2017a.<br />
#Yang You, Zhao Zhang, C Hsieh, James Demmel, and Kurt Keutzer. Imagenet training in minutes. CoRR, abs/1709.05011, 2017b.<br />
#Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016.<br />
#Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. arXiv preprint arXiv:1611.03530, 2016.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=DON%27T_DECAY_THE_LEARNING_RATE_,_INCREASE_THE_BATCH_SIZE&diff=42072DON'T DECAY THE LEARNING RATE , INCREASE THE BATCH SIZE2018-11-30T16:20:50Z<p>Ka2khan: /* INTRODUCTION */</p>
<hr />
<div>Summary of the ICLR 2018 paper: '''Don't Decay the learning Rate, Increase the Batch Size ''' <br />
<br />
Link: [https://arxiv.org/pdf/1711.00489.pdf]<br />
<br />
Summarized by: Afify, Ahmed [ID: 20700841]<br />
<br />
==INTUITION==<br />
Nowadays, it is a common practice not to have a singular steady learning rate for the learning phase of neural network models. Instead, we use adaptive learning rates with the standard gradient descent method. The intuition behind this is that when we are far away from the minima, it is beneficial for us to take large steps towards the minima, as it would require a lesser number of steps to converge, but as we approach the minima, our step size should decrease, otherwise we may just keep oscillating around the minima. In practice, this is generally achieved by methods like SGD with momentum, Nesterov momentum, and Adam. However, the core claim of this paper is that the same effect can be achieved by increasing the batch size during the gradient descent process while keeping the learning rate constant throughout. In addition, the paper argues that such an approach also reduces the parameter updates required to reach the minima, thus leading to greater parallelism and shorter training times.<br />
<br />
== INTRODUCTION ==<br />
Stochastic gradient descent (SGD) is the most widely used optimization technique for training deep learning models. The reason for this is that the minima found using this process generalizes well (Zhang et al., 2016; Wilson et al., 2017), but the optimization process is slow and time consuming as each parameter update corresponds to a small step towards the gooal. According to (Goyal et al., 2017; Hoffer et al., 2017; You et al., 2017a), this has motivated researchers to try to speed up this optimization process by taking bigger steps, and hence reduce the number of parameter updates in training a model. This can be achieved by using large batch training, which can be divided across many machines. <br />
<br />
However, increasing the batch size leads to decreasing the test set accuracy (Keskar et al., 2016; Goyal et al., 2017). Smith and Le (2017) believed that SGD has a scale of random fluctuations <math> g = \epsilon (\frac{N}{B}-1) </math>, where <math> \epsilon </math> is the learning rate, N number of training samples, and B batch size. They concluded that there is an optimal batch size proportional to the learning rate when <math> B \ll N </math>, and optimum fluctuation scale <math>g</math> at constant learning rate which maximizes test set accuracy.<br />
<br />
In this paper, the authors' main goal is to provide evidence that increasing the batch size is quantitatively equivalent to decreasing the learning rate with the same number of training epochs in decreasing the scale of random fluctuations, but with remarkably less number of parameter updates. Moreover, an additional reduction in the number of parameter updates can be attained by increasing the learning rate and scaling <math> B \propto \epsilon </math> or even more reduction by increasing the momentum coefficient and scaling <math> B \propto \frac{1}{1-m} </math> although the latter decreases the test accuracy. This has been demonstrated by several experiments on the ImageNet and CIFAR-10 datasets using ResNet-50 and Inception-ResNet-V2 architectures respectively.<br />
<br />
== STOCHASTIC GRADIENT DESCENT AND CONVEX OPTIMIZATION ==<br />
As mentioned in the previous section, the drawback of SGD when compared to full-batch training is the noise that it introduces that hinders optimization. According to (Robbins & Monro, 1951), there are two equations that govern how to reach the minimum of a convex function: (<math> \epsilon_i </math> denotes the learning rate at the <math> i^{th} </math> gradient update)<br />
<br />
<math> \sum_{i=1}^{\infty} \epsilon_i = \infty </math>. This equation guarantees that we will reach the minimum. <br />
<br />
<math> \sum_{i=1}^{\infty} \epsilon^2_i < \infty </math>. This equation, which is valid only for a fixed batch size, guarantees that learning rate decays fast enough allowing us to reach the minimum rather than bouncing due to noise.<br />
<br />
These equations indicate that the learning rate must decay during training, and second equation is only available when the batch size is constant. To change the batch size, Smith and Le (2017) proposed to interpret SGD as integrating this stochastic differential equation <math> \frac{dw}{dt} = -\frac{dC}{dw} + \eta(t) </math>, where <math>C</math> represents cost function, <math>w</math> represents the parameters, and <math>\eta</math> represents the Gaussian random noise. Furthermore, they proved that noise scale <math>g</math> controls the magnitude of random fluctuations in the training dynamics by this formula: <math> g = \epsilon (\frac{N}{B}-1) </math>, where <math> \epsilon </math> is the learning rate, N is the training set size and <math>B</math> is the batch size. As we usually have <math> B \ll N </math>, we can define <math> g \approx \epsilon \frac{N}{B} </math>. This explains why when the learning rate decreases, noise <math>g</math> decreases, enabling us to converge to the minimum of the cost function. However, increasing the batch size has the same effect and makes <math>g</math> decays with constant learning rate. In this work, the batch size is increased until <math> B \approx \frac{N}{10} </math>, then the conventional way of decaying the learning rate is followed.<br />
<br />
== SIMULATED ANNEALING AND THE GENERALIZATION GAP ==<br />
'''Simulated Annealing:''' decaying learning rates are empirically successful. To understand this, they note that introducing random fluctuations<br />
whose scale falls during training is also a well established technique in non-convex optimization; simulated annealing. The initial noisy optimization phase allows to explore a larger fraction of the parameter space without becoming trapped in local minima. Once a promising region of parameter space is located, the noise is reduced to fine-tune the parameters.<br />
<br />
For more info: Simulated annealing (SA) is a probabilistic technique for approximating the global optimum of a given function. Specifically, it is a metaheuristic to approximate global optimization in a large search space for an optimization problem. It is often used when the search space is discrete (e.g., all tours that visit a given set of cities). For problems where finding an approximate global optimum is more important than finding a precise local optimum in a fixed amount of time, simulated annealing may be preferable to alternatives such as gradient descent. [https://en.wikipedia.org/wiki/Simulated_annealing [Reference]]<br />
<br />
'''Generalization Gap:''' Small batch data generalizes better to the test set than large batch data.<br />
<br />
Smith and Le (2017) found that there is an optimal batch size which corresponds to optimal noise scale g <math> (g \approx \epsilon \frac{N}{B}) </math> and concluded that <math> B_{opt} \propto \epsilon N </math> that corresponds to maximum test set accuracy. This means that gradient noise is helpful as it makes SGD escape sharp minima, which does not generalize well. <br />
<br />
Simulated Annealing is a famous technique in non-convex optimization. Starting with noise in the training process helps us to discover a wide range of parameters then once we are near the optimum value, noise is reduced to fine tune our final parameters. However, more and more researches like to use the sharper decay schedules like cosine decay or step-function drops. In physical sciences, slowly annealing (or decaying) the temperature (which is the noise scale in this situation) helps to converge to the global minimum, which is sharp. But decaying the temperature in discrete steps can make the system stuck in a local minimum, which lead to higher cost and lower curvature. The authors think that deep learning has the same intuition.<br />
.<br />
<br />
== THE EFFECTIVE LEARNING RATE AND THE ACCUMULATION VARIABLE ==<br />
'''The Effective Learning Rate''' : <math> \epsilon_{eff} = \frac{\epsilon}{1-m} </math><br />
<br />
Smith and Le (2017) included momentum to the equation of the vanilla SGD noise scale that was defined above to be: <math> g = \frac{\epsilon}{1-m}(\frac{N}{B}-1)\approx \frac{\epsilon N}{B(1-m)} </math>, which is the same as the previous equation when m goes to 0. They found that increasing the learning rate and momentum coefficient and scaling <math> B \propto \frac{\epsilon }{1-m} </math> reduces the number of parameter updates, but the test accuracy decreases when the momentum coefficient is increased. <br />
<br />
To understand the reasons behind this, we need to analyze momentum update equations below:<br />
<br />
<center><math><br />
\Delta A = -(1-m)A + \frac{d\widehat{C}}{dw} <br />
</math><br />
<br />
<math><br />
\Delta w = -A\epsilon<br />
</math><br />
</center><br />
<br />
We can see that the Accumulation variable A, which is initially set to 0, then increases exponentially to reach its steady state value during <math> \frac{B}{N(1-m)} </math> training epochs while <math> \Delta w </math> is suppressed that can reduce the rate of convergence. Moreover, at high momentum, we have three challenges:<br />
<br />
1- Additional epochs are needed to catch up with the accumulation.<br />
<br />
2- Accumulation needs more time <math> \frac{B}{N(1-m)} </math> to forget old gradients. <br />
<br />
3- After this time, however, the accumulation cannot adapt to changes in the loss landscape.<br />
<br />
4- In the early stage, large batch size will lead to the instabilities.<br />
<br />
== EXPERIMENTS ==<br />
=== SIMULATED ANNEALING IN A WIDE RESNET ===<br />
<br />
'''Dataset:''' CIFAR-10 (50,000 training images)<br />
<br />
'''Network Architecture:''' “16-4” wide ResNet<br />
<br />
'''Training Schedules used as in the below figure:''' . These demonstrate the equivalence between decreasing the learning rate and increasing the batch size.<br />
<br />
- Decaying learning rate: learning rate decays by a factor of 5 at a sequence of “steps”, and the batch size is constant<br />
<br />
- Increasing batch size: learning rate is constant, and the batch size is increased by a factor of 5 at every step.<br />
<br />
- Hybrid: At the beginning, the learning rate is constant and batch size is increased by a factor of 5. Then, the learning rate decays by a factor of 5 at each subsequent step, and the batch size is constant. This is the schedule that will be used if there is a hardware limit affecting a maximum batch size limit.<br />
<br />
If the learning rate itself must decay during training, then these schedules should show different learning curves (as a function of the number of training epochs) and reach different final test set accuracies. Meanwhile if it is the noise scale which should decay, all three schedules should be indistinguishable.<br />
[[File:Paper_40_Fig_1.png | 800px|center]]<br />
<br />
As shown in the below figure: in the left figure (2a), we can observe that for the training set, the three learning curves are exactly the same while in figure 2b, increasing the batch size has a huge advantage of reducing the number of parameter updates.<br />
This concludes that noise scale is the one that needs to be decayed and not the learning rate itself<br />
[[File:Paper_40_Fig_2.png | 800px|center]] <br />
<br />
To make sure that these results are the same for the test set as well, in figure 3, we can see that the three learning curves are exactly the same for SGD with momentum, and Nesterov momentum<br />
[[File:Paper_40_Fig_3.png | 800px|center]]<br />
<br />
To check for other optimizers as well. the below figure shows the same experiment as in figure 3, which is the three learning curves for test set, but for vanilla SGD and Adam, and showing <br />
[[File:Paper_40_Fig_4.png | 800px|center]]<br />
<br />
'''Conclusion:''' Decreasing the learning rate and increasing the batch size during training are equivalent<br />
<br />
=== INCREASING THE EFFECTIVE LEARNING RATE===<br />
<br />
Here, the focus is on minimizing the number of parameter updates required to train a model. As shown above, the first step is to replace decaying learning rates by increasing batch sizes. Now, the authors show here that we can also increase the effective learning rate <math>\epsilon_{eff} = \epsilon/(1 − m) </math> at the start of training, while scaling the initial batch size <math>B \propto \epsilon_{eff} </math> . All experiments are conducted using SGD with momentum. There are 50000 images in the CIFAR-10 training set, and since the scaling rules only hold when <math>B << N </math> , we decided to set a maximum batch size <math>B_{max} </math>= 5120 .<br />
<br />
'''Dataset:''' CIFAR-10 (50,000 training images)<br />
<br />
'''Network Architecture:''' “16-4” wide ResNet<br />
<br />
'''Training Parameters:''' Optimization Algorithm: SGD with momentum / Maximum batch size = 5120<br />
<br />
'''Training Schedules:''' <br />
<br />
The authors consider four training schedules, all of which decay the noise scale by a factor of five in a series of three steps with the same number of epochs.<br />
<br />
Original training schedule: initial learning rate of 0.1 which decays by a factor of 5 at each step, a momentum coefficient of 0.9, and a batch size of 128. Follows the implementation of Zagoruyko & Komodakis (2016).<br />
<br />
Increasing batch size: learning rate of 0.1, momentum coefficient of 0.9, initial batch size of 128 that increases by a factor of 5 at each step. <br />
<br />
Increased initial learning rate: initial learning rate of 0.5, initial batch size of 640 that increase during training.<br />
<br />
Increased momentum coefficient: increased initial learning rate of 0.5, initial batch size of 3200 that increase during training, and an increased momentum coefficient of 0.98.<br />
<br />
The results of all training schedules, which are presented in the below figure, are documented in the following table:<br />
<br />
[[File:Paper_40_Table_1.png | 800px|center]]<br />
<br />
[[File:Paper_40_Fig_5.png | 800px|center]]<br />
<br />
<br />
<br />
'''Conclusion:''' Increasing the effective learning rate and scaling the batch size results in further reduction in the number of parameter updates<br />
<br />
=== TRAINING IMAGENET IN 2500 PARAMETER UPDATES===<br />
<br />
'''A) Experiment Goal:''' Control Batch Size<br />
<br />
'''Dataset:''' ImageNet (1.28 million training images)<br />
<br />
The paper modified the setup of Goyal et al. (2017), and used the following configuration:<br />
<br />
'''Network Architecture:''' Inception-ResNet-V2 <br />
<br />
'''Training Parameters:''' <br />
<br />
90 epochs / noise decayed at epoch 30, 60, and 80 by a factor of 10 / Initial ghost batch size = 32 / Learning rate = 3 / momentum coefficient = 0.9 / Initial batch size = 8192<br />
<br />
Two training schedules were used:<br />
<br />
“Decaying learning rate”, where batch size is fixed and the learning rate is decayed<br />
<br />
“Increasing batch size”, where batch size is increased to 81920 then the learning rate is decayed at two steps.<br />
<br />
[[File:Paper_40_Table_2.png | 800px|center]]<br />
<br />
[[File:Paper_40_Fig_6.png | 800px|center]]<br />
<br />
'''Conclusion:''' Increasing the batch size resulted in reducing the number of parameter updates from 14,000 to 6,000.<br />
<br />
'''B) Experiment Goal:''' Control Batch Size and Momentum Coefficient<br />
<br />
'''Training Parameters:''' Ghost batch size = 64 / noise decayed at epoch 30, 60, and 80 by a factor of 10. <br />
<br />
The below table shows the number of parameter updates and accuracy for different set of training parameters:<br />
<br />
[[File:Paper_40_Table_3.png | 800px|center]]<br />
<br />
[[File:Paper_40_Fig_7.png | 800px|center]]<br />
<br />
'''Conclusion:''' Increasing the momentum reduces the number of parameter updates, but leads to a drop in the test accuracy.<br />
<br />
=== TRAINING IMAGENET IN 30 MINUTES===<br />
<br />
'''Dataset:''' ImageNet (Already introduced in the previous section)<br />
<br />
'''Network Architecture:''' ResNet-50<br />
<br />
The paper replicated the setup of Goyal et al. (2017) while modifying the number of TPU devices, batch size, learning rate, and then calculating the time to complete 90 epochs, and measuring the accuracy, and performed the following experiments below:<br />
<br />
[[File:Paper_40_Table_4.png | 800px|center]]<br />
<br />
'''Conclusion:''' Model training times can be reduced by increasing the batch size during training.<br />
<br />
== RELATED WORK ==<br />
Main related work mentioned in the paper is as follows:<br />
<br />
- Smith & Le (2017) interpreted Stochastic gradient descent as stochastic differential equation; the paper built on this idea to include decaying learning rate.<br />
<br />
- Mandt et al. (2017) analyzed how to modify SGD for the task of Bayesian posterior sampling.<br />
<br />
- Keskar et al. (2016) focused on the analysis of noise once the training is started.<br />
<br />
- Moreover, the proportional relationship between batch size and learning rate was first discovered by Goyal et al. (2017) and successfully trained ResNet-50 on ImageNet in one hour after discovering the proportionality relationship between batch size and learning rate.<br />
<br />
- Furthermore, You et al. (2017a) presented Layer-wise Adaptive Rate Scaling (LARS), which is applying different learning rates to train ImageNet in 14 minutes and 74.9% accuracy. <br />
<br />
- Wilson et al. (2017) argued that adaptive optimization methods tend to generalize less well than SGD and SGD with momentum (although<br />
they did not include K-FAC in their study), while the authors' work reduces the gap in convergence speed.<br />
<br />
- Finally, another strategy called Asynchronous-SGD that allowed (Recht et al., 2011; Dean et al., 2012) to use multiple GPUs even with small batch sizes.<br />
<br />
== CONCLUSIONS ==<br />
Increasing the batch size during training has the same benefits of decaying the learning rate in addition to reducing the number of parameter updates, which corresponds to faster training time. Experiments were performed on different image datasets and various optimizers with different training schedules to prove this result. The paper proposed to increase the learning rate and momentum parameter <math>m</math>, while scaling <math> B \propto \frac{\epsilon}{1-m} </math>, which achieves fewer parameter updates, but slightly less test set accuracy as mentioned in detail in the experiments’ section. In summary, on ImageNet dataset, Inception-ResNet-V2 achieved 77% validation accuracy in under 2500 parameter updates, and ResNet-50 achieved 76.1% validation set accuracy on TPU in less than 30 minutes. One of the great finding of this paper is that all the methods use the hyper-parameters directly from previous works in the literature, and no additional hyper-parameter tuning was performed.<br />
<br />
== CRITIQUE ==<br />
'''Pros:'''<br />
<br />
- The paper showed empirically that increasing batch size and decaying learning rate are equivalent.<br />
<br />
- Several experiments were performed on different optimizers such as SGD and Adam.<br />
<br />
- Had several comparisons with previous experimental setups.<br />
<br />
'''Cons:'''<br />
<br />
<br />
- All datasets used are image datasets. Other experiments should have been done on datasets from different domains to ensure generalization. <br />
<br />
- The number of parameter updates was used as a comparison criterion, but wall-clock times could have provided additional measurable judgment although they depend on the hardware used.<br />
<br />
- Special hardware is needed for large batch training, which is not always feasible. As batch-size increases, we generally need more RAM to train the same model. However, if learning rate is decreased, the RAM use remains constant. As a result, learning rate decay will allow us to train bigger models.<br />
<br />
- In section 5.2 (Increasing the Effective Learning rate), the authors did not test a range of learning rate values and used only (0.1 and 0.5). Additional results from varying the initial learning rate values from 0.1 to 3.2 are provided in the appendix, which indicates that the test accuracy begins to fall for initial learning rates greater than ~0.4. The appended results do not show validation set accuracy curves like in Figure 6, however. It would be beneficial to see if they were similar to the original 0.1 and 0.5 initial learning rate baselines.<br />
<br />
- Although the main idea of the paper is interesting, its results does not seem to be too surprising in comparison with other recent papers in the subject.<br />
<br />
- The paper could benefit from using some other models to demonstrate its claim and generalize its idea by adding some comparisons with other models as well as other recent methods to increase batch size.<br />
<br />
- The paper presents interesting ideas. However, it lacks of mathematical and theoretical analysis beyond the idea. Since the experiment is primary on image dataset and it does not provide sufficient theories, the paper itself presents limited applicability to other types. <br />
<br />
- Also, in experimental setting, only single training runs from one random initialization is used. It would be better to take the best of many runs or to show confidence intervals.<br />
<br />
- It is proposed that we should compare learning rate decay with batch-size increase under the setting that total budget / number of training samples is fixed.<br />
<br />
== REFERENCES ==<br />
# Takuya Akiba, Shuji Suzuki, and Keisuke Fukuda. Extremely large minibatch sgd: Training resnet-50 on imagenet in 15 minutes. arXiv preprint arXiv:1711.04325, 2017.<br />
#Lukas Balles, Javier Romero, and Philipp Hennig. Coupling adaptive batch sizes with learning rates.arXiv preprint arXiv:1612.05086, 2016.<br />
#L´eon Bottou, Frank E Curtis, and Jorge Nocedal. Optimization methods for large-scale machine learning.arXiv preprint arXiv:1606.04838, 2016.<br />
#Richard H Byrd, Gillian M Chin, Jorge Nocedal, and Yuchen Wu. Sample size selection in optimization methods for machine learning. Mathematical programming, 134(1):127–155, 2012.<br />
#Pratik Chaudhari, Anna Choromanska, Stefano Soatto, and Yann LeCun. Entropy-SGD: Biasing gradient descent into wide valleys. arXiv preprint arXiv:1611.01838, 2016.<br />
#Soham De, Abhay Yadav, David Jacobs, and Tom Goldstein. Automated inference with adaptive batches. In Artificial Intelligence and Statistics, pp. 1504–1513, 2017.<br />
#Jeffrey Dean, Greg Corrado, Rajat Monga, Kai Chen, Matthieu Devin, Mark Mao, Andrew Senior, Paul Tucker, Ke Yang, Quoc V Le, et al. Large scale distributed deep networks. In Advances in neural information processing systems, pp. 1223–1231, 2012.<br />
#Michael P Friedlander and Mark Schmidt. Hybrid deterministic-stochastic methods for data fitting.SIAM Journal on Scientific Computing, 34(3):A1380–A1405, 2012.<br />
#Priya Goyal, Piotr Doll´ar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch SGD: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017.<br />
#Sepp Hochreiter and J¨urgen Schmidhuber. Flat minima. Neural Computation, 9(1):1–42, 1997.<br />
#Elad Hoffer, Itay Hubara, and Daniel Soudry. Train longer, generalize better: closing the generalization gap in large batch training of neural networks. arXiv preprint arXiv:1705.08741, 2017.<br />
#Norman P Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder Bajwa, Sarah Bates, Suresh Bhatia, Nan Boden, Al Borchers, et al. In-datacenter performance analysis of a tensor processing unit. In Proceedings of the 44th Annual International Symposium on Computer Architecture, pp. 1–12. ACM, 2017.<br />
#Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. arXiv preprint arXiv:1609.04836, 2016.<br />
#Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.<br />
#Alex Krizhevsky. One weird trick for parallelizing convolutional neural networks. arXiv preprint arXiv:1404.5997, 2014.<br />
#Qianxiao Li, Cheng Tai, and E Weinan. Stochastic modified equations and adaptive stochastic gradient algorithms. arXiv preprint arXiv:1511.06251, 2017.<br />
#Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with restarts. arXiv preprint arXiv:1608.03983, 2016.<br />
#Stephan Mandt, Matthew D Hoffman, and DavidMBlei. Stochastic gradient descent as approximate bayesian inference. arXiv preprint arXiv:1704.04289, 2017.<br />
#James Martens and Roger Grosse. Optimizing neural networks with kronecker-factored approximate curvature. In International Conference on Machine Learning, pp. 2408–2417, 2015.<br />
#Yurii Nesterov. A method of solving a convex programming problem with convergence rate o (1/k2). In Soviet Mathematics Doklady, volume 27, pp. 372–376, 1983.<br />
#Lutz Prechelt. Early stopping-but when? Neural Networks: Tricks of the trade, pp. 553–553, 1998.<br />
#Benjamin Recht, Christopher Re, Stephen Wright, and Feng Niu. Hogwild: A lock-free approach to parallelizing stochastic gradient descent. In Advances in neural information processing systems, pp. 693–701, 2011.<br />
#Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pp. 400–407, 1951.<br />
#Samuel L. Smith and Quoc V. Le. A bayesian perspective on generalization and stochastic gradient descent. arXiv preprint arXiv:1710.06451, 2017.<br />
#Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A Alemi. Inception-v4, Inception-ResNet and the impact of residual connections on learning. In AAAI, pp. 4278–4284, 2017.<br />
#Max Welling and Yee W Teh. Bayesian learning via stochastic gradient langevin dynamics. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pp. 681–688, 2011.<br />
#Ashia C Wilson, Rebecca Roelofs, Mitchell Stern, Nathan Srebro, and Benjamin Recht. The marginal value of adaptive gradient methods in machine learning. arXiv preprint arXiv:1705.08292, 2017.<br />
#Yang You, Igor Gitman, and Boris Ginsburg. Scaling SGD batch size to 32k for imagenet training. arXiv preprint arXiv:1708.03888, 2017a.<br />
#Yang You, Zhao Zhang, C Hsieh, James Demmel, and Kurt Keutzer. Imagenet training in minutes. CoRR, abs/1709.05011, 2017b.<br />
#Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016.<br />
#Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. arXiv preprint arXiv:1611.03530, 2016.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=DETECTING_STATISTICAL_INTERACTIONS_FROM_NEURAL_NETWORK_WEIGHTS&diff=42061DETECTING STATISTICAL INTERACTIONS FROM NEURAL NETWORK WEIGHTS2018-11-30T16:05:31Z<p>Ka2khan: /* Experiment */</p>
<hr />
<div>=Introduction=<br />
<br />
It has been commonly believed that one major advantage of neural networks is their capability of modelling complex statistical interactions between features for automatic feature learning. Statistical interactions capture important information on where features often have joint effects with other features on predicting an outcome. The discovery of interactions is especially useful for scientific discoveries and hypothesis validation. For example, physicists may be interested in understanding what joint factors provide evidence for new elementary particles; doctors may want to know what interactions are accounted for in risk prediction models, to compare against known interactions from existing medical literature.<br />
<br />
With the growth in the computational power available Neural Networks have been able to solve many of the complex tasks in a wide variety of fields. This is mainly due to their ability to model complex and non-linear interactions. Neural networks have traditionally been treated as “black box” models, preventing their adoption in many application domains, such as those where explainability is desirable. It has been noted that complex machine learning models can learn unintended patterns from data, raising significant risks to stakeholders [14]. Therefore, in applications where machine learning models are intended for making critical decisions, such as healthcare or finance, it is paramount to understand how they make predictions [9]. Within several areas, like eg: computation social science, interpretability is of utmost importance. Since we do not understand how a neural network comes to its decision, practitioners in these areas tend to prefer simpler models like linear regression, decision trees, etc. which are much more interpretable. In this paper, we are going to present one way of implementing interpretability in a neural network.<br />
<br />
Existing approaches to interpreting neural networks can be summarized into two types. One type is direct interpretation, which focuses on 1) explaining individual feature importance, for example by computing input gradients [13] and decomposing predictions [8], 2) developing attention-based models, which illustrate where neural networks focus during inference [11], and 3) providing model-specific visualizations, such as feature map and gate activation visualizations [15]. The other type is indirect interpretation, for example post-hoc interpretations of feature importance [12] and knowledge distillation to simpler interpretable models [10].<br />
<br />
In this paper, the authors propose Neural Interaction Detection (NID), which can detect any order or form of statistical interaction captured by the feedforward neural network by examining its weight matrix.<br />
<br />
Note that in this paper, we only consider one specific types of neural network, feedforward neural network. Based on the methodology discussed here, the authors suggest that we can build an interpretation method for other types of networks also.<br />
<br />
=Related Work=<br />
<br />
1. Interaction Detection approaches: <br />
* Conduct individual tests for all features' combination such as ANOVA and Additive Groves.<br />
* Define all interaction forms of interest, then later finds the important ones.<br />
- The paper's goal is to detect interactions without compromising the functional forms. Our method accomplishes higher-order interaction detection, which has the benefit of avoiding a high false positive or false discovery rate.<br />
<br />
2. Interpretability: A lot of work has also been done in this particular area and it can be divided it the following broad categories:<br />
* Feature Importance through Decomposition: Methods like Input Gradient(Sundararajan et al., 2017) learns the importance of features through a gradient-based approach similar to backpropagation. Works like Li et al(2017), Murdoch(2017) and Murdoch(2018) study interpretability of LSTMs by looking at phrase and word level importance scores. Bach et al. 2015 and Shrikumar et al. 2016 (DeepLift) study pixel importance in CNNs.<br />
* Studying Visualizations in Models - Karpathy et al. (2015) worked with character generating LSTMs and tried to study activation and firing in certain hidden units for meaningful attributes. (Yosinski et al., 2015 studies feature map visualizations. <br />
* Attention-Based Models: Bahdanau et al. (2014) - These are a different class of models which use attention modules(different architectures) to help focus the neural network to decide the parts of the input that it should look more closely or give more importance to. Looking at the results of these type of model an indirect sense of interpretability can be gauged.<br />
<br />
The approach in this paper is to extract non-additive interactions between variables from the neural network weights.<br />
<br />
=Notations=<br />
Before we dive in to methodology, we are going to define a few notations here. Most of them will be trivial.<br />
<br />
1. Vector: Vectors are defined with bold-lowercases, '''v, w'''<br />
<br />
2. Matrix: Matrice are defined with blod-uppercases, '''V, W'''<br />
<br />
3. Interger Set: For some interger p <math>\in</math> Z, we define [p] := {1,2,3,...,p}<br />
<br />
=Interaction=<br />
First of all, in order to explain the model, we need to be able to explain the interactions and their effects to output. Therefore, we define 'interacion' between variables as below. <br />
<br />
[[File:def_interaction.PNG|900px|center]]<br />
<br />
From the definition above, for a function like, <math>x_1x_2 + sin(x_3 + x_4 + x_5)</math>, we have <math>{[x_1, x_2]}</math> and <math>{[x_3, x_4, x_5]}</math> interactions. And we say that the latter interaction to be 3-way interaction.<br />
<br />
Note that from the definition above, we can naturally deduce that d-way interaction can exist if and only if all of its (d-1) interactions exist. For example, 3-way interaction above shows that we have 2-way interactions <math>{[3,4], [4,5]}</math> and <math>{[3,5]}</math>.<br />
<br />
One thing that we need to keep in mind is that for models like neural network, most of interactions are happening within hidden layers. This means that we needa proper way of measuring interaction strength.<br />
<br />
The key observation is that for any kinds of interaction, at a some hidden unit of some hidden layer, two interacting features the ancestors. In graph-theoretical language, interaction map can be viewed as an associated directed graph and for any interaction <math>\Gamma \in [p]</math>, there exists at least one vertix that has all of features of <math>\Gamma</math> as ancestors. The statement can be rigorized as the following:<br />
<br />
<br />
[[File:prop2.PNG|900px|center]]<br />
<br />
Now, the above mathematical statement gurantees us to measure interaction strengths at ANY hidden layers. For example, if we want to study about interactions at some specific hidden layer, now we now that there exists corresponding vertices between the hidden layer and output layer. Therefore all we need to do is now to find approprite measure which can summarize the information between those two layers.<br />
<br />
Before doing so, let's think about a single-layered neural network. For any one hidden unit, we can have possibly, <math>2^{||W_i,:||}</math>, number of interactions. This means that our search space might be too huge for multi-layered networks. Therefore, we need a some descent way of approximate out search space. Moreover, the authors realized a fast interaction detection by limiting the search complexity of the task by only quantifying interactions created at the first hidden layer.<br />
[[File:network1.PNG|500px|center]]<br />
<br />
==Measuring influence in hidden layers==<br />
As we discussed above, in order to consider interaction between units in any layers, we need to think about their out-going paths. However, we soon encountered the fact that for some fully-connected multi-layer neural network, the search space might be too huge to compare. Therefore, we use information about out-going paths gredient upper bond. To represent the influence of out-going paths at <math>l</math>-hidden layer, we define cumulative impact of weights between output layer and <math>l+1</math>. We define aggregated weights as, <br />
<br />
[[File:def3.PNG|900px|center]]<br />
<br />
<br />
Note that <math>z^{(l)} \in R^{(p_l)}</math> where <math>p_l</math> is the number of hidden units in <math>l</math>-layer.<br />
Moreover, this is the lipschitz constant of gredients. Gredient has been an import variable of measuring influence of features, especially when we consider that input layer's derivative computes the direction normal to decision boundaries.<br />
<br />
==Quantifying influence==<br />
For some <math>i</math> hidden unit at the first hidden layer, which is the closet layer to the input layer, we define the influence strength of some interaction as, <br />
<br />
[[File:measure1.PNG|900px|center]]<br />
<br />
The function <math>\mu</math> will be defined later. Essentially, the formula shows that the strength of influence is defined as the product of the aggregated weight on the first hidden layer and some measure of influence between the first hidden layer and the input layer. <br />
<br />
For the function, <math>\mu</math>, any positive-real valued functions such as max, min and average can be candidates. The effects of those candidates will be tested later.<br />
<br />
Now based on the specifications above, the author suggested the algorithm for searching influential interactions between input layer units as follows:<br />
<br />
It was pointed out that restricting to the first hidden layer might miss some important feature interactions, however, the author state that it is not straightforward how to incorporate the idea of hidden units at intermediate layers to get better interaction detection performance.<br />
<br />
[[File:algorithm1.PNG|850px|center]]<br />
<br />
=Cut-off Model=<br />
Now using the greedy algorithm defined above, we can rank the interactions by their strength. However, in order to access true interactions, we are building the cut-off model which is a generalized additive model (GAM) as below,<br />
<br />
<center><math><br />
c_K('''x''') = \sum_{i=1}^{p}g_i(x_i) + \sum_{i=1}^{K}{g_i}^\prime(x_\chi)<br />
</math></center><br />
<br />
From the above model, each <math>g</math> and <math>g^*</math> are Feed-Forward neural network. We are keep adding interactions until the performance reaches plateaus.<br />
<br />
=Experiment=<br />
For the experiment, the authors have compared three neural network model with traditional statistical interaction detecting algorithms. For the nueral network models, first model will be MLP, second model will be MLP-M, which is MLP with additional univariate network at the output. The last one is the cut-off model defined above, which is denoted by MLP-cutoff. In the experiments that the authors performed, all the networks which modelled feature interactions consisted of four hidden layers containing 140, 100, 60, and 20 units respectively. Whereas, all the individual univariate networks contained three hidden layers with each layer containing 10 units. All of these networks used ReLu activation and backpropagation for training. The MLP-M model is graphically represented below.<br />
<br />
[[File:output11.PNG|300px|center]]<br />
<br />
For the experiment, the authors study our interaction detection framework on both simulated and real-world experiments. For simulated experiments, the authors are going to test on 10 synthetic functions as shown in table I.<br />
<br />
[[File:synthetic.PNG|900px|center]]<br />
<br />
The authors use four real-world datasets, of which two are regression datasets, and the other two are binary classification datasets. The datasets are a mixture of common prediction tasks in the cal housing<br />
and bike sharing datasets, a scientific discovery task in the higgs boson dataset, and an example of very-high order interaction detection in the letter dataset.<br />
<br />
And the authors also reported the results of comparisons between the models. As you can see, neural network based models are performing better on average. Compare to the traditional methods like ANOVA, MLP and MLP-M method shows 20% increases in performance.<br />
<br />
[[File:performance_mlpm.PNG|900px|center]]<br />
<br />
<br />
[[File:performance2_mlpm.PNG|900px|center]]<br />
<br />
The above result shows that MLP-M almost perfectly capture the most influential pair-wise interactions.<br />
<br />
=Limitations=<br />
Even though for the above synthetic experiment MLP methods showed superior performances, the method still have some limitations. For example, fir the function like, <math>x_1x_2 + x_2x_3 + x_1x_3</math>, neural network fails to distinguish between interlinked interactions to single higher order interaction. Moreoever, correlation between features deteriorates the ability of the network to distinguish interactions. However, correlation issues are presented most of interaction detection algorithms. <br />
<br />
Because this method relies on the neural network fitting the data well, there are some additional concerns. Notably, if the NN is unable to make an appropriate fit (under/overfitting), the resulting interactions will be flawed. This can occur if the datasets that are too small or too noisy, which often occurs in practical settings. <br />
<br />
=Conclusion=<br />
Here we presented the method of detecting interactions using MLP. Compared to other state-of-the-art methods like Additive Groves (AG), the performances are competitive yet computational powers required is far less. Therefore, it is safe to claim that the method will be extremly useful for practitioners with (comparably) less computational powers. Moreover, the NIP algorithm successfully reduced the computation sizes. After all, the most important aspect of this algorithm is that now users of nueral networks can impose interpretability in the model usage, which will change the level of usability to another level for most of practitioners outside of those working in machine learning and deep learning areas.<br />
<br />
For future work, the authors want to detect feature interactions by using the common units in the intermediate hidden layers of feedforward networks, and also want to use such interaction detection to interpret weights in other deep neural networks. Also, it was pointed out that the neural network weights heavily depend on L-1 regularized neural network training, but a group lasso penalty may work better.<br />
<br />
=Critique=<br />
1. Authors need to do large-scale experiments, instead of just conducting experiments on some synthetic dataset with small feature dimensionality, to make their claim stronger.<br />
<br />
2. Although the method proposed in this paper is interesting, the paper would benefit from providing some more explanations to support its idea and fill the possible gaps in its experimental evaluation. In some parts there are repetitive explanations that could be replaced by other essential clarifications.<br />
<br />
3. Greedy algorithm is implemented by nothing is mentioned about the speed of this algorithm which is definitely not fast. So, this has the potential to be a weak point of the study.<br />
<br />
=Reference=<br />
<br />
[1] Jacob Bien, Jonathan Taylor, and Robert Tibshirani. A lasso for hierarchical interactions. Annals of statistics, 41(3):1111, 2013. <br />
<br />
[2] G David Garson. Interpreting neural-network connection weights. AI Expert, 6(4):46–51, 1991.<br />
<br />
[3] Yotam Hechtlinger. Interpretation of prediction models using the input gradient. arXiv preprint arXiv:1611.07634, 2016.<br />
<br />
[4] Shiyu Liang and R Srikant. Why deep neural networks for function approximation? 2016. <br />
<br />
[5] David Rolnick and Max Tegmark. The power of deeper networks for expressing natural functions. International Conference on Learning Representations, 2018. <br />
<br />
[6] Daria Sorokina, Rich Caruana, and Mirek Riedewald. Additive groves of regression trees. Machine Learning: ECML 2007, pp. 323–334, 2007.<br />
<br />
[7] Simon Wood. Generalized additive models: an introduction with R. CRC press, 2006<br />
<br />
[8] Sebastian Bach, Alexander Binder, Gre ́goire Montavon, Frederick Klauschen, Klaus-Robert Mu ̈ller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PloS one, 10(7):e0130140, 2015.<br />
<br />
[9] Rich Caruana, Yin Lou, Johannes Gehrke, Paul Koch, Marc Sturm, and Noemie Elhadad. Intel- ligible models for healthcare: Predicting pneumonia risk and hospital 30-day readmission. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 1721–1730. ACM, 2015.<br />
<br />
[10] Zhengping Che, Sanjay Purushotham, Robinder Khemani, and Yan Liu. Interpretable deep models for icu outcome prediction. In AMIA Annual Symposium Proceedings, volume 2016, pp. 371. American Medical Informatics Association, 2016.<br />
<br />
[11] Laurent Itti, Christof Koch, and Ernst Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on pattern analysis and machine intelligence, 20(11):1254– 1259, 1998.<br />
<br />
[12] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. Why should i trust you?: Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 1135–1144. ACM, 2016.<br />
<br />
[13]Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Vi- sualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013.<br />
<br />
[14] Kush R Varshney and Homa Alemzadeh. On the safety of machine learning: Cyber-physical sys- tems, decision sciences, and data products. arXiv preprint arXiv:1610.01256, 2016.<br />
<br />
[15] Jason Yosinski, Jeff Clune, Anh Nguyen, Thomas Fuchs, and Hod Lipson. Understanding neural networks through deep visualization. arXiv preprint arXiv:1506.06579, 2015.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=DETECTING_STATISTICAL_INTERACTIONS_FROM_NEURAL_NETWORK_WEIGHTS&diff=42060DETECTING STATISTICAL INTERACTIONS FROM NEURAL NETWORK WEIGHTS2018-11-30T16:03:32Z<p>Ka2khan: /* Experiment */</p>
<hr />
<div>=Introduction=<br />
<br />
It has been commonly believed that one major advantage of neural networks is their capability of modelling complex statistical interactions between features for automatic feature learning. Statistical interactions capture important information on where features often have joint effects with other features on predicting an outcome. The discovery of interactions is especially useful for scientific discoveries and hypothesis validation. For example, physicists may be interested in understanding what joint factors provide evidence for new elementary particles; doctors may want to know what interactions are accounted for in risk prediction models, to compare against known interactions from existing medical literature.<br />
<br />
With the growth in the computational power available Neural Networks have been able to solve many of the complex tasks in a wide variety of fields. This is mainly due to their ability to model complex and non-linear interactions. Neural networks have traditionally been treated as “black box” models, preventing their adoption in many application domains, such as those where explainability is desirable. It has been noted that complex machine learning models can learn unintended patterns from data, raising significant risks to stakeholders [14]. Therefore, in applications where machine learning models are intended for making critical decisions, such as healthcare or finance, it is paramount to understand how they make predictions [9]. Within several areas, like eg: computation social science, interpretability is of utmost importance. Since we do not understand how a neural network comes to its decision, practitioners in these areas tend to prefer simpler models like linear regression, decision trees, etc. which are much more interpretable. In this paper, we are going to present one way of implementing interpretability in a neural network.<br />
<br />
Existing approaches to interpreting neural networks can be summarized into two types. One type is direct interpretation, which focuses on 1) explaining individual feature importance, for example by computing input gradients [13] and decomposing predictions [8], 2) developing attention-based models, which illustrate where neural networks focus during inference [11], and 3) providing model-specific visualizations, such as feature map and gate activation visualizations [15]. The other type is indirect interpretation, for example post-hoc interpretations of feature importance [12] and knowledge distillation to simpler interpretable models [10].<br />
<br />
In this paper, the authors propose Neural Interaction Detection (NID), which can detect any order or form of statistical interaction captured by the feedforward neural network by examining its weight matrix.<br />
<br />
Note that in this paper, we only consider one specific types of neural network, feedforward neural network. Based on the methodology discussed here, the authors suggest that we can build an interpretation method for other types of networks also.<br />
<br />
=Related Work=<br />
<br />
1. Interaction Detection approaches: <br />
* Conduct individual tests for all features' combination such as ANOVA and Additive Groves.<br />
* Define all interaction forms of interest, then later finds the important ones.<br />
- The paper's goal is to detect interactions without compromising the functional forms. Our method accomplishes higher-order interaction detection, which has the benefit of avoiding a high false positive or false discovery rate.<br />
<br />
2. Interpretability: A lot of work has also been done in this particular area and it can be divided it the following broad categories:<br />
* Feature Importance through Decomposition: Methods like Input Gradient(Sundararajan et al., 2017) learns the importance of features through a gradient-based approach similar to backpropagation. Works like Li et al(2017), Murdoch(2017) and Murdoch(2018) study interpretability of LSTMs by looking at phrase and word level importance scores. Bach et al. 2015 and Shrikumar et al. 2016 (DeepLift) study pixel importance in CNNs.<br />
* Studying Visualizations in Models - Karpathy et al. (2015) worked with character generating LSTMs and tried to study activation and firing in certain hidden units for meaningful attributes. (Yosinski et al., 2015 studies feature map visualizations. <br />
* Attention-Based Models: Bahdanau et al. (2014) - These are a different class of models which use attention modules(different architectures) to help focus the neural network to decide the parts of the input that it should look more closely or give more importance to. Looking at the results of these type of model an indirect sense of interpretability can be gauged.<br />
<br />
The approach in this paper is to extract non-additive interactions between variables from the neural network weights.<br />
<br />
=Notations=<br />
Before we dive in to methodology, we are going to define a few notations here. Most of them will be trivial.<br />
<br />
1. Vector: Vectors are defined with bold-lowercases, '''v, w'''<br />
<br />
2. Matrix: Matrice are defined with blod-uppercases, '''V, W'''<br />
<br />
3. Interger Set: For some interger p <math>\in</math> Z, we define [p] := {1,2,3,...,p}<br />
<br />
=Interaction=<br />
First of all, in order to explain the model, we need to be able to explain the interactions and their effects to output. Therefore, we define 'interacion' between variables as below. <br />
<br />
[[File:def_interaction.PNG|900px|center]]<br />
<br />
From the definition above, for a function like, <math>x_1x_2 + sin(x_3 + x_4 + x_5)</math>, we have <math>{[x_1, x_2]}</math> and <math>{[x_3, x_4, x_5]}</math> interactions. And we say that the latter interaction to be 3-way interaction.<br />
<br />
Note that from the definition above, we can naturally deduce that d-way interaction can exist if and only if all of its (d-1) interactions exist. For example, 3-way interaction above shows that we have 2-way interactions <math>{[3,4], [4,5]}</math> and <math>{[3,5]}</math>.<br />
<br />
One thing that we need to keep in mind is that for models like neural network, most of interactions are happening within hidden layers. This means that we needa proper way of measuring interaction strength.<br />
<br />
The key observation is that for any kinds of interaction, at a some hidden unit of some hidden layer, two interacting features the ancestors. In graph-theoretical language, interaction map can be viewed as an associated directed graph and for any interaction <math>\Gamma \in [p]</math>, there exists at least one vertix that has all of features of <math>\Gamma</math> as ancestors. The statement can be rigorized as the following:<br />
<br />
<br />
[[File:prop2.PNG|900px|center]]<br />
<br />
Now, the above mathematical statement gurantees us to measure interaction strengths at ANY hidden layers. For example, if we want to study about interactions at some specific hidden layer, now we now that there exists corresponding vertices between the hidden layer and output layer. Therefore all we need to do is now to find approprite measure which can summarize the information between those two layers.<br />
<br />
Before doing so, let's think about a single-layered neural network. For any one hidden unit, we can have possibly, <math>2^{||W_i,:||}</math>, number of interactions. This means that our search space might be too huge for multi-layered networks. Therefore, we need a some descent way of approximate out search space. Moreover, the authors realized a fast interaction detection by limiting the search complexity of the task by only quantifying interactions created at the first hidden layer.<br />
[[File:network1.PNG|500px|center]]<br />
<br />
==Measuring influence in hidden layers==<br />
As we discussed above, in order to consider interaction between units in any layers, we need to think about their out-going paths. However, we soon encountered the fact that for some fully-connected multi-layer neural network, the search space might be too huge to compare. Therefore, we use information about out-going paths gredient upper bond. To represent the influence of out-going paths at <math>l</math>-hidden layer, we define cumulative impact of weights between output layer and <math>l+1</math>. We define aggregated weights as, <br />
<br />
[[File:def3.PNG|900px|center]]<br />
<br />
<br />
Note that <math>z^{(l)} \in R^{(p_l)}</math> where <math>p_l</math> is the number of hidden units in <math>l</math>-layer.<br />
Moreover, this is the lipschitz constant of gredients. Gredient has been an import variable of measuring influence of features, especially when we consider that input layer's derivative computes the direction normal to decision boundaries.<br />
<br />
==Quantifying influence==<br />
For some <math>i</math> hidden unit at the first hidden layer, which is the closet layer to the input layer, we define the influence strength of some interaction as, <br />
<br />
[[File:measure1.PNG|900px|center]]<br />
<br />
The function <math>\mu</math> will be defined later. Essentially, the formula shows that the strength of influence is defined as the product of the aggregated weight on the first hidden layer and some measure of influence between the first hidden layer and the input layer. <br />
<br />
For the function, <math>\mu</math>, any positive-real valued functions such as max, min and average can be candidates. The effects of those candidates will be tested later.<br />
<br />
Now based on the specifications above, the author suggested the algorithm for searching influential interactions between input layer units as follows:<br />
<br />
It was pointed out that restricting to the first hidden layer might miss some important feature interactions, however, the author state that it is not straightforward how to incorporate the idea of hidden units at intermediate layers to get better interaction detection performance.<br />
<br />
[[File:algorithm1.PNG|850px|center]]<br />
<br />
=Cut-off Model=<br />
Now using the greedy algorithm defined above, we can rank the interactions by their strength. However, in order to access true interactions, we are building the cut-off model which is a generalized additive model (GAM) as below,<br />
<br />
<center><math><br />
c_K('''x''') = \sum_{i=1}^{p}g_i(x_i) + \sum_{i=1}^{K}{g_i}^\prime(x_\chi)<br />
</math></center><br />
<br />
From the above model, each <math>g</math> and <math>g^*</math> are Feed-Forward neural network. We are keep adding interactions until the performance reaches plateaus.<br />
<br />
=Experiment=<br />
For the experiment, the authors have compared three neural network model with traditional statistical interaction detecting algorithms. For the nueral network models, first model will be MLP, second model will be MLP-M, which is MLP with additional univariate network at the output. The last one is the cut-off model defined above, which is denoted by MLP-cutoff. In the experiments that the authors performed, all the networks which modelled feature interactions consisted of four hidden layers containing 140, 100, 60, and 20 units respectively. Whereas, all the individual univariate networks contained three hidden layers with each layer containing 10 units. All of these networks used ReLu activation and backpropagation for training. The MLP-M model is graphically represented below.<br />
<br />
[[File:output11.PNG|300px|center]]<br />
<br />
For the experiment, We study our interaction detection framework on both simulated and real-world experiments. For simulated experiments, we are going to test on 10 synthetic functions as shown in table I.<br />
<br />
[[File:synthetic.PNG|900px|center]]<br />
<br />
We use four real-world datasets, of which two are regression datasets, and the other two are binary classification datasets. The datasets are a mixture of common prediction tasks in the cal housing<br />
and bike sharing datasets, a scientific discovery task in the higgs boson dataset, and an example of very-high order interaction detection in the letter dataset.<br />
<br />
And the author also reported the results of comparisons between the models. As you can see, neural network based models are performing better in average. Compare to the traditional methods liek ANOVA, MLP and MLP-M method shows 20% increases in performance.<br />
<br />
[[File:performance_mlpm.PNG|900px|center]]<br />
<br />
<br />
[[File:performance2_mlpm.PNG|900px|center]]<br />
<br />
The above result shows that MLP-M almost perfectly catch the most influential pair-wise interactions.<br />
<br />
=Limitations=<br />
Even though for the above synthetic experiment MLP methods showed superior performances, the method still have some limitations. For example, fir the function like, <math>x_1x_2 + x_2x_3 + x_1x_3</math>, neural network fails to distinguish between interlinked interactions to single higher order interaction. Moreoever, correlation between features deteriorates the ability of the network to distinguish interactions. However, correlation issues are presented most of interaction detection algorithms. <br />
<br />
Because this method relies on the neural network fitting the data well, there are some additional concerns. Notably, if the NN is unable to make an appropriate fit (under/overfitting), the resulting interactions will be flawed. This can occur if the datasets that are too small or too noisy, which often occurs in practical settings. <br />
<br />
=Conclusion=<br />
Here we presented the method of detecting interactions using MLP. Compared to other state-of-the-art methods like Additive Groves (AG), the performances are competitive yet computational powers required is far less. Therefore, it is safe to claim that the method will be extremly useful for practitioners with (comparably) less computational powers. Moreover, the NIP algorithm successfully reduced the computation sizes. After all, the most important aspect of this algorithm is that now users of nueral networks can impose interpretability in the model usage, which will change the level of usability to another level for most of practitioners outside of those working in machine learning and deep learning areas.<br />
<br />
For future work, the authors want to detect feature interactions by using the common units in the intermediate hidden layers of feedforward networks, and also want to use such interaction detection to interpret weights in other deep neural networks. Also, it was pointed out that the neural network weights heavily depend on L-1 regularized neural network training, but a group lasso penalty may work better.<br />
<br />
=Critique=<br />
1. Authors need to do large-scale experiments, instead of just conducting experiments on some synthetic dataset with small feature dimensionality, to make their claim stronger.<br />
<br />
2. Although the method proposed in this paper is interesting, the paper would benefit from providing some more explanations to support its idea and fill the possible gaps in its experimental evaluation. In some parts there are repetitive explanations that could be replaced by other essential clarifications.<br />
<br />
3. Greedy algorithm is implemented by nothing is mentioned about the speed of this algorithm which is definitely not fast. So, this has the potential to be a weak point of the study.<br />
<br />
=Reference=<br />
<br />
[1] Jacob Bien, Jonathan Taylor, and Robert Tibshirani. A lasso for hierarchical interactions. Annals of statistics, 41(3):1111, 2013. <br />
<br />
[2] G David Garson. Interpreting neural-network connection weights. AI Expert, 6(4):46–51, 1991.<br />
<br />
[3] Yotam Hechtlinger. Interpretation of prediction models using the input gradient. arXiv preprint arXiv:1611.07634, 2016.<br />
<br />
[4] Shiyu Liang and R Srikant. Why deep neural networks for function approximation? 2016. <br />
<br />
[5] David Rolnick and Max Tegmark. The power of deeper networks for expressing natural functions. International Conference on Learning Representations, 2018. <br />
<br />
[6] Daria Sorokina, Rich Caruana, and Mirek Riedewald. Additive groves of regression trees. Machine Learning: ECML 2007, pp. 323–334, 2007.<br />
<br />
[7] Simon Wood. Generalized additive models: an introduction with R. CRC press, 2006<br />
<br />
[8] Sebastian Bach, Alexander Binder, Gre ́goire Montavon, Frederick Klauschen, Klaus-Robert Mu ̈ller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PloS one, 10(7):e0130140, 2015.<br />
<br />
[9] Rich Caruana, Yin Lou, Johannes Gehrke, Paul Koch, Marc Sturm, and Noemie Elhadad. Intel- ligible models for healthcare: Predicting pneumonia risk and hospital 30-day readmission. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 1721–1730. ACM, 2015.<br />
<br />
[10] Zhengping Che, Sanjay Purushotham, Robinder Khemani, and Yan Liu. Interpretable deep models for icu outcome prediction. In AMIA Annual Symposium Proceedings, volume 2016, pp. 371. American Medical Informatics Association, 2016.<br />
<br />
[11] Laurent Itti, Christof Koch, and Ernst Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on pattern analysis and machine intelligence, 20(11):1254– 1259, 1998.<br />
<br />
[12] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. Why should i trust you?: Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 1135–1144. ACM, 2016.<br />
<br />
[13]Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Vi- sualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013.<br />
<br />
[14] Kush R Varshney and Homa Alemzadeh. On the safety of machine learning: Cyber-physical sys- tems, decision sciences, and data products. arXiv preprint arXiv:1610.01256, 2016.<br />
<br />
[15] Jason Yosinski, Jeff Clune, Anh Nguyen, Thomas Fuchs, and Hod Lipson. Understanding neural networks through deep visualization. arXiv preprint arXiv:1506.06579, 2015.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Fix_your_classifier:_the_marginal_value_of_training_the_last_weight_layer&diff=41931Fix your classifier: the marginal value of training the last weight layer2018-11-29T23:14:16Z<p>Ka2khan: /* Language Modelling */</p>
<hr />
<div>=Introduction=<br />
<br />
Deep neural networks have become a widely used model for machine learning, achieving state-of-the-art results on many tasks. The most common task these models are used for is to perform classification, as in the case of convolutional neural networks (CNNs) being used to classify images to a semantic category. Typically, a learned affine transformation is placed at the end of such models, yielding a per-class value used for classification. This classifier can have a vast number of parameters, which grows linearly with the number of possible classes, thus requiring increasingly more computational resources.<br />
<br />
=Brief Overview=<br />
<br />
In order to alleviate the aforementioned problem, the authors propose that the final layer of the classifier be fixed (upto a global scale constant). They argue that with little or no loss of accuracy for most classification tasks, the method provides significant memory and computational benefits. In addition, they show that by initializing the classifier with a Hadamard matrix the inference could be made faster as well.<br />
<br />
=Previous Work=<br />
<br />
Training NN models and using them for inference requires large amounts of memory and computational resources; thus, extensive amount of research has been done lately to reduce the size of networks which are as follows:<br />
<br />
* Weight sharing and specification (Han et al., 2015)<br />
<br />
* Mixed precision to reduce the size of the neural networks by half (Micikevicius et al., 2017)<br />
<br />
* Low-rank approximations to speed up CNN (Tai et al., 2015)<br />
<br />
* Quantization of weights, activations and gradients to further reduce computation during training (Hubara et al., 2016b; Li et al., 2016 and Zhou et al., 2016)<br />
<br />
Some of the past works have also put forward the fact that predefined (Park & Sandberg, 1991) and random (Huang et al., 2006) projections can be used together with a learned affine transformation to achieve competitive results on many of the classification tasks. However, the authors' proposal in the current paper is quite reversed.<br />
<br />
=Background=<br />
<br />
Convolutional neural networks (CNNs) are commonly used to solve a variety of spatial and temporal tasks. CNNs are usually composed of a stack of convolutional parameterized layers, spatial pooling layers and fully connected layers, separated by non-linear activation functions. Earlier architectures of CNNs (LeCun et al., 1998; Krizhevsky et al., 2012) used a set of fully-connected layers at later stage of the network, presumably to allow classification based on global features of an image.<br />
<br />
== Shortcomings of the final classification layer and its solution ==<br />
<br />
Despite the enormous number of trainable parameters these layers added to the model, they are known to have a rather marginal impact on the final performance of the network (Zeiler & Fergus, 2014).<br />
<br />
It has been shown previously that these layers could be easily compressed and reduced after a model was trained by simple means such as matrix decomposition and sparsification (Han et al., 2015). Modern architecture choices are characterized with the removal of most of the fully connected layers (Lin et al., 2013; Szegedy et al., 2015; He et al., 2016), that lead to better generalization and overall accuracy, together with a huge decrease in the number of trainable parameters. Additionally, numerous works showed that CNNs can be trained in a metric learning regime (Bromley et al., 1994; Schroff et al., 2015; Hoffer & Ailon, 2015), where no explicit classification layer was introduced and the objective regarded only distance measures between intermediate representations. Hardt & Ma (2017) suggested an all-convolutional network variant, where they kept the original initialization of the classification layer fixed with no negative impact on performance on the CIFAR-10 dataset.<br />
<br />
=Proposed Method=<br />
<br />
The aforementioned works provide evidence that fully-connected layers are in fact redundant and play a small role in learning and generalization. In this work, the authors have suggested that parameters used for the final classification transform are completely redundant, and can be replaced with a predetermined linear transform. This holds for even in large-scale models and classification tasks, such as recent architectures trained on the ImageNet benchmark (Deng et al., 2009).<br />
<br />
==Using a fixed classifier==<br />
<br />
Suppose the final representation obtained by the network (the last hidden layer) is represented as <math>x = F(z;\theta)</math> where <math>F</math> is assumed to be a deep neural network with input z and parameters θ, e.g., a convolutional network, trained by backpropagation.<br />
<br />
In common NN models, this representation is followed by an additional affine transformation, <math>y = W^T x + b</math> ,where <math>W</math> and <math>b</math> are also trained by back-propagation.<br />
<br />
For input <math>x</math> of <math>N</math> length, and <math>C</math> different possible outputs, <math>W</math> is required to be a matrix of <math>N ×<br />
C</math>. Training is done using cross-entropy loss, by feeding the network outputs through a softmax activation<br />
<br />
<math><br />
v_i = \frac{e^{y_i}}{\sum_{j}^{C}{e^{y_j}}}, i &isin; </math> { <math> {1, . . . , C} </math> }<br />
<br />
and reducing the expected negative log likelihood with respect to ground-truth target <math> t &isin; </math> { <math> {1, . . . , C} </math> },<br />
by minimizing the loss function:<br />
<br />
<math><br />
L(x, t) = −\text{log}\ {v_t} = −{w_t}·{x} − b_t + \text{log} ({\sum_{j}^{C}e^{w_j . x + b_j}})<br />
</math><br />
<br />
where <math>w_i</math> is the <math>i</math>-th column of <math>W</math>.<br />
<br />
==Choosing the projection matrix==<br />
<br />
To evaluate the conjecture regarding the importance of the final classification transformation, the trainable parameter matrix <math>W</math> is replaced with a fixed orthonormal projection <math> Q &isin; R^{N×C} </math>, such that <math> &forall; i &ne; j : q_i · q_j = 0 </math> and <math> || q_i ||_{2} = 1 </math>, where <math>q_i</math> is the <math>i</math>th column of <math>Q</math>. This is ensured by a simple random sampling and singular-value decomposition<br />
<br />
As the rows of classifier weight matrix are fixed with an equally valued <math>L_{2}</math> norm, we find it beneficial<br />
to also restrict the representation of <math>x</math> by normalizing it to reside on the <math>n</math>-dimensional sphere:<br />
<br />
<center><math><br />
\hat{x} = \frac{x}{||x||_{2}}<br />
</math></center><br />
<br />
This allows faster training and convergence, as the network does not need to account for changes in the scale of its weights. However, it has now an issue that <math>q_i · \hat{x} </math> is bounded between −1 and 1. This causes convergence issues, as the softmax function is scale sensitive, and the network is affected by the inability to re-scale its input. This issue is amended with a fixed scale <math>T</math> applied to softmax inputs <math>f(y) = softmax(\frac{1}{T}y)</math>, also known as the ''softmax temperature''. However, this introduces an additional hyper-parameter which may differ between networks and datasets. So, the authors propose to introduce a single scalar parameter <math>\alpha</math> to learn the softmax scale, effectively functioning as an inverse of the softmax temperature <math>\frac{1}{T}</math>. The normalized weights and an additional scale coefficient are also used, specially using a single scale for all entries in the weight matrix. The additional vector of bias parameters <math>b &isin; \mathbb{R}^{C}</math> is kept the same and the model is trained using the traditional negative-log-likelihood criterion. Explicitly, the classifier output is now:<br />
<br />
<center><br />
<math><br />
v_i=\frac{e^{\alpha q_i &middot; \hat{x} + b_i}}{\sum_{j}^{C} e^{\alpha q_j &middot; \hat{x} + b_j}}, i &isin; </math> { <math> {1,...,C} </math>}<br />
</center><br />
<br />
and the loss to be minimized is:<br />
<br />
<center><math><br />
L(x, t) = -\alpha q_t &middot; \frac{x}{||x||_{2}} + b_t + \text{log} (\sum_{i=1}^{C} \text{exp}((\alpha q_i &middot; \frac{x}{||x||_{2}} + b_i)))<br />
</math></center><br />
<br />
where <math>x</math> is the final representation obtained by the network for a specific sample, and <math> t &isin; </math> { <math> {1, . . . , C} </math> } is the ground-truth label for that sample. The behaviour of the parameter <math> \alpha </math> over time, which is logarithmic in nature and has the same behavior exhibited by the norm of a learned classifier, is shown in<br />
[[Media: figure1_log_behave.png| Figure 1]].<br />
<br />
<center>[[File:figure1_log_behave.png]]</center><br />
<br />
When <math> -1 \le q_i · \hat{x} \le 1 </math>, a possible cosine angle loss is <br />
<br />
<center>[[File:caloss.png]]</center><br />
<br />
But its final validation accuracy has slight decrease, compared to original models.<br />
<br />
==Using a Hadmard matrix==<br />
<br />
To recall, Hadmard matrix (Hedayat et al., 1978) <math> H </math> is an <math> n × n </math> matrix, where all of its entries are either +1 or −1.<br />
Furthermore, <math> H </math> is orthogonal, such that <math> HH^{T} = nI_n </math> where <math>I_n</math> is the identity matrix. Instead of using the entire Hadmard matrix <math>H</math>, a truncated version, <math> \hat{H} &isin; </math> {<math> {-1, 1}</math>}<math>^{C \times N}</math> where all <math>C</math> rows are orthogonal as the final classification layer is such that:<br />
<br />
<center><math><br />
y = \hat{H} \hat{x} + b<br />
</math></center><br />
<br />
This usage allows two main benefits:<br />
* A deterministic, low-memory and easily generated matrix that can be used for classification.<br />
* Removal of the need to perform a full matrix-matrix multiplication - as multiplying by a Hadamard matrix can be done by simple sign manipulation and addition.<br />
<br />
Here, <math>n</math> must be a multiple of 4, but it can be easily truncated to fit normally defined networks. Also, as the classifier weights are fixed to need only 1-bit precision, it is now possible to focus our attention on the features preceding it.<br />
<br />
=Experimental Results=<br />
<br />
The authors have evaluated their proposed model on the following datasets:<br />
<br />
==CIFAR-10/100==<br />
<br />
===About the dataset===<br />
<br />
CIFAR-10 is an image classification benchmark dataset containing 50,000 training images and 10,000 test images. The images are in color and contain 32×32 pixels. There are 10 possible classes of various animals and vehicles. CIFAR-100 holds the same number of images of same size, but contains 100 different classes.<br />
<br />
===Training Details===<br />
<br />
The authors trained a residual network ( He et al., 2016) on the CIFAR-10 dataset. The network depth was 56 and the same hyper-parameters as in the original work were used. A comparison of the two variants, i.e., the learned classifier and the proposed classifier with a fixed transformation is shown in [[Media: figure1_resnet_cifar10.png | Figure 2]].<br />
<br />
<center>[[File: figure1_resnet_cifar10.png]]</center><br />
<br />
These results demonstrate that although the training error is considerably lower for the network with learned classifier, both models achieve the same classification accuracy on the validation set. The authors conjecture is that with the new fixed parameterization, the network can no longer increase the<br />
norm of a given sample’s representation - thus learning its label requires more effort. As this may happen for specific seen samples - it affects only training error.<br />
<br />
The authors also compared using a fixed scale variable <math>\alpha </math> at different values vs. the learned parameter. Results for <math> \alpha = </math> {0.1, 1, 10} are depicted in [[Media: figure3_alpha_resnet_cifar.png| Figure 3]] for both training and validation error and as can be seen, similar validation accuracy can be obtained using a fixed scale value (in this case <math>\alpha </math>= 1 or 10 will suffice) at the expense of another hyper-parameter to seek. In all the further experiments the scaling parameter <math> \alpha </math> was regularized with the same weight decay coefficient used on original classifier.<br />
<br />
<center>[[File: figure3_alpha_resnet_cifar.png]]</center><br />
<br />
The authors then train the model on CIFAR-100 dataset. They used the DenseNet-BC model from Huang et al. (2017) with depth of 100 layers and k = 12. The higher number of classes caused the number of parameters to grow and encompassed about 4% of the whole model. However, validation accuracy for the fixed-classifier model remained equally good as the original model, and the same training curve was observed as earlier.<br />
<br />
==IMAGENET==<br />
<br />
===About the dataset===<br />
<br />
The Imagenet dataset introduced by Deng et al. (2009) spans over 1000 visual classes, and over 1.2 million samples. This is supposedly a more challenging dataset to work on as compared to CIFAR-10/100.<br />
<br />
===Experiment Details===<br />
<br />
The authors evaluated their fixed classifier method on Imagenet using Resnet50 by He et al. (2016) and Densenet169 model (Huang et al., 2017) as described in the original work. Using a fixed classifier removed approximately 2-million parameters were from the model, accounting for about 8% and 12 % of the model parameters respectively. The experiments revealed similar trends as observed on CIFAR-10.<br />
<br />
For a more stricter evaluation, the authors also trained a Shufflenet architecture (Zhang et al., 2017b), which was designed to be used in low memory and limited computing platforms and has parameters making up the majority of the model. They were able to reduce the parameters to 0.86 million as compared to 0.96 million parameters in the final layer of the original model. Again, the proposed modification in the original model gave similar convergence results on validation accuracy.<br />
<br />
The overall results of the fixed-classifier are summarized in [[Media: table1_fixed_results.png | Table 1]].<br />
<br />
<center>[[File: table1_fixed_results.png]]</center><br />
<br />
==Language Modelling==<br />
<br />
Recent works have empirically found that using the same weights for both word embedding and classifier can yield equal or better results than using a separate pair of weights. So the authors experimented with fix-classifiers on language modelling as it also requires classification of all possible tokens available in the task vocabulary. They trained a recurrent model with 2-layers of LSTM (Hochreiter & Schmidhuber, 1997) and embedding + hidden size of 512 on the WikiText2 dataset (Merity et al., 2016), using same settings as in Merity et al. (2017). WikiText2 dataset contains about 33K different words, so the number of parameters expected in the embedding and classifier layer was about 34-million. This number is about 89% of the total number of parameters used for the whole model which is 38-million. However, using a random orthogonal transform yielded poor results compared to learned embedding. This was suspected to be due to semantic relationships captured in the embedding layer of language models, which is not the case in image classification task. The intuition was further confirmed by the much better results when pre-trained embeddings using word2vec algorithm by Mikolov et al. (2013) or PMI factorization as suggested by Levy & Goldberg (2014), were used.<br />
<br />
=Discussion=<br />
<br />
==Implications and use cases==<br />
<br />
With the increasing number of classes in the benchmark datasets, computational demands for the final classifier will increase as well. In order to understand the problem better, the authors observe the work by Sun et al. (2017), which introduced JFT-300M - an internal Google dataset with over 18K different classes. Using a Resnet50 (He et al., 2016), with a 2048 sized representation led to a model with over 36M parameters meaning that over 60% of the model parameters resided in the final classification layer. Sun et al. (2017) also describe the difficulty in distributing so many parameters over the training servers involving a non-trivial overhead during synchronization of the model for update. The authors claim that the fixed-classifier would help considerably in this kind of scenario - where using a fixed classifier removes the need to do any gradient synchronization for the final layer. Furthermore, introduction of Hadamard matrix removes the need to save the transformation altogether, thereby, making it more efficient and allowing considerable memory and computational savings.<br />
<br />
==Possible Caveats==<br />
<br />
The good performance of fixed-classifiers relies on the ability of the preceding layers to learn separable representations. This could be affected when when the ratio between learned features and number of classes is small – that is, when <math> C > N</math>. However, they tested their method in such cases and their model performed well and provided good results.<br />
Another factor that can affect the performance of their model using a fixed classifier is when the classes are highly correlated. In that case, the fixed classifier actually cannot support correlated classes and thus, the network could have some difficulty to learn. For a language model, word classes tend to have highly correlated instances, which also lead to difficult learning process.<br />
<br />
==Future Work==<br />
<br />
<br />
The use of fixed classifiers might be further simplified in Binarized Neural Networks (Hubara et al., 2016a), where the activations and weights are restricted to ±1 during propagations. In that case the norm of the last hidden layer would be constant for all samples (equal to the square root of the hidden layer width). The constant could then be absorbed into the scale constant <math>\alpha</math>, and there is no need in a per-sample normalization.<br />
<br />
Additionally, more efficient ways to learn a word embedding should also be explored where similar redundancy in classifier weights may suggest simpler forms of token representations - such as low-rank or sparse versions.<br />
<br />
A related paper was published that claims that fixing most of the parameters of the neural network achieves comparable results with learning all of them [A. Rosenfeld and J. K. Tsotsos]<br />
<br />
=Conclusion=<br />
<br />
In this work, the authors argue that the final classification layer in deep neural networks is redundant and suggest removing the parameters from the classification layer. The empirical results from experiments on the CIFAR and IMAGENET datasets suggest that such a change lead to little or almost no decline in the performance of the architecture. Furthermore, using a Hadmard matrix as classifier might lead to some computational benefits when properly implemented, and save memory otherwise spent on large amount of transformation coefficients.<br />
<br />
Another possible scope of research that could be pointed out for future could be to find new efficient methods to create pre-defined word embeddings, which require huge amount of parameters that can possibly be avoided when learning a new task. Therefore, more emphasis should be given to the representations learned by the non-linear parts of the neural networks - upto the final classifier, as it seems highly redundant.<br />
<br />
=Critique=<br />
<br />
The paper proposes an interesting idea that has a potential use case when designing memory-efficient neural networks. The experiments shown in the paper are quite rigorous and provide support to the authors' claim. However, it would have been more helpful if the authors had described a bit more about efficient implementation of the Hadamard matrix and how to scale this method for larger datasets (cases with <math> C >N</math>).<br />
<br />
The paper presents a very interesting idea which can be applied for most of networks in deep learning area. However, technical proofs of the effect of the algorithm should be considered in order to generalize and be appreciated further.<br />
<br />
=References=<br />
<br />
The code for the proposed model is available at https://github.com/eladhoffer/fix_your_classifier.<br />
<br />
Madhu S Advani and Andrew M Saxe. High-dimensional dynamics of generalization error in neural networks. arXiv preprint arXiv:1710.03667, 2017.<br />
<br />
Peter Bartlett, Dylan J Foster, and Matus Telgarsky. Spectrally-normalized margin bounds for neural networks. arXiv preprint arXiv:1706.08498, 2017.<br />
<br />
Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Sackinger, and Roopak Shah. Signature verification using a” siamese” time delay neural network. In Advances in Neural Information Processing Systems, pp. 737–744, 1994.<br />
<br />
Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In Advances in Neural Information Processing Systems, pp. 3123–3131, 2015.<br />
<br />
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pp. 248–255. IEEE, 2009.<br />
<br />
Suriya Gunasekar, Blake Woodworth, Srinadh Bhojanapalli, Behnam Neyshabur, and Nathan Srebro. Implicit regularization in matrix factorization. arXiv preprint arXiv:1705.09280, 2017.<br />
<br />
Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015.<br />
<br />
Moritz Hardt and Tengyu Ma. Identity matters in deep learning. 2017.<br />
<br />
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016.<br />
<br />
A Hedayat, WD Wallis, et al. Hadamard matrices and their applications. The Annals of Statistics, 6<br />
(6):1184–1238, 1978.<br />
<br />
Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8): 1735–1780, 1997.<br />
<br />
Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In International Workshop on Similarity-Based Pattern Recognition, pp. 84–92. Springer, 2015.<br />
<br />
Elad Hoffer, Itay Hubara, and Daniel Soudry. Train longer, generalize better: closing the generalization gap in large batch training of neural networks. 2017.<br />
<br />
Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017.<br />
<br />
Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017.<br />
<br />
Guang-Bin Huang, Qin-Yu Zhu, and Chee-Kheong Siew. Extreme learning machine: theory and applications. Neurocomputing, 70(1):489–501, 2006.<br />
<br />
Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized neural networks. In Advances in Neural Information Processing Systems 29 (NIPS’16), 2016a.<br />
<br />
Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Quantized neural networks: Training neural networks with low precision weights and activations. arXiv preprint arXiv:1609.07061, 2016b.<br />
<br />
Hakan Inan, Khashayar Khosravi, and Richard Socher. Tying word vectors and word classifiers: A loss framework for language modeling. arXiv preprint arXiv:1611.01462, 2016.<br />
<br />
Max Jaderberg, Andrea Vedaldi, and Andrew Zisserman. Speeding up convolutional neural networks with low rank expansions. arXiv preprint arXiv:1405.3866, 2014.<br />
<br />
Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009.<br />
<br />
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012.<br />
<br />
Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´ document recognition. Proceedings of the IEEE, 86(11):2278 2324, 1998.<br />
<br />
Omer Levy and Yoav Goldberg. Neural word embedding as implicit matrix factorization. In Advances in neural information processing systems, pp. 2177–2185, 2014.<br />
<br />
Fengfu Li, Bo Zhang, and Bin Liu. Ternary weight networks. arXiv preprint arXiv:1605.04711, 2016.<br />
<br />
Min Lin, Qiang Chen, and Shuicheng Yan. Network in network. arXiv preprint arXiv:1312.4400, 2013.<br />
<br />
Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016.<br />
<br />
Stephen Merity, Nitish Shirish Keskar, and Richard Socher. Regularizing and Optimizing LSTM Language Models. arXiv preprint arXiv:1708.02182, 2017.<br />
<br />
Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaev, Ganesh Venkatesh, et al. Mixed precision training. arXiv preprint arXiv:1710.03740, 2017.<br />
<br />
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed tations of words and phrases and their compositionality. In Advances in neural information processing systems, pp. 3111–3119, 2013.<br />
<br />
Behnam Neyshabur, Srinadh Bhojanapalli, David McAllester, and Nathan Srebro. Exploring generalization in deep learning. arXiv preprint arXiv:1706.08947, 2017.<br />
Jooyoung Park and Irwin W Sandberg. Universal approximation using radial-basis-function networks. Neural computation, 3(2):246–257, 1991.<br />
<br />
Ofir Press and Lior Wolf. Using the output embedding to improve language models. EACL 2017,<br />
pp. 157, 2017.<br />
<br />
Itay Safran and Ohad Shamir. On the quality of the initial basin in overspecified neural networks. In International Conference on Machine Learning, pp. 774–782, 2016.<br />
<br />
Tim Salimans and Diederik P Kingma. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. In Advances in Neural Information Processing Systems, pp. 901–909, 2016.<br />
<br />
Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 815–823, 2015.<br />
<br />
Mahdi Soltanolkotabi, Adel Javanmard, and Jason D Lee. Theoretical insights into the optimization landscape of over-parameterized shallow neural networks. arXiv preprint arXiv:1707.04926, 2017.<br />
<br />
Daniel Soudry and Yair Carmon. No bad local minima: Data independent training error guarantees for multilayer neural networks. arXiv preprint arXiv:1605.08361, 2016.<br />
<br />
Daniel Soudry and Elad Hoffer. Exponentially vanishing sub-optimal local minima in multilayer neural networks. arXiv preprint arXiv:1702.05777, 2017.<br />
<br />
Daniel Soudry, Elad Hoffer, and Nathan Srebro. The implicit bias of gradient descent on separable data. 2018.<br />
<br />
Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806, 2014.<br />
<br />
Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. arXiv preprint arXiv:1707.02968, 2017.<br />
<br />
Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1–9, 2015.<br />
<br />
Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2818–2826, 2016.<br />
<br />
Cheng Tai, Tong Xiao, Yi Zhang, Xiaogang Wang, et al. Convolutional neural networks with lowrank regularization. arXiv preprint arXiv:1511.06067, 2015.<br />
<br />
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. 2017.<br />
Ashia C Wilson, Rebecca Roelofs, Mitchell Stern, Nathan Srebro, and Benjamin Recht. The marginal value of adaptive gradient methods in machine learning. arXiv preprint arXiv:1705.08292, 2017.<br />
<br />
Bo Xie, Yingyu Liang, and Le Song. Diversity leads to generalization in neural networks. arXiv preprint arXiv:1611.03131, 2016.<br />
<br />
Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In European conference on computer vision, pp. 818–833. Springer, 2014. Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. In ICLR, 2017a. URL https://arxiv.org/abs/1611.03530.<br />
<br />
Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. arXiv preprint arXiv:1707.01083, 2017b.<br />
<br />
Shuchang Zhou, Zekun Ni, Xinyu Zhou, He Wen, Yuxin Wu, and Yuheng Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160, 2016.<br />
<br />
A. Rosenfeld and J. K. Tsotsos, “Intriguing properties of randomly weighted networks: Generalizing while learning next to nothing,” arXiv preprint arXiv:1802.00844, 2018.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Fix_your_classifier:_the_marginal_value_of_training_the_last_weight_layer&diff=41924Fix your classifier: the marginal value of training the last weight layer2018-11-29T23:02:54Z<p>Ka2khan: /* Choosing the projection matrix */</p>
<hr />
<div>=Introduction=<br />
<br />
Deep neural networks have become a widely used model for machine learning, achieving state-of-the-art results on many tasks. The most common task these models are used for is to perform classification, as in the case of convolutional neural networks (CNNs) used to classify images to a semantic category. Typically, a learned affine transformation is placed at the end of such models, yielding a per-class value used for classification. This classifier can have<br />
a vast number of parameters, which grows linearly with the number of possible classes, thus requiring increasingly more computational resources.<br />
<br />
=Brief Overview=<br />
<br />
In order to alleviate the aforementioned problem, the authors propose that the final layer of the classifier be fixed (upto a global scale constant). They argue that with little or no loss of accuracy for most classification tasks, the method provides significant memory and computational benefits. In addition, they show that by initializing the classifier with a Hadamard matrix the inference could be made faster as well.<br />
<br />
=Previous Work=<br />
<br />
Training NN models and using them for inference requires large amounts of memory and computational resources; thus, extensive amount of research has been done lately to reduce the size of networks which are as follows:<br />
<br />
* Weight sharing and specification (Han et al., 2015)<br />
<br />
* Mixed precision to reduce the size of the neural networks by half (Micikevicius et al., 2017)<br />
<br />
* Low-rank approximations to speed up CNN (Tai et al., 2015)<br />
<br />
* Quantization of weights, activations and gradients to further reduce computation during training (Hubara et al., 2016b; Li et al., 2016 and Zhou et al., 2016)<br />
<br />
Some of the past works have also put forward the fact that predefined (Park & Sandberg, 1991) and random (Huang et al., 2006) projections can be used together with a learned affine transformation to achieve competitive results on many of the classification tasks. However, the authors' proposal in the current paper is quite reversed.<br />
<br />
=Background=<br />
<br />
Convolutional neural networks (CNNs) are commonly used to solve a variety of spatial and temporal tasks. CNNs are usually composed of a stack of convolutional parameterized layers, spatial pooling layers and fully connected layers, separated by non-linear activation functions. Earlier architectures of CNNs (LeCun et al., 1998; Krizhevsky et al., 2012) used a set of fully-connected layers at later stage of the network, presumably to allow classification based on global features of an image.<br />
<br />
== Shortcomings of the final classification layer and its solution ==<br />
<br />
Despite the enormous number of trainable parameters these layers added to the model, they are known to have a rather marginal impact on the final performance of the network (Zeiler & Fergus, 2014).<br />
<br />
It has been shown previously that these layers could be easily compressed and reduced after a model was trained by simple means such as matrix decomposition and sparsification (Han et al., 2015). Modern architecture choices are characterized with the removal of most of the fully connected layers (Lin et al., 2013; Szegedy et al., 2015; He et al., 2016), that lead to better generalization and overall accuracy, together with a huge decrease in the number of trainable parameters. Additionally, numerous works showed that CNNs can be trained in a metric learning regime (Bromley et al., 1994; Schroff et al., 2015; Hoffer & Ailon, 2015), where no explicit classification layer was introduced and the objective regarded only distance measures between intermediate representations. Hardt & Ma (2017) suggested an all-convolutional network variant, where they kept the original initialization of the classification layer fixed with no negative impact on performance on the CIFAR-10 dataset.<br />
<br />
=Proposed Method=<br />
<br />
The aforementioned works provide evidence that fully-connected layers are in fact redundant and play a small role in learning and generalization. In this work, the authors have suggested that parameters used for the final classification transform are completely redundant, and can be replaced with a predetermined linear transform. This holds for even in large-scale models and classification tasks, such as recent architectures trained on the ImageNet benchmark (Deng et al., 2009).<br />
<br />
==Using a fixed classifier==<br />
<br />
Suppose the final representation obtained by the network (the last hidden layer) is represented as <math>x = F(z;\theta)</math> where <math>F</math> is assumed to be a deep neural network with input z and parameters θ, e.g., a convolutional network, trained by backpropagation.<br />
<br />
In common NN models, this representation is followed by an additional affine transformation, <math>y = W^T x + b</math> ,where <math>W</math> and <math>b</math> are also trained by back-propagation.<br />
<br />
For input <math>x</math> of <math>N</math> length, and <math>C</math> different possible outputs, <math>W</math> is required to be a matrix of <math>N ×<br />
C</math>. Training is done using cross-entropy loss, by feeding the network outputs through a softmax activation<br />
<br />
<math><br />
v_i = \frac{e^{y_i}}{\sum_{j}^{C}{e^{y_j}}}, i &isin; </math> { <math> {1, . . . , C} </math> }<br />
<br />
and reducing the expected negative log likelihood with respect to ground-truth target <math> t &isin; </math> { <math> {1, . . . , C} </math> },<br />
by minimizing the loss function:<br />
<br />
<math><br />
L(x, t) = −\text{log}\ {v_t} = −{w_t}·{x} − b_t + \text{log} ({\sum_{j}^{C}e^{w_j . x + b_j}})<br />
</math><br />
<br />
where <math>w_i</math> is the <math>i</math>-th column of <math>W</math>.<br />
<br />
==Choosing the projection matrix==<br />
<br />
To evaluate the conjecture regarding the importance of the final classification transformation, the trainable parameter matrix <math>W</math> is replaced with a fixed orthonormal projection <math> Q &isin; R^{N×C} </math>, such that <math> &forall; i &ne; j : q_i · q_j = 0 </math> and <math> || q_i ||_{2} = 1 </math>, where <math>q_i</math> is the <math>i</math>th column of <math>Q</math>. This is ensured by a simple random sampling and singular-value decomposition<br />
<br />
As the rows of classifier weight matrix are fixed with an equally valued <math>L_{2}</math> norm, we find it beneficial<br />
to also restrict the representation of <math>x</math> by normalizing it to reside on the <math>n</math>-dimensional sphere:<br />
<br />
<center><math><br />
\hat{x} = \frac{x}{||x||_{2}}<br />
</math></center><br />
<br />
This allows faster training and convergence, as the network does not need to account for changes in the scale of its weights. However, it has now an issue that <math>q_i · \hat{x} </math> is bounded between −1 and 1. This causes convergence issues, as the softmax function is scale sensitive, and the network is affected by the inability to re-scale its input. This issue is amended with a fixed scale <math>T</math> applied to softmax inputs <math>f(y) = softmax(\frac{1}{T}y)</math>, also known as the ''softmax temperature''. However, this introduces an additional hyper-parameter which may differ between networks and datasets. So, the authors propose to introduce a single scalar parameter <math>\alpha</math> to learn the softmax scale, effectively functioning as an inverse of the softmax temperature <math>\frac{1}{T}</math>. The normalized weights and an additional scale coefficient are also used, specially using a single scale for all entries in the weight matrix. The additional vector of bias parameters <math>b &isin; \mathbb{R}^{C}</math> is kept the same and the model is trained using the traditional negative-log-likelihood criterion. Explicitly, the classifier output is now:<br />
<br />
<center><br />
<math><br />
v_i=\frac{e^{\alpha q_i &middot; \hat{x} + b_i}}{\sum_{j}^{C} e^{\alpha q_j &middot; \hat{x} + b_j}}, i &isin; </math> { <math> {1,...,C} </math>}<br />
</center><br />
<br />
and the loss to be minimized is:<br />
<br />
<center><math><br />
L(x, t) = -\alpha q_t &middot; \frac{x}{||x||_{2}} + b_t + \text{log} (\sum_{i=1}^{C} \text{exp}((\alpha q_i &middot; \frac{x}{||x||_{2}} + b_i)))<br />
</math></center><br />
<br />
where <math>x</math> is the final representation obtained by the network for a specific sample, and <math> t &isin; </math> { <math> {1, . . . , C} </math> } is the ground-truth label for that sample. The behaviour of the parameter <math> \alpha </math> over time, which is logarithmic in nature and has the same behavior exhibited by the norm of a learned classifier, is shown in<br />
[[Media: figure1_log_behave.png| Figure 1]].<br />
<br />
<center>[[File:figure1_log_behave.png]]</center><br />
<br />
When <math> -1 \le q_i · \hat{x} \le 1 </math>, a possible cosine angle loss is <br />
<br />
<center>[[File:caloss.png]]</center><br />
<br />
But its final validation accuracy has slight decrease, compared to original models.<br />
<br />
==Using a Hadmard matrix==<br />
<br />
To recall, Hadmard matrix (Hedayat et al., 1978) <math> H </math> is an <math> n × n </math> matrix, where all of its entries are either +1 or −1.<br />
Furthermore, <math> H </math> is orthogonal, such that <math> HH^{T} = nI_n </math> where <math>I_n</math> is the identity matrix. Instead of using the entire Hadmard matrix <math>H</math>, a truncated version, <math> \hat{H} &isin; </math> {<math> {-1, 1}</math>}<math>^{C \times N}</math> where all <math>C</math> rows are orthogonal as the final classification layer is such that:<br />
<br />
<center><math><br />
y = \hat{H} \hat{x} + b<br />
</math></center><br />
<br />
This usage allows two main benefits:<br />
* A deterministic, low-memory and easily generated matrix that can be used for classification.<br />
* Removal of the need to perform a full matrix-matrix multiplication - as multiplying by a Hadamard matrix can be done by simple sign manipulation and addition.<br />
<br />
Here, <math>n</math> must be a multiple of 4, but it can be easily truncated to fit normally defined networks. Also, as the classifier weights are fixed to need only 1-bit precision, it is now possible to focus our attention on the features preceding it.<br />
<br />
=Experimental Results=<br />
<br />
The authors have evaluated their proposed model on the following datasets:<br />
<br />
==CIFAR-10/100==<br />
<br />
===About the dataset===<br />
<br />
CIFAR-10 is an image classification benchmark dataset containing 50,000 training images and 10,000 test images. The images are in color and contain 32×32 pixels. There are 10 possible classes of various animals and vehicles. CIFAR-100 holds the same number of images of same size, but contains 100 different classes.<br />
<br />
===Training Details===<br />
<br />
The authors trained a residual network ( He et al., 2016) on the CIFAR-10 dataset. The network depth was 56 and the same hyper-parameters as in the original work were used. A comparison of the two variants, i.e., the learned classifier and the proposed classifier with a fixed transformation is shown in [[Media: figure1_resnet_cifar10.png | Figure 2]].<br />
<br />
<center>[[File: figure1_resnet_cifar10.png]]</center><br />
<br />
These results demonstrate that although the training error is considerably lower for the network with learned classifier, both models achieve the same classification accuracy on the validation set. The authors conjecture is that with the new fixed parameterization, the network can no longer increase the<br />
norm of a given sample’s representation - thus learning its label requires more effort. As this may happen for specific seen samples - it affects only training error.<br />
<br />
The authors also compared using a fixed scale variable <math>\alpha </math> at different values vs. the learned parameter. Results for <math> \alpha = </math> {0.1, 1, 10} are depicted in [[Media: figure3_alpha_resnet_cifar.png| Figure 3]] for both training and validation error and as can be seen, similar validation accuracy can be obtained using a fixed scale value (in this case <math>\alpha </math>= 1 or 10 will suffice) at the expense of another hyper-parameter to seek. In all the further experiments the scaling parameter <math> \alpha </math> was regularized with the same weight decay coefficient used on original classifier.<br />
<br />
<center>[[File: figure3_alpha_resnet_cifar.png]]</center><br />
<br />
The authors then train the model on CIFAR-100 dataset. They used the DenseNet-BC model from Huang et al. (2017) with depth of 100 layers and k = 12. The higher number of classes caused the number of parameters to grow and encompassed about 4% of the whole model. However, validation accuracy for the fixed-classifier model remained equally good as the original model, and the same training curve was observed as earlier.<br />
<br />
==IMAGENET==<br />
<br />
===About the dataset===<br />
<br />
The Imagenet dataset introduced by Deng et al. (2009) spans over 1000 visual classes, and over 1.2 million samples. This is supposedly a more challenging dataset to work on as compared to CIFAR-10/100.<br />
<br />
===Experiment Details===<br />
<br />
The authors evaluated their fixed classifier method on Imagenet using Resnet50 by He et al. (2016) and Densenet169 model (Huang et al., 2017) as described in the original work. Using a fixed classifier removed approximately 2-million parameters were from the model, accounting for about 8% and 12 % of the model parameters respectively. The experiments revealed similar trends as observed on CIFAR-10.<br />
<br />
For a more stricter evaluation, the authors also trained a Shufflenet architecture (Zhang et al., 2017b), which was designed to be used in low memory and limited computing platforms and has parameters making up the majority of the model. They were able to reduce the parameters to 0.86 million as compared to 0.96 million parameters in the final layer of the original model. Again, the proposed modification in the original model gave similar convergence results on validation accuracy.<br />
<br />
The overall results of the fixed-classifier are summarized in [[Media: table1_fixed_results.png | Table 1]].<br />
<br />
<center>[[File: table1_fixed_results.png]]</center><br />
<br />
==Language Modelling==<br />
<br />
Recent works have empirically found that using the same weights for both word embedding and classifier can yield equal or better results than using a separate pair of weights. So the authors experimented with fix-classifiers on language modelling as it also requires classification of all possible tokens available in the task vocabulary. They trained a recurrent model with 2-layers of LSTM (Hochreiter & Schmidhuber, 1997) and embedding + hidden size of 512 on the WikiText2 dataset (Merity et al., 2016), using same settings as in Merity et al. (2017). However, using a random orthogonal transform yielded poor results compared to learned embedding. This was suspected to be due to semantic relationships captured in the embedding layer of language models, which is not the case in image classification task. The intuition was further confirmed by the much better results when pre-trained embeddings using word2vec algorithm by Mikolov et al. (2013) or PMI factorization as suggested by Levy & Goldberg (2014), were used.<br />
<br />
=Discussion=<br />
<br />
==Implications and use cases==<br />
<br />
With the increasing number of classes in the benchmark datasets, computational demands for the final classifier will increase as well. In order to understand the problem better, the authors observe the work by Sun et al. (2017), which introduced JFT-300M - an internal Google dataset with over 18K different classes. Using a Resnet50 (He et al., 2016), with a 2048 sized representation led to a model with over 36M parameters meaning that over 60% of the model parameters resided in the final classification layer. Sun et al. (2017) also describe the difficulty in distributing so many parameters over the training servers involving a non-trivial overhead during synchronization of the model for update. The authors claim that the fixed-classifier would help considerably in this kind of scenario - where using a fixed classifier removes the need to do any gradient synchronization for the final layer. Furthermore, introduction of Hadamard matrix removes the need to save the transformation altogether, thereby, making it more efficient and allowing considerable memory and computational savings.<br />
<br />
==Possible Caveats==<br />
<br />
The good performance of fixed-classifiers relies on the ability of the preceding layers to learn separable representations. This could be affected when when the ratio between learned features and number of classes is small – that is, when <math> C > N</math>. However, they tested their method in such cases and their model performed well and provided good results.<br />
Another factor that can affect the performance of their model using a fixed classifier is when the classes are highly correlated. In that case, the fixed classifier actually cannot support correlated classes and thus, the network could have some difficulty to learn. For a language model, word classes tend to have highly correlated instances, which also lead to difficult learning process.<br />
<br />
==Future Work==<br />
<br />
<br />
The use of fixed classifiers might be further simplified in Binarized Neural Networks (Hubara et al., 2016a), where the activations and weights are restricted to ±1 during propagations. In that case the norm of the last hidden layer would be constant for all samples (equal to the square root of the hidden layer width). The constant could then be absorbed into the scale constant <math>\alpha</math>, and there is no need in a per-sample normalization.<br />
<br />
Additionally, more efficient ways to learn a word embedding should also be explored where similar redundancy in classifier weights may suggest simpler forms of token representations - such as low-rank or sparse versions.<br />
<br />
A related paper was published that claims that fixing most of the parameters of the neural network achieves comparable results with learning all of them [A. Rosenfeld and J. K. Tsotsos]<br />
<br />
=Conclusion=<br />
<br />
In this work, the authors argue that the final classification layer in deep neural networks is redundant and suggest removing the parameters from the classification layer. The empirical results from experiments on the CIFAR and IMAGENET datasets suggest that such a change lead to little or almost no decline in the performance of the architecture. Furthermore, using a Hadmard matrix as classifier might lead to some computational benefits when properly implemented, and save memory otherwise spent on large amount of transformation coefficients.<br />
<br />
Another possible scope of research that could be pointed out for future could be to find new efficient methods to create pre-defined word embeddings, which require huge amount of parameters that can possibly be avoided when learning a new task. Therefore, more emphasis should be given to the representations learned by the non-linear parts of the neural networks - upto the final classifier, as it seems highly redundant.<br />
<br />
=Critique=<br />
<br />
The paper proposes an interesting idea that has a potential use case when designing memory-efficient neural networks. The experiments shown in the paper are quite rigorous and provide support to the authors' claim. However, it would have been more helpful if the authors had described a bit more about efficient implementation of the Hadamard matrix and how to scale this method for larger datasets (cases with <math> C >N</math>).<br />
<br />
The paper presents a very interesting idea which can be applied for most of networks in deep learning area. However, technical proofs of the effect of the algorithm should be considered in order to generalize and be appreciated further.<br />
<br />
=References=<br />
<br />
The code for the proposed model is available at https://github.com/eladhoffer/fix_your_classifier.<br />
<br />
Madhu S Advani and Andrew M Saxe. High-dimensional dynamics of generalization error in neural networks. arXiv preprint arXiv:1710.03667, 2017.<br />
<br />
Peter Bartlett, Dylan J Foster, and Matus Telgarsky. Spectrally-normalized margin bounds for neural networks. arXiv preprint arXiv:1706.08498, 2017.<br />
<br />
Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Sackinger, and Roopak Shah. Signature verification using a” siamese” time delay neural network. In Advances in Neural Information Processing Systems, pp. 737–744, 1994.<br />
<br />
Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In Advances in Neural Information Processing Systems, pp. 3123–3131, 2015.<br />
<br />
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pp. 248–255. IEEE, 2009.<br />
<br />
Suriya Gunasekar, Blake Woodworth, Srinadh Bhojanapalli, Behnam Neyshabur, and Nathan Srebro. Implicit regularization in matrix factorization. arXiv preprint arXiv:1705.09280, 2017.<br />
<br />
Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015.<br />
<br />
Moritz Hardt and Tengyu Ma. Identity matters in deep learning. 2017.<br />
<br />
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016.<br />
<br />
A Hedayat, WD Wallis, et al. Hadamard matrices and their applications. The Annals of Statistics, 6<br />
(6):1184–1238, 1978.<br />
<br />
Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8): 1735–1780, 1997.<br />
<br />
Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In International Workshop on Similarity-Based Pattern Recognition, pp. 84–92. Springer, 2015.<br />
<br />
Elad Hoffer, Itay Hubara, and Daniel Soudry. Train longer, generalize better: closing the generalization gap in large batch training of neural networks. 2017.<br />
<br />
Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017.<br />
<br />
Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017.<br />
<br />
Guang-Bin Huang, Qin-Yu Zhu, and Chee-Kheong Siew. Extreme learning machine: theory and applications. Neurocomputing, 70(1):489–501, 2006.<br />
<br />
Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized neural networks. In Advances in Neural Information Processing Systems 29 (NIPS’16), 2016a.<br />
<br />
Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Quantized neural networks: Training neural networks with low precision weights and activations. arXiv preprint arXiv:1609.07061, 2016b.<br />
<br />
Hakan Inan, Khashayar Khosravi, and Richard Socher. Tying word vectors and word classifiers: A loss framework for language modeling. arXiv preprint arXiv:1611.01462, 2016.<br />
<br />
Max Jaderberg, Andrea Vedaldi, and Andrew Zisserman. Speeding up convolutional neural networks with low rank expansions. arXiv preprint arXiv:1405.3866, 2014.<br />
<br />
Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009.<br />
<br />
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012.<br />
<br />
Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´ document recognition. Proceedings of the IEEE, 86(11):2278 2324, 1998.<br />
<br />
Omer Levy and Yoav Goldberg. Neural word embedding as implicit matrix factorization. In Advances in neural information processing systems, pp. 2177–2185, 2014.<br />
<br />
Fengfu Li, Bo Zhang, and Bin Liu. Ternary weight networks. arXiv preprint arXiv:1605.04711, 2016.<br />
<br />
Min Lin, Qiang Chen, and Shuicheng Yan. Network in network. arXiv preprint arXiv:1312.4400, 2013.<br />
<br />
Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016.<br />
<br />
Stephen Merity, Nitish Shirish Keskar, and Richard Socher. Regularizing and Optimizing LSTM Language Models. arXiv preprint arXiv:1708.02182, 2017.<br />
<br />
Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaev, Ganesh Venkatesh, et al. Mixed precision training. arXiv preprint arXiv:1710.03740, 2017.<br />
<br />
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed tations of words and phrases and their compositionality. In Advances in neural information processing systems, pp. 3111–3119, 2013.<br />
<br />
Behnam Neyshabur, Srinadh Bhojanapalli, David McAllester, and Nathan Srebro. Exploring generalization in deep learning. arXiv preprint arXiv:1706.08947, 2017.<br />
Jooyoung Park and Irwin W Sandberg. Universal approximation using radial-basis-function networks. Neural computation, 3(2):246–257, 1991.<br />
<br />
Ofir Press and Lior Wolf. Using the output embedding to improve language models. EACL 2017,<br />
pp. 157, 2017.<br />
<br />
Itay Safran and Ohad Shamir. On the quality of the initial basin in overspecified neural networks. In International Conference on Machine Learning, pp. 774–782, 2016.<br />
<br />
Tim Salimans and Diederik P Kingma. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. In Advances in Neural Information Processing Systems, pp. 901–909, 2016.<br />
<br />
Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 815–823, 2015.<br />
<br />
Mahdi Soltanolkotabi, Adel Javanmard, and Jason D Lee. Theoretical insights into the optimization landscape of over-parameterized shallow neural networks. arXiv preprint arXiv:1707.04926, 2017.<br />
<br />
Daniel Soudry and Yair Carmon. No bad local minima: Data independent training error guarantees for multilayer neural networks. arXiv preprint arXiv:1605.08361, 2016.<br />
<br />
Daniel Soudry and Elad Hoffer. Exponentially vanishing sub-optimal local minima in multilayer neural networks. arXiv preprint arXiv:1702.05777, 2017.<br />
<br />
Daniel Soudry, Elad Hoffer, and Nathan Srebro. The implicit bias of gradient descent on separable data. 2018.<br />
<br />
Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806, 2014.<br />
<br />
Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. arXiv preprint arXiv:1707.02968, 2017.<br />
<br />
Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1–9, 2015.<br />
<br />
Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2818–2826, 2016.<br />
<br />
Cheng Tai, Tong Xiao, Yi Zhang, Xiaogang Wang, et al. Convolutional neural networks with lowrank regularization. arXiv preprint arXiv:1511.06067, 2015.<br />
<br />
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. 2017.<br />
Ashia C Wilson, Rebecca Roelofs, Mitchell Stern, Nathan Srebro, and Benjamin Recht. The marginal value of adaptive gradient methods in machine learning. arXiv preprint arXiv:1705.08292, 2017.<br />
<br />
Bo Xie, Yingyu Liang, and Le Song. Diversity leads to generalization in neural networks. arXiv preprint arXiv:1611.03131, 2016.<br />
<br />
Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In European conference on computer vision, pp. 818–833. Springer, 2014. Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. In ICLR, 2017a. URL https://arxiv.org/abs/1611.03530.<br />
<br />
Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. arXiv preprint arXiv:1707.01083, 2017b.<br />
<br />
Shuchang Zhou, Zekun Ni, Xinyu Zhou, He Wen, Yuxin Wu, and Yuheng Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160, 2016.<br />
<br />
A. Rosenfeld and J. K. Tsotsos, “Intriguing properties of randomly weighted networks: Generalizing while learning next to nothing,” arXiv preprint arXiv:1802.00844, 2018.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Fix_your_classifier:_the_marginal_value_of_training_the_last_weight_layer&diff=41920Fix your classifier: the marginal value of training the last weight layer2018-11-29T22:59:12Z<p>Ka2khan: /* Using a fixed classifier */</p>
<hr />
<div>=Introduction=<br />
<br />
Deep neural networks have become a widely used model for machine learning, achieving state-of-the-art results on many tasks. The most common task these models are used for is to perform classification, as in the case of convolutional neural networks (CNNs) used to classify images to a semantic category. Typically, a learned affine transformation is placed at the end of such models, yielding a per-class value used for classification. This classifier can have<br />
a vast number of parameters, which grows linearly with the number of possible classes, thus requiring increasingly more computational resources.<br />
<br />
=Brief Overview=<br />
<br />
In order to alleviate the aforementioned problem, the authors propose that the final layer of the classifier be fixed (upto a global scale constant). They argue that with little or no loss of accuracy for most classification tasks, the method provides significant memory and computational benefits. In addition, they show that by initializing the classifier with a Hadamard matrix the inference could be made faster as well.<br />
<br />
=Previous Work=<br />
<br />
Training NN models and using them for inference requires large amounts of memory and computational resources; thus, extensive amount of research has been done lately to reduce the size of networks which are as follows:<br />
<br />
* Weight sharing and specification (Han et al., 2015)<br />
<br />
* Mixed precision to reduce the size of the neural networks by half (Micikevicius et al., 2017)<br />
<br />
* Low-rank approximations to speed up CNN (Tai et al., 2015)<br />
<br />
* Quantization of weights, activations and gradients to further reduce computation during training (Hubara et al., 2016b; Li et al., 2016 and Zhou et al., 2016)<br />
<br />
Some of the past works have also put forward the fact that predefined (Park & Sandberg, 1991) and random (Huang et al., 2006) projections can be used together with a learned affine transformation to achieve competitive results on many of the classification tasks. However, the authors' proposal in the current paper is quite reversed.<br />
<br />
=Background=<br />
<br />
Convolutional neural networks (CNNs) are commonly used to solve a variety of spatial and temporal tasks. CNNs are usually composed of a stack of convolutional parameterized layers, spatial pooling layers and fully connected layers, separated by non-linear activation functions. Earlier architectures of CNNs (LeCun et al., 1998; Krizhevsky et al., 2012) used a set of fully-connected layers at later stage of the network, presumably to allow classification based on global features of an image.<br />
<br />
== Shortcomings of the final classification layer and its solution ==<br />
<br />
Despite the enormous number of trainable parameters these layers added to the model, they are known to have a rather marginal impact on the final performance of the network (Zeiler & Fergus, 2014).<br />
<br />
It has been shown previously that these layers could be easily compressed and reduced after a model was trained by simple means such as matrix decomposition and sparsification (Han et al., 2015). Modern architecture choices are characterized with the removal of most of the fully connected layers (Lin et al., 2013; Szegedy et al., 2015; He et al., 2016), that lead to better generalization and overall accuracy, together with a huge decrease in the number of trainable parameters. Additionally, numerous works showed that CNNs can be trained in a metric learning regime (Bromley et al., 1994; Schroff et al., 2015; Hoffer & Ailon, 2015), where no explicit classification layer was introduced and the objective regarded only distance measures between intermediate representations. Hardt & Ma (2017) suggested an all-convolutional network variant, where they kept the original initialization of the classification layer fixed with no negative impact on performance on the CIFAR-10 dataset.<br />
<br />
=Proposed Method=<br />
<br />
The aforementioned works provide evidence that fully-connected layers are in fact redundant and play a small role in learning and generalization. In this work, the authors have suggested that parameters used for the final classification transform are completely redundant, and can be replaced with a predetermined linear transform. This holds for even in large-scale models and classification tasks, such as recent architectures trained on the ImageNet benchmark (Deng et al., 2009).<br />
<br />
==Using a fixed classifier==<br />
<br />
Suppose the final representation obtained by the network (the last hidden layer) is represented as <math>x = F(z;\theta)</math> where <math>F</math> is assumed to be a deep neural network with input z and parameters θ, e.g., a convolutional network, trained by backpropagation.<br />
<br />
In common NN models, this representation is followed by an additional affine transformation, <math>y = W^T x + b</math> ,where <math>W</math> and <math>b</math> are also trained by back-propagation.<br />
<br />
For input <math>x</math> of <math>N</math> length, and <math>C</math> different possible outputs, <math>W</math> is required to be a matrix of <math>N ×<br />
C</math>. Training is done using cross-entropy loss, by feeding the network outputs through a softmax activation<br />
<br />
<math><br />
v_i = \frac{e^{y_i}}{\sum_{j}^{C}{e^{y_j}}}, i &isin; </math> { <math> {1, . . . , C} </math> }<br />
<br />
and reducing the expected negative log likelihood with respect to ground-truth target <math> t &isin; </math> { <math> {1, . . . , C} </math> },<br />
by minimizing the loss function:<br />
<br />
<math><br />
L(x, t) = −\text{log}\ {v_t} = −{w_t}·{x} − b_t + \text{log} ({\sum_{j}^{C}e^{w_j . x + b_j}})<br />
</math><br />
<br />
where <math>w_i</math> is the <math>i</math>-th column of <math>W</math>.<br />
<br />
==Choosing the projection matrix==<br />
<br />
To evaluate the conjecture regarding the importance of the final classification transformation, the trainable parameter matrix <math>W</math> is replaced with a fixed orthonormal projection <math> Q &isin; R^{N×C} </math>, such that <math> &forall; i &ne; j : q_i · q_j = 0 </math> and <math> || q_i ||_{2} = 1 </math>, where <math>q_i</math> is the <math>i</math>th column of <math>Q</math>. This is ensured by a simple random sampling and singular-value decomposition<br />
<br />
As the rows of classifier weight matrix are fixed with an equally valued <math>L_{2}</math> norm, we find it beneficial<br />
to also restrict the representation of <math>x</math> by normalizing it to reside on the <math>n</math>-dimensional sphere:<br />
<br />
<center><math><br />
\hat{x} = \frac{x}{||x||_{2}}<br />
</math></center><br />
<br />
This allows faster training and convergence, as the network does not need to account for changes in the scale of its weights. However, it has now an issue that <math>q_i · \hat{x} </math> is bounded between −1 and 1. This causes convergence issues, as the softmax function is scale sensitive, and the network is affected by the inability to re-scale its input. This issue is amended with a fixed scale <math>T</math> applied to softmax inputs <math>f(y) = softmax(\frac{1}{T}y)</math>, also known as the ''softmax temperature''. However, this introduces an additional hyper-parameter which may differ between networks and datasets. So, the authors propose to introduce a single scalar parameter <math>\alpha</math> to learn the softmax scale, effectively functioning as an inverse of the softmax temperature <math>\frac{1}{T}</math>. The normalized weights and an additional scale coefficient are also used, specially using a single scale for all entries in the weight matrix. The additional vector of bias parameters <math>b &isin; R^{C}</math> is kept the same and the model is trained using the traditional negative-log-likelihood criterion. Explicitly, the classifier output is now:<br />
<br />
<center><br />
<math><br />
v_i=\frac{e^{\alpha q_i &middot; \hat{x} + b_i}}{\sum_{j}^{C} e^{\alpha q_j &middot; \hat{x} + b_j}}, i &isin; </math> { <math> {1,...,C} </math>}<br />
</center><br />
<br />
and the loss to be minimized is:<br />
<br />
<center><math><br />
L(x, t) = -\alpha q_t &middot; \frac{x}{||x||_{2}} + b_t + log (\sum_{i=1}^{C} exp((\alpha q_i &middot; \frac{x}{||x||_{2}} + b_i)))<br />
</math></center><br />
<br />
where <math>x</math> is the final representation obtained by the network for a specific sample, and <math> t &isin; </math> { <math> {1, . . . , C} </math> } is the ground-truth label for that sample. The behaviour of the parameter <math> \alpha </math> over time, which is logarithmic in nature and has the same behavior exhibited by the norm of a learned classifier, is shown in<br />
[[Media: figure1_log_behave.png| Figure 1]].<br />
<br />
<center>[[File:figure1_log_behave.png]]</center><br />
<br />
When <math> -1 \le q_i · \hat{x} \le 1 </math>, a possible cosine angle loss is <br />
<br />
<center>[[File:caloss.png]]</center><br />
<br />
But its final validation accuracy has slight decrease, compared to original models.<br />
<br />
==Using a Hadmard matrix==<br />
<br />
To recall, Hadmard matrix (Hedayat et al., 1978) <math> H </math> is an <math> n × n </math> matrix, where all of its entries are either +1 or −1.<br />
Furthermore, <math> H </math> is orthogonal, such that <math> HH^{T} = nI_n </math> where <math>I_n</math> is the identity matrix. Instead of using the entire Hadmard matrix <math>H</math>, a truncated version, <math> \hat{H} &isin; </math> {<math> {-1, 1}</math>}<math>^{C \times N}</math> where all <math>C</math> rows are orthogonal as the final classification layer is such that:<br />
<br />
<center><math><br />
y = \hat{H} \hat{x} + b<br />
</math></center><br />
<br />
This usage allows two main benefits:<br />
* A deterministic, low-memory and easily generated matrix that can be used for classification.<br />
* Removal of the need to perform a full matrix-matrix multiplication - as multiplying by a Hadamard matrix can be done by simple sign manipulation and addition.<br />
<br />
Here, <math>n</math> must be a multiple of 4, but it can be easily truncated to fit normally defined networks. Also, as the classifier weights are fixed to need only 1-bit precision, it is now possible to focus our attention on the features preceding it.<br />
<br />
=Experimental Results=<br />
<br />
The authors have evaluated their proposed model on the following datasets:<br />
<br />
==CIFAR-10/100==<br />
<br />
===About the dataset===<br />
<br />
CIFAR-10 is an image classification benchmark dataset containing 50,000 training images and 10,000 test images. The images are in color and contain 32×32 pixels. There are 10 possible classes of various animals and vehicles. CIFAR-100 holds the same number of images of same size, but contains 100 different classes.<br />
<br />
===Training Details===<br />
<br />
The authors trained a residual network ( He et al., 2016) on the CIFAR-10 dataset. The network depth was 56 and the same hyper-parameters as in the original work were used. A comparison of the two variants, i.e., the learned classifier and the proposed classifier with a fixed transformation is shown in [[Media: figure1_resnet_cifar10.png | Figure 2]].<br />
<br />
<center>[[File: figure1_resnet_cifar10.png]]</center><br />
<br />
These results demonstrate that although the training error is considerably lower for the network with learned classifier, both models achieve the same classification accuracy on the validation set. The authors conjecture is that with the new fixed parameterization, the network can no longer increase the<br />
norm of a given sample’s representation - thus learning its label requires more effort. As this may happen for specific seen samples - it affects only training error.<br />
<br />
The authors also compared using a fixed scale variable <math>\alpha </math> at different values vs. the learned parameter. Results for <math> \alpha = </math> {0.1, 1, 10} are depicted in [[Media: figure3_alpha_resnet_cifar.png| Figure 3]] for both training and validation error and as can be seen, similar validation accuracy can be obtained using a fixed scale value (in this case <math>\alpha </math>= 1 or 10 will suffice) at the expense of another hyper-parameter to seek. In all the further experiments the scaling parameter <math> \alpha </math> was regularized with the same weight decay coefficient used on original classifier.<br />
<br />
<center>[[File: figure3_alpha_resnet_cifar.png]]</center><br />
<br />
The authors then train the model on CIFAR-100 dataset. They used the DenseNet-BC model from Huang et al. (2017) with depth of 100 layers and k = 12. The higher number of classes caused the number of parameters to grow and encompassed about 4% of the whole model. However, validation accuracy for the fixed-classifier model remained equally good as the original model, and the same training curve was observed as earlier.<br />
<br />
==IMAGENET==<br />
<br />
===About the dataset===<br />
<br />
The Imagenet dataset introduced by Deng et al. (2009) spans over 1000 visual classes, and over 1.2 million samples. This is supposedly a more challenging dataset to work on as compared to CIFAR-10/100.<br />
<br />
===Experiment Details===<br />
<br />
The authors evaluated their fixed classifier method on Imagenet using Resnet50 by He et al. (2016) and Densenet169 model (Huang et al., 2017) as described in the original work. Using a fixed classifier removed approximately 2-million parameters were from the model, accounting for about 8% and 12 % of the model parameters respectively. The experiments revealed similar trends as observed on CIFAR-10.<br />
<br />
For a more stricter evaluation, the authors also trained a Shufflenet architecture (Zhang et al., 2017b), which was designed to be used in low memory and limited computing platforms and has parameters making up the majority of the model. They were able to reduce the parameters to 0.86 million as compared to 0.96 million parameters in the final layer of the original model. Again, the proposed modification in the original model gave similar convergence results on validation accuracy.<br />
<br />
The overall results of the fixed-classifier are summarized in [[Media: table1_fixed_results.png | Table 1]].<br />
<br />
<center>[[File: table1_fixed_results.png]]</center><br />
<br />
==Language Modelling==<br />
<br />
Recent works have empirically found that using the same weights for both word embedding and classifier can yield equal or better results than using a separate pair of weights. So the authors experimented with fix-classifiers on language modelling as it also requires classification of all possible tokens available in the task vocabulary. They trained a recurrent model with 2-layers of LSTM (Hochreiter & Schmidhuber, 1997) and embedding + hidden size of 512 on the WikiText2 dataset (Merity et al., 2016), using same settings as in Merity et al. (2017). However, using a random orthogonal transform yielded poor results compared to learned embedding. This was suspected to be due to semantic relationships captured in the embedding layer of language models, which is not the case in image classification task. The intuition was further confirmed by the much better results when pre-trained embeddings using word2vec algorithm by Mikolov et al. (2013) or PMI factorization as suggested by Levy & Goldberg (2014), were used.<br />
<br />
=Discussion=<br />
<br />
==Implications and use cases==<br />
<br />
With the increasing number of classes in the benchmark datasets, computational demands for the final classifier will increase as well. In order to understand the problem better, the authors observe the work by Sun et al. (2017), which introduced JFT-300M - an internal Google dataset with over 18K different classes. Using a Resnet50 (He et al., 2016), with a 2048 sized representation led to a model with over 36M parameters meaning that over 60% of the model parameters resided in the final classification layer. Sun et al. (2017) also describe the difficulty in distributing so many parameters over the training servers involving a non-trivial overhead during synchronization of the model for update. The authors claim that the fixed-classifier would help considerably in this kind of scenario - where using a fixed classifier removes the need to do any gradient synchronization for the final layer. Furthermore, introduction of Hadamard matrix removes the need to save the transformation altogether, thereby, making it more efficient and allowing considerable memory and computational savings.<br />
<br />
==Possible Caveats==<br />
<br />
The good performance of fixed-classifiers relies on the ability of the preceding layers to learn separable representations. This could be affected when when the ratio between learned features and number of classes is small – that is, when <math> C > N</math>. However, they tested their method in such cases and their model performed well and provided good results.<br />
Another factor that can affect the performance of their model using a fixed classifier is when the classes are highly correlated. In that case, the fixed classifier actually cannot support correlated classes and thus, the network could have some difficulty to learn. For a language model, word classes tend to have highly correlated instances, which also lead to difficult learning process.<br />
<br />
==Future Work==<br />
<br />
<br />
The use of fixed classifiers might be further simplified in Binarized Neural Networks (Hubara et al., 2016a), where the activations and weights are restricted to ±1 during propagations. In that case the norm of the last hidden layer would be constant for all samples (equal to the square root of the hidden layer width). The constant could then be absorbed into the scale constant <math>\alpha</math>, and there is no need in a per-sample normalization.<br />
<br />
Additionally, more efficient ways to learn a word embedding should also be explored where similar redundancy in classifier weights may suggest simpler forms of token representations - such as low-rank or sparse versions.<br />
<br />
A related paper was published that claims that fixing most of the parameters of the neural network achieves comparable results with learning all of them [A. Rosenfeld and J. K. Tsotsos]<br />
<br />
=Conclusion=<br />
<br />
In this work, the authors argue that the final classification layer in deep neural networks is redundant and suggest removing the parameters from the classification layer. The empirical results from experiments on the CIFAR and IMAGENET datasets suggest that such a change lead to little or almost no decline in the performance of the architecture. Furthermore, using a Hadmard matrix as classifier might lead to some computational benefits when properly implemented, and save memory otherwise spent on large amount of transformation coefficients.<br />
<br />
Another possible scope of research that could be pointed out for future could be to find new efficient methods to create pre-defined word embeddings, which require huge amount of parameters that can possibly be avoided when learning a new task. Therefore, more emphasis should be given to the representations learned by the non-linear parts of the neural networks - upto the final classifier, as it seems highly redundant.<br />
<br />
=Critique=<br />
<br />
The paper proposes an interesting idea that has a potential use case when designing memory-efficient neural networks. The experiments shown in the paper are quite rigorous and provide support to the authors' claim. However, it would have been more helpful if the authors had described a bit more about efficient implementation of the Hadamard matrix and how to scale this method for larger datasets (cases with <math> C >N</math>).<br />
<br />
The paper presents a very interesting idea which can be applied for most of networks in deep learning area. However, technical proofs of the effect of the algorithm should be considered in order to generalize and be appreciated further.<br />
<br />
=References=<br />
<br />
The code for the proposed model is available at https://github.com/eladhoffer/fix_your_classifier.<br />
<br />
Madhu S Advani and Andrew M Saxe. High-dimensional dynamics of generalization error in neural networks. arXiv preprint arXiv:1710.03667, 2017.<br />
<br />
Peter Bartlett, Dylan J Foster, and Matus Telgarsky. Spectrally-normalized margin bounds for neural networks. arXiv preprint arXiv:1706.08498, 2017.<br />
<br />
Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Sackinger, and Roopak Shah. Signature verification using a” siamese” time delay neural network. In Advances in Neural Information Processing Systems, pp. 737–744, 1994.<br />
<br />
Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In Advances in Neural Information Processing Systems, pp. 3123–3131, 2015.<br />
<br />
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pp. 248–255. IEEE, 2009.<br />
<br />
Suriya Gunasekar, Blake Woodworth, Srinadh Bhojanapalli, Behnam Neyshabur, and Nathan Srebro. Implicit regularization in matrix factorization. arXiv preprint arXiv:1705.09280, 2017.<br />
<br />
Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015.<br />
<br />
Moritz Hardt and Tengyu Ma. Identity matters in deep learning. 2017.<br />
<br />
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016.<br />
<br />
A Hedayat, WD Wallis, et al. Hadamard matrices and their applications. The Annals of Statistics, 6<br />
(6):1184–1238, 1978.<br />
<br />
Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8): 1735–1780, 1997.<br />
<br />
Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In International Workshop on Similarity-Based Pattern Recognition, pp. 84–92. Springer, 2015.<br />
<br />
Elad Hoffer, Itay Hubara, and Daniel Soudry. Train longer, generalize better: closing the generalization gap in large batch training of neural networks. 2017.<br />
<br />
Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017.<br />
<br />
Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017.<br />
<br />
Guang-Bin Huang, Qin-Yu Zhu, and Chee-Kheong Siew. Extreme learning machine: theory and applications. Neurocomputing, 70(1):489–501, 2006.<br />
<br />
Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized neural networks. In Advances in Neural Information Processing Systems 29 (NIPS’16), 2016a.<br />
<br />
Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Quantized neural networks: Training neural networks with low precision weights and activations. arXiv preprint arXiv:1609.07061, 2016b.<br />
<br />
Hakan Inan, Khashayar Khosravi, and Richard Socher. Tying word vectors and word classifiers: A loss framework for language modeling. arXiv preprint arXiv:1611.01462, 2016.<br />
<br />
Max Jaderberg, Andrea Vedaldi, and Andrew Zisserman. Speeding up convolutional neural networks with low rank expansions. arXiv preprint arXiv:1405.3866, 2014.<br />
<br />
Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009.<br />
<br />
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012.<br />
<br />
Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´ document recognition. Proceedings of the IEEE, 86(11):2278 2324, 1998.<br />
<br />
Omer Levy and Yoav Goldberg. Neural word embedding as implicit matrix factorization. In Advances in neural information processing systems, pp. 2177–2185, 2014.<br />
<br />
Fengfu Li, Bo Zhang, and Bin Liu. Ternary weight networks. arXiv preprint arXiv:1605.04711, 2016.<br />
<br />
Min Lin, Qiang Chen, and Shuicheng Yan. Network in network. arXiv preprint arXiv:1312.4400, 2013.<br />
<br />
Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016.<br />
<br />
Stephen Merity, Nitish Shirish Keskar, and Richard Socher. Regularizing and Optimizing LSTM Language Models. arXiv preprint arXiv:1708.02182, 2017.<br />
<br />
Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaev, Ganesh Venkatesh, et al. Mixed precision training. arXiv preprint arXiv:1710.03740, 2017.<br />
<br />
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed tations of words and phrases and their compositionality. In Advances in neural information processing systems, pp. 3111–3119, 2013.<br />
<br />
Behnam Neyshabur, Srinadh Bhojanapalli, David McAllester, and Nathan Srebro. Exploring generalization in deep learning. arXiv preprint arXiv:1706.08947, 2017.<br />
Jooyoung Park and Irwin W Sandberg. Universal approximation using radial-basis-function networks. Neural computation, 3(2):246–257, 1991.<br />
<br />
Ofir Press and Lior Wolf. Using the output embedding to improve language models. EACL 2017,<br />
pp. 157, 2017.<br />
<br />
Itay Safran and Ohad Shamir. On the quality of the initial basin in overspecified neural networks. In International Conference on Machine Learning, pp. 774–782, 2016.<br />
<br />
Tim Salimans and Diederik P Kingma. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. In Advances in Neural Information Processing Systems, pp. 901–909, 2016.<br />
<br />
Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 815–823, 2015.<br />
<br />
Mahdi Soltanolkotabi, Adel Javanmard, and Jason D Lee. Theoretical insights into the optimization landscape of over-parameterized shallow neural networks. arXiv preprint arXiv:1707.04926, 2017.<br />
<br />
Daniel Soudry and Yair Carmon. No bad local minima: Data independent training error guarantees for multilayer neural networks. arXiv preprint arXiv:1605.08361, 2016.<br />
<br />
Daniel Soudry and Elad Hoffer. Exponentially vanishing sub-optimal local minima in multilayer neural networks. arXiv preprint arXiv:1702.05777, 2017.<br />
<br />
Daniel Soudry, Elad Hoffer, and Nathan Srebro. The implicit bias of gradient descent on separable data. 2018.<br />
<br />
Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806, 2014.<br />
<br />
Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. arXiv preprint arXiv:1707.02968, 2017.<br />
<br />
Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1–9, 2015.<br />
<br />
Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2818–2826, 2016.<br />
<br />
Cheng Tai, Tong Xiao, Yi Zhang, Xiaogang Wang, et al. Convolutional neural networks with lowrank regularization. arXiv preprint arXiv:1511.06067, 2015.<br />
<br />
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. 2017.<br />
Ashia C Wilson, Rebecca Roelofs, Mitchell Stern, Nathan Srebro, and Benjamin Recht. The marginal value of adaptive gradient methods in machine learning. arXiv preprint arXiv:1705.08292, 2017.<br />
<br />
Bo Xie, Yingyu Liang, and Le Song. Diversity leads to generalization in neural networks. arXiv preprint arXiv:1611.03131, 2016.<br />
<br />
Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In European conference on computer vision, pp. 818–833. Springer, 2014. Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. In ICLR, 2017a. URL https://arxiv.org/abs/1611.03530.<br />
<br />
Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. arXiv preprint arXiv:1707.01083, 2017b.<br />
<br />
Shuchang Zhou, Zekun Ni, Xinyu Zhou, He Wen, Yuxin Wu, and Yuheng Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160, 2016.<br />
<br />
A. Rosenfeld and J. K. Tsotsos, “Intriguing properties of randomly weighted networks: Generalizing while learning next to nothing,” arXiv preprint arXiv:1802.00844, 2018.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=learn_what_not_to_learn&diff=41912learn what not to learn2018-11-29T22:21:35Z<p>Ka2khan: /* Action elimination with contextual bandits */</p>
<hr />
<div>=Introduction=<br />
In reinforcement learning, it is often difficult for an agent to learn when the action space is large, especially the difficulties from function approximation and exploration. In some cases many actions are irrelevant and it is sometimes easier for the algorithm to learn which action not to take. The paper proposes a new reinforcement learning approach for dealing with large action spaces based on action elimination by restricting the available actions in each state to a subset of the most likely ones. There is a core assumption being made in the proposed method that it is easier to predict which actions in each state are invalid or inferior and use that information for control. More specifically, it proposes a system that learns the approximation of a Q-function and concurrently learns to eliminate actions. The method utilizes an external elimination signal which incorporates domain-specific prior knowledge. For example, in parser-based text games, the parser gives feedback regarding irrelevant actions after the action is played (e.g., Player: "Climb the tree." Parser: "There are no trees to climb"). Then a machine learning model can be trained to generalize to unseen states. <br />
<br />
The paper focuses on tasks where both states and the actions are natural language. It introduces a novel deep reinforcement learning approach which has a Deep Q-Network (DQN) and an Action Elimination Network (AEN), both using the Convolutional Neural Networks (CNN) for Natural Language Processing (NLP) tasks. The AEN is trained to predict invalid actions, supervised by the elimination signal from the environment. The proposed method uses the final layer activations of AEN to build a linear contextual bandit model which allows the elimination of sub-optimal actions with high probability. '''Note that the core assumption is that it is easy to predict which actions are invalid or inferior in each state and leverage that information for control.'''<br />
<br />
The text-based game called "Zork", which lets players to interact with a virtual world through a text based interface, is tested by using the elimination framework. The AEN algorithm has achieved faster learning rate than the baseline agents through eliminating irrelevant actions.<br />
<br />
Below shows an example for the Zork interface:<br />
<br />
[[File:lnottol_fig1.png|500px|center]]<br />
<br />
All states and actions are given in natural language. Input for the game contains more than a thousand possible actions in each state since player can type anything.<br />
<br />
=Related Work=<br />
Text-Based Games(TBG): The state of the environment in TBG is described by simple language. The player interacts with the environment with text command which respects a pre-defined grammar. A popular example is Zork which has been tested in the paper. TBG is a good research intersection of RL and NLP, it requires language understanding, long-term memory, planning, exploration, affordability extraction and common sense. It also often introduce stochastic dynamics to increase randomness.<br />
<br />
Representations for TBG: Good word representation is necessary in order to learn control policies from texts. Previous work on TBG used pre-trained embeddings directly for control. other works combined pre-trained embedding with neural networks.<br />
<br />
DRL with linear function approximation: DRL methods such as the DQN have achieved state-of-the-art results in a variety of challenging, high-dimensional domains. This is mainly because neural networks can learn rich domain representations for value function and policy. On the other hand, linear representation batch reinforcement learning methods are more stable and accurate, while feature engineering is necessary.<br />
<br />
RL in Large Action Spaces: Prior work concentrated on factorizing the action space into binary subspace(Pazis and Parr, 2011; Dulac-Arnold et al., 2012; Lagoudakis and Parr, 2003), other works proposed to embed the discrete actions into a continuous space, then choose the nearest discrete action according to the optimal actions in the continuous space(Dulac-Arnold et al., 2015; Van Hasselt and Wiering, 2009). He et. al. (2015)extended DQN to unbounded(natural language) action spaces.<br />
Learning to eliminate actions was first mentioned by (Even-Dar, Mannor, and Mansour, 2003). They proposed to learn confidence intervals around the value function in each state. Lipton et al.(2016a) proposed to learn a classifier that detects hazardous state and then use it to shape the reward. Fulda et al.(2017) presented a method for affordability extraction via inner products of pre-trained word embedding.<br />
<br />
=Action Elimination=<br />
<br />
The approach in the paper builds on the standard Reinforcement Learning formulation. At each time step <math>t</math>, the agent observes state <math display="inline">s_t </math> and chooses a discrete action <math display="inline">a_t\in\{1,...,|A|\} </math>. Then, after action execution, the agent obtains a reward <math display="inline">r_t(s_t,a_t) </math> and observes next state <math display="inline">s_{t+1} </math> according to a transition kernel <math>P(s_{t+1}|s_t,a_t)</math>. The goal of the algorithm is to learn a policy <math display="inline">\pi(a|s) </math> which maximizes the expected future discounted cumulative return <math display="inline">V^\pi(s)=E^\pi[\sum_{t=0}^{\infty}\gamma^tr(s_t,a_t)|s_0=s]</math>, where <math> 0< \gamma <1 </math>. The Q-function is <math display="inline">Q^\pi(s,a)=E^\pi[\sum_{t=0}^{\infty}\gamma^tr(s_t,a_t)|s_0=s,a_0=a]</math>, and it can be optimized by Q-learning algorithm.<br />
<br />
After executing an action, the agent observes a binary elimination signal <math>e(s, a)</math> to determine which actions not to take. It equals 1 if action <math>a</math> may be eliminated in state <math>s</math> (and 0 otherwise). The signal helps mitigating the problem of large discrete action spaces. We start with the following definitions:<br />
<br />
'''Definition 1:''' <br />
<br />
Valid state-action pairs with respect to an elimination signal are state action pairs which the elimination process should not eliminate. <br />
<br />
The set of valid state-action pairs contains all of the state-action pairs that are a part of some optimal policy, i.e., only strictly suboptimal state-actions can be invalid.<br />
<br />
'''Definition 2:'''<br />
<br />
Admissible state-action pairs with respect to an elimination algorithm are state action pairs which the elimination algorithm does not eliminate.<br />
<br />
'''Definition 3:'''<br />
<br />
Action Elimination Q-learning is a Q-learning algorithm which updates only admissible state-action pairs and chooses the best action in the next state from its admissible actions. We allow the base Q-learning algorithm to be any algorithm that converges to <math display="inline">Q^*</math> with probability 1 after observing each state-action infinitely often.<br />
<br />
==Advantages of Action Elimination==<br />
The main advantages of action elimination is that it allows the agent to overcome some of the main difficulties in large action spaces which are Function Approximation and Sample Complexity. <br />
<br />
Function approximation: Errors in the Q-function estimates may cause the learning algorithm to converge to a suboptimal policy, this phenomenon becomes more noticeable when the action space is large. Action elimination mitigates this effect by taking the max operator only on valid actions, thus, reducing potential overestimation errors. Besides, by ignoring the invalid actions, the function approximation can also learn a simpler mapping (i.e., only the Q-values of the valid state-action pairs) leading to faster convergence and better solution.<br />
<br />
Sample complexity: The sample complexity measures the number of steps during learning, in which the policy is not <math display="inline">\epsilon</math>-optimal. Assume that there are <math>A'</math> actions that should be eliminated and are <math>\epsilon</math>-optimal, i.e. their value is at least <math>V^*(s)-\epsilon</math>. The invalid action often returns no reward and doesn't change the state, (Lattimore and Hutter, 2012)resulting in an action gap of <math display="inline">\epsilon=(1-\gamma)V^*(s)</math>, and this translates to <math display="inline">V^*(s)^{-2}(1-\gamma)^{-5}log(1/\delta)</math> wasted samples for learning each invalid state-action pair. Practically, elimination algorithm can eliminate these invalid actions and therefore speed up the learning process approximately by <math display="inline">A/A'</math>.<br />
<br />
Because it is difficult to embed the elimination signal into the MDP, the authors use contextual multi-armed bandits to decouple the elimination signal from the MDP, which can correctly eliminate actions when applying standard Q learning into learning process.<br />
<br />
==Action elimination with contextual bandits==<br />
<br />
Let <math display="inline">x(s_t)\in R^d </math> be the feature representation of <math display="inline">s_t </math>. We assume that under this representation there exists a set of parameters <math display="inline">\theta_a^*\in \mathbb{R}^d </math> such that the elimination signal in state <math display="inline">s_t </math> is <math display="inline">e_t(s_t,a) = \theta_a^{*T}x(s_t)+\eta_t </math>, where <math display="inline"> \Vert\theta_a^*\Vert_2\leq S</math>. <math display="inline">\eta_t</math> is an R-subgaussian random variable with zero mean that models additive noise to the elimination signal. When there is no noise in the elimination signal, R=0. Otherwise, <math display="inline">R\leq 1</math> since the elimination signal is bounded in [0,1]. Assume the elimination signal satisfies: <math display="inline">0\leq E[e_t(s_t,a)]\leq l </math> for any valid action and <math display="inline"> u\leq E[e_t(s_t, a)]\leq 1</math> for any invalid action. And <math display="inline"> l\leq u</math>. Denote by <math display="inline">X_{t,a}</math> as the matrix whose rows are the observed state representation vectors in which action a was chosen, up to time t. <math display="inline">E_{t,a}</math> as the vector whose elements are the observed state representation elimination signals in which action a was chosen, up to time t. Denote the solution to the regularized linear regression <math display="inline">\Vert X_{t,a}\theta_{t,a}-E_{t,a}\Vert_2^2+\lambda\Vert \theta_{t,a}\Vert_2^2 </math> (for some <math display="inline">\lambda>0</math>) by <math display="inline">\hat{\theta}_{t,a}=\bar{V}_{t,a}^{-1}X_{t,a}^TE_{t,a} </math>, where <math display="inline">\bar{V}_{t,a}=\lambda I + X_{t,a}^TX_{t,a}</math>.<br />
<br />
<br />
According to Theorem 2 in (Abbasi-Yadkori, Pal, and Szepesvari, 2011), <math display="inline">|\hat{\theta}_{t,a}^{T}x(s_t)-\theta_a^{*T}x(s_t)|\leq\sqrt{\beta_t(\delta)x(s_t)^T\bar{V}_{t,a}^{-1}x(s_t)}\ \forall t>0</math>, where <math display="inline">\sqrt{\beta_t(\delta)}=R\sqrt{2\ \text{log}(\text{det}(\bar{V}_{t,a})^{1/2}\text{det}(\lambda I)^{-1/2}/\delta)}+\lambda^{1/2}S</math>, with probability of at least <math display="inline">1-\delta</math>. If <math display="inline">\forall s\ ,\Vert x(s)\Vert_2 \leq L</math>, then <math display="inline">\beta_t</math> can be bounded by <math display="inline">\sqrt{\beta_t(\delta)} \leq R \sqrt{d\ \text{log}(1+tL^2/\lambda/\delta)}+\lambda^{1/2}S</math>. Next, define <math display="inline">\tilde{\delta}=\delta/k</math> and bound this probability for all the actions. i.e., <math display="inline">\forall a,t>0</math><br />
<br />
<math display="inline">Pr(|\hat{\theta}_{t-1,a}^{T}x(s_t)-\theta_{t-1, a}^{*T}x(s_t)|\leq\sqrt{\beta_t(\tilde\delta)x(s_t)^T\bar{V}_{t - 1,a}^{-1}x(s_t)}) \leq 1-\delta</math><br />
<br />
Recall that <math display="inline">E[e_t(s,a)]=\theta_a^{*T}x(s_t)\leq l</math> if a is a valid action. Then we can eliminate action a at state <math display="inline">s_t</math> if it satisfies:<br />
<br />
<math display="inline">\hat{\theta}_{t-1,a}^{T}x(s_t)-\sqrt{\beta_{t-1}(\tilde\delta)x(s_t)^T\bar{V}_{t-1,a}^{-1}x(s_t)})>l</math><br />
<br />
with probability <math display="inline">1-\delta</math> that we never eliminate any valid action. Note that <math display="inline">l, u</math> are not known. In practice, choosing <math display="inline">l</math> to be 0.5 should suffice.<br />
<br />
==Concurrent Learning==<br />
In fact, Q-learning and contextual bandit algorithms can learn simultaneously, resulting in the convergence of both algorithms, i.e., finding an optimal policy and a minimal valid action space. <br />
<br />
If the elimination is done based on the concentration bounds of the linear contextual bandits, it can be ensured that Action Elimination Q-learning converges, as shown in Proposition 1.<br />
<br />
'''Proposition 1:'''<br />
<br />
Assume that all state action pairs (s,a) are visited infinitely often, unless eliminated according to <math display="inline">\hat{\theta}_{t-1,a}^Tx(s)-\sqrt{\beta_{t-1}(\tilde{\delta})x(s)^T\bar{V}_{t-1,a}^{-1}x(s))}>l</math>. Then, with a probability of at least <math display="inline">1-\delta</math>, action elimination Q-learning converges to the optimal Q-function for any valid state-action pairs. In addition, actions which should be eliminated are visited at most <math display="inline">T_{s,a}(t)\leq 4\beta_t/(u-l)^2<br />
+1</math> times.<br />
<br />
Notice that when there is no noise in the elimination signal(R=0), we correctly eliminate actions with probability 1. so invalid actions will be sampled a finite number of times.<br />
<br />
=Method=<br />
The assumption that <math display="inline">e_t(s_t,a)=\theta_a^{*T}x(s_t)+\eta_t </math> might not hold when using raw features like word2vec. So the paper proposes to use the neural network's last layer as features. A practical challenge here is that the features must be fixed over time when used by the contextual bandit. So batch-updates framework(Levine et al., 2017;Riquelme, Tucker, and Snoek, 2018) is used, where a new contextual bandit model is learned for every few steps that uses the last layer activation of the AEN as features.<br />
<br />
==Architecture of action elimination framework==<br />
<br />
[[File:lnottol_fig1b.png|300px|center]]<br />
<br />
After taking action <math display="inline">a_t</math>, the agent observes <math display="inline">(r_t,s_{t+1},e_t)</math>. The agent use it to learn two function approximation deep neural networks: A DQN and an AEN. AEN provides an admissible actions set <math display="inline">A'</math> to the DQN, which uses this set to decide how to act and learn. The architecture for both the AEN and DQN is an NLP CNN(100 convolutional filters for AEN and 500 for DQN, with three different 1D kernels of length (1,2,3)), based on(Kim, 2014). The state is represented as a sequence of words, composed of the game descriptor and the player's inventory. These are truncated or zero padded to a length of 50 descriptor + 15 inventory words and each word is embedded into continuous vectors using word2vec in <math display="inline">R^{300}</math>. The features of the last four states are then concatenated together such that the final state representations s are in <math display="inline">R^{78000}</math>. The AEN is trained to minimize the MSE loss, using the elimination signal as a label. The code, the Zork domain, and the implementation of the elimination signal can be found [https://github.com/TomZahavy/CB_AE_DQN here.]<br />
<br />
==Psuedocode of the Algorithm==<br />
<br />
[[File:lnottol_fig2.png|750px|center]]<br />
<br />
AE-DQN trains two networks: a DQN denoted by Q and an AEN denoted by E. The algorithm creates a linear contextual bandit model from it every L iterations with procedure AENUpdate(). This procedure uses the activations of the last hidden layer of E as features, which are then used to create a contextual linear bandit model.AENUpdate() then solved this model and plugin it into the target AEN. The contextual linear bandit model <math display="inline">(E^-,V)</math> is then used to eliminate actions via the ACT() and Target() functions. ACT() follows an <math display="inline">\epsilon</math>-greedy mechanism on the admissible actions set. For exploitation, it selects the action with highest Q-value by taking an argmax on Q-values among <math display="inline">A'</math>. For exploration, it selects an action uniformly from <math display="inline">A'</math>. The targets() procedure is estimating the value function by taking max over Q-values only among admissible actions, hence, reducing function approximation errors.<br />
<br />
=Experiment=<br />
==Zork domain==<br />
The world of Zork presents a rich environment with a large state and action space. <br />
Zork players describe their actions using natural language instructions. For example, "open the mailbox". Then their actions were processed by a sophisticated natural language parser. Based on the results, the game presents the outcome of the action. The goal of Zork is to collect the Twenty Treasures of Zork and install them in the trophy case. Points that are generated from the game's scoring system are given to the agent as the reward. For example, the player gets the points when solving the puzzles. Placing all treasures in the trophy will get 350 points. The elimination signal is given in two forms, "wrong parse" flag, and text feedback "you cannot take that". These two signals are grouped together into a single binary signal which then provided to the algorithm. <br />
<br />
Experiments begin with the two subdomains of Zork domains: Egg Quest and the Troll Quest. For these subdomains, an additional reward signal is provided to guide the agent towards solving specific tasks and make the results more visible. A reward of -1 is applied at every time step to encourage the agent to favor short paths. Each trajectory terminates is upon completing the quest or after T steps are taken. The discounted factor for training is <math display="inline">\gamma=0.8</math> and <math display="inline">\gamma=1</math> during evaluation. Also <math display="inline">\beta=0.5, l=0.6</math> in all experiments. <br />
<br />
===Egg Quest===<br />
The goal for this quest is to find and open the jewel-encrusted egg hidden on a tree in the forest. The agent will get 100 points upon completing this task. For action space, there are 9 fixed actions for navigation, and a second subset which consisting <math display="inline">N_{Take}</math> actions for taking possible objects in the game. <math display="inline">N_{Take}=200 (set A_1), N_{Take}=300 (set A_2)</math> has been tested separately.<br />
AE-DQN (blue) and a vanilla DQN agent (green) has been tested in this quest.<br />
<br />
[[File:AEF_zork_comparison.png|1200px|thumb|center|Performance of agents in the egg quest.]]<br />
<br />
Figure a) corresponds to the set <math display="inline">A_1</math>, with T=100, b) corresponds to the set <math display="inline">A_2</math>, with T=100, and c) corresponds to the set <math display="inline">A_2</math>, with T=200. Both agents has performed well on sets a and c. However the AE-DQN agent has learned much faster than the DQN on set b, which implies that action elimination is more robust to hyperparameter optimization when the action space is large. One important observation to note is that the three figures have different scales for the cumulative reward. While the AE-DQN outperformed the standard DQN in figure b, both models performed significantly better with the hyperparameter configuration in figure c.<br />
<br />
===Troll Quest===<br />
The goal of this quest is to find the troll. To do it the agent need to find the way to the house, use a lantern to expose the hidden entrance to the underworld. It will get 100 points upon achieving the goal. This quest is a larger problem than Egg Quest. The action set <math display="inline">A_1</math> is 200 take actions and 15 necessary actions, 215 in total.<br />
<br />
[[File:AEF_troll_comparison.png|400px|thumb|center|Results in the Troll Quest.]]<br />
<br />
The red line above is an "optimal elimination" baseline which consists of only 35 actions(15 essential, and 20 relevant take actions). We can see that AE-DQN still outperforms DQN and its improvement over DQN is more significant in the Troll Quest than the Egg quest. Also, it achieves compatible performance to the "optimal elimination" baseline.<br />
<br />
===Open Zork===<br />
Lastly, the "Open Zork" domain has been tested which only the environment reward has been used. 1M steps has been trained. Each trajectory terminates after T=200 steps. Two action sets have been used:<math display="inline">A_3</math>, the "Minimal Zork" action set, which is the minimal set of actions (131) that is required to solve the game. <math display="inline">A_4</math>, the "Open Zork" action set (1227) which composed of {Verb, Object} tuples for all the verbs and objects in the game.<br />
<br />
[[]]<br />
<br />
[[File:AEF_open_zork_comparison.png|600px|thumb|center|Results in "Open Zork".]]<br />
<br />
<br />
The above Figure shows the learning curve for both AE-DQN and DQN. We can see that AE-DQN (blue) still outperform the DQN (blue) in terms of speed and cumulative reward.<br />
<br />
=Conclusion=<br />
In this paper, the authors proposed a Deep Reinforcement Learning model for sub-optimal actions while performing Q-learning. Moreover, they showed that by eliminating actions, using linear contextual bandits with theoretical guarantees of convergence, the size of the action space is reduced, exploration is more effective, and learning is improved when tested on Zork, a text-based game.<br />
<br />
For future work the authors aim to investigate more sophisticated architectures and tackle learning shared representations for elimination and control which may boost performance on both tasks.<br />
<br />
They also hope to to investigate other mechanisms for action elimination, such as eliminating actions that result from low Q-values as in Even-Dar, Mannor, and Mansour, 2003.<br />
<br />
The authors also hope to generate elimination signals in real-world domains and achieve the purpose of eliminating the signal implicitly.<br />
<br />
=Critique=<br />
The paper is not a significant algorithmic contribution and it merely adds an extra layer of complexity to the most famous DQN algorithm. All the experimental domains considered in the paper are discrete action problems that have so many actions that it could have been easily extended to a continuous action problem. In continuous action space there are several policy gradient based RL algorithms that have provided stronger performances. The authors should have ideally compared their methods to such algorithms like PPO or DRPO.<br />
Even with the critique above, the paper presents mathematical/theoretical justifications of the methodology. Moreover, since the methodology is built on the standard RL framework, this means that other variant RL algorithms can apply the idea to decrease the complexity and increase the performance. Moreover, the there are some rooms for applying technical variantions for the algorithm.<br />
<br />
=Reference=<br />
1. Chu, W.; Li, L.; Reyzin, L.; and Schapire, R. 2011. Contextual bandits with linear payoff functions. In Proceedings of the Fourteenth International Conference on Artiﬁcial Intelligence and Statistics.<br />
<br />
2. Côté,M.-A.;Kádár,Á.;Yuan,X.;Kybartas,B.;Barnes,T.;Fine,E.;Moore,J.;Hausknecht,M.;Asri, L. E.; Adada, M.; et al. 2018. Textworld: A learning environment for text-based games. arXiv.<br />
<br />
3. Dulac-Arnold, G.; Evans, R.; van Hasselt, H.; Sunehag, P.; Lillicrap, T.; Hunt, J.; Mann, T.; Weber, T.; Degris, T.; and Coppin, B. 2015. Deep reinforcement learning in large discrete action spaces. arXiv.<br />
<br />
4. He, J.; Chen, J.; He, X.; Gao, J.; Li, L.; Deng, L.; and Ostendorf, M. 2015. Deep reinforcement learning with an unbounded action space. CoRR abs/1511.04636.<br />
<br />
5. Kim, Y. 2014. Convolutional neural networks for sentence classiﬁcation. arXiv preprint.<br />
<br />
6. VanHasselt,H.,andWiering,M.A. 2009. Usingcontinuousactionspacestosolvediscreteproblems. In Neural Networks, 2009. IJCNN 2009. International Joint Conference on, 1149–1156. IEEE.<br />
<br />
7. Watkins, C. J., and Dayan, P. 1992. Q-learning. Machine learning 8(3-4):279–292.<br />
<br />
8. Su, P.-H.; Gasic, M.; Mrksic, N.; Rojas-Barahona, L.; Ultes, S.; Vandyke, D.; Wen, T.-H.; and Young, S. 2016. Continuously learning neural dialogue management. arXiv preprint.<br />
<br />
9. Wu, Y.; Schuster, M.; Chen, Z.; Le, Q. V.; Norouzi, M.; Macherey, W.; Krikun, M.; Cao, Y.; Gao, Q.; Macherey, K.; et al. 2016. Google’s neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint.<br />
<br />
10. Yuan, X.; Côté, M.-A.; Sordoni, A.; Laroche, R.; Combes, R. T. d.; Hausknecht, M.; and Trischler, A. 2018. Counting to explore and generalize in text-based games. arXiv preprint arXiv:1806.1152</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=learn_what_not_to_learn&diff=41907learn what not to learn2018-11-29T22:09:18Z<p>Ka2khan: /* Action Elimination */</p>
<hr />
<div>=Introduction=<br />
In reinforcement learning, it is often difficult for an agent to learn when the action space is large, especially the difficulties from function approximation and exploration. In some cases many actions are irrelevant and it is sometimes easier for the algorithm to learn which action not to take. The paper proposes a new reinforcement learning approach for dealing with large action spaces based on action elimination by restricting the available actions in each state to a subset of the most likely ones. There is a core assumption being made in the proposed method that it is easier to predict which actions in each state are invalid or inferior and use that information for control. More specifically, it proposes a system that learns the approximation of a Q-function and concurrently learns to eliminate actions. The method utilizes an external elimination signal which incorporates domain-specific prior knowledge. For example, in parser-based text games, the parser gives feedback regarding irrelevant actions after the action is played (e.g., Player: "Climb the tree." Parser: "There are no trees to climb"). Then a machine learning model can be trained to generalize to unseen states. <br />
<br />
The paper focuses on tasks where both states and the actions are natural language. It introduces a novel deep reinforcement learning approach which has a Deep Q-Network (DQN) and an Action Elimination Network (AEN), both using the Convolutional Neural Networks (CNN) for Natural Language Processing (NLP) tasks. The AEN is trained to predict invalid actions, supervised by the elimination signal from the environment. The proposed method uses the final layer activations of AEN to build a linear contextual bandit model which allows the elimination of sub-optimal actions with high probability. '''Note that the core assumption is that it is easy to predict which actions are invalid or inferior in each state and leverage that information for control.'''<br />
<br />
The text-based game called "Zork", which lets players to interact with a virtual world through a text based interface, is tested by using the elimination framework. The AEN algorithm has achieved faster learning rate than the baseline agents through eliminating irrelevant actions.<br />
<br />
Below shows an example for the Zork interface:<br />
<br />
[[File:lnottol_fig1.png|500px|center]]<br />
<br />
All states and actions are given in natural language. Input for the game contains more than a thousand possible actions in each state since player can type anything.<br />
<br />
=Related Work=<br />
Text-Based Games(TBG): The state of the environment in TBG is described by simple language. The player interacts with the environment with text command which respects a pre-defined grammar. A popular example is Zork which has been tested in the paper. TBG is a good research intersection of RL and NLP, it requires language understanding, long-term memory, planning, exploration, affordability extraction and common sense. It also often introduce stochastic dynamics to increase randomness.<br />
<br />
Representations for TBG: Good word representation is necessary in order to learn control policies from texts. Previous work on TBG used pre-trained embeddings directly for control. other works combined pre-trained embedding with neural networks.<br />
<br />
DRL with linear function approximation: DRL methods such as the DQN have achieved state-of-the-art results in a variety of challenging, high-dimensional domains. This is mainly because neural networks can learn rich domain representations for value function and policy. On the other hand, linear representation batch reinforcement learning methods are more stable and accurate, while feature engineering is necessary.<br />
<br />
RL in Large Action Spaces: Prior work concentrated on factorizing the action space into binary subspace(Pazis and Parr, 2011; Dulac-Arnold et al., 2012; Lagoudakis and Parr, 2003), other works proposed to embed the discrete actions into a continuous space, then choose the nearest discrete action according to the optimal actions in the continuous space(Dulac-Arnold et al., 2015; Van Hasselt and Wiering, 2009). He et. al. (2015)extended DQN to unbounded(natural language) action spaces.<br />
Learning to eliminate actions was first mentioned by (Even-Dar, Mannor, and Mansour, 2003). They proposed to learn confidence intervals around the value function in each state. Lipton et al.(2016a) proposed to learn a classifier that detects hazardous state and then use it to shape the reward. Fulda et al.(2017) presented a method for affordability extraction via inner products of pre-trained word embedding.<br />
<br />
=Action Elimination=<br />
<br />
The approach in the paper builds on the standard Reinforcement Learning formulation. At each time step <math>t</math>, the agent observes state <math display="inline">s_t </math> and chooses a discrete action <math display="inline">a_t\in\{1,...,|A|\} </math>. Then, after action execution, the agent obtains a reward <math display="inline">r_t(s_t,a_t) </math> and observes next state <math display="inline">s_{t+1} </math> according to a transition kernel <math>P(s_{t+1}|s_t,a_t)</math>. The goal of the algorithm is to learn a policy <math display="inline">\pi(a|s) </math> which maximizes the expected future discounted cumulative return <math display="inline">V^\pi(s)=E^\pi[\sum_{t=0}^{\infty}\gamma^tr(s_t,a_t)|s_0=s]</math>, where <math> 0< \gamma <1 </math>. The Q-function is <math display="inline">Q^\pi(s,a)=E^\pi[\sum_{t=0}^{\infty}\gamma^tr(s_t,a_t)|s_0=s,a_0=a]</math>, and it can be optimized by Q-learning algorithm.<br />
<br />
After executing an action, the agent observes a binary elimination signal <math>e(s, a)</math> to determine which actions not to take. It equals 1 if action <math>a</math> may be eliminated in state <math>s</math> (and 0 otherwise). The signal helps mitigating the problem of large discrete action spaces. We start with the following definitions:<br />
<br />
'''Definition 1:''' <br />
<br />
Valid state-action pairs with respect to an elimination signal are state action pairs which the elimination process should not eliminate. <br />
<br />
The set of valid state-action pairs contains all of the state-action pairs that are a part of some optimal policy, i.e., only strictly suboptimal state-actions can be invalid.<br />
<br />
'''Definition 2:'''<br />
<br />
Admissible state-action pairs with respect to an elimination algorithm are state action pairs which the elimination algorithm does not eliminate.<br />
<br />
'''Definition 3:'''<br />
<br />
Action Elimination Q-learning is a Q-learning algorithm which updates only admissible state-action pairs and chooses the best action in the next state from its admissible actions. We allow the base Q-learning algorithm to be any algorithm that converges to <math display="inline">Q^*</math> with probability 1 after observing each state-action infinitely often.<br />
<br />
==Advantages of Action Elimination==<br />
The main advantages of action elimination is that it allows the agent to overcome some of the main difficulties in large action spaces which are Function Approximation and Sample Complexity. <br />
<br />
Function approximation: Errors in the Q-function estimates may cause the learning algorithm to converge to a suboptimal policy, this phenomenon becomes more noticeable when the action space is large. Action elimination mitigates this effect by taking the max operator only on valid actions, thus, reducing potential overestimation errors. Besides, by ignoring the invalid actions, the function approximation can also learn a simpler mapping (i.e., only the Q-values of the valid state-action pairs) leading to faster convergence and better solution.<br />
<br />
Sample complexity: The sample complexity measures the number of steps during learning, in which the policy is not <math display="inline">\epsilon</math>-optimal. Assume that there are <math>A'</math> actions that should be eliminated and are <math>\epsilon</math>-optimal, i.e. their value is at least <math>V^*(s)-\epsilon</math>. The invalid action often returns no reward and doesn't change the state, (Lattimore and Hutter, 2012)resulting in an action gap of <math display="inline">\epsilon=(1-\gamma)V^*(s)</math>, and this translates to <math display="inline">V^*(s)^{-2}(1-\gamma)^{-5}log(1/\delta)</math> wasted samples for learning each invalid state-action pair. Practically, elimination algorithm can eliminate these invalid actions and therefore speed up the learning process approximately by <math display="inline">A/A'</math>.<br />
<br />
Because it is difficult to embed the elimination signal into the MDP, the authors use contextual multi-armed bandits to decouple the elimination signal from the MDP, which can correctly eliminate actions when applying standard Q learning into learning process.<br />
<br />
==Action elimination with contextual bandits==<br />
<br />
Let <math display="inline">x(s_t)\in R^d </math> be the feature representation of <math display="inline">s_t </math>. We assume that under this representation there exists a set of parameters <math display="inline">\theta_a^*\in R_d </math> such that the elimination signal in state <math display="inline">s_t </math> is <math display="inline">e_t(s_t,a) = \theta_a^Tx(s_t)+\eta_t </math>, where <math display="inline"> \Vert\theta_a^*\Vert_2\leq S</math>. <math display="inline">\eta_t</math> is an R-subgaussian random variable with zero mean that models additive noise to the elimination signal. When there is no noise in the elimination signal, R=0. Otherwise, <math display="inline">R\leq 1</math> since the elimination signal is bounded in [0,1]. Assume the elimination signal satisfies: <math display="inline">0\leq E[e_t(s_t,a)]\leq l </math> for any valid action and <math display="inline"> u\leq E[e_t(s_t, a)]\leq 1</math> for any invalid action. And <math display="inline"> l\leq u</math>. Denote by <math display="inline">X_{t,a}</math> as the matrix whose rows are the observed state representation vectors in which action a was chosen, up to time t. <math display="inline">E_{t,a}</math> as the vector whose elements are the observed state representation elimination signals in which action a was chosen, up to time t. Denote the solution to the regularized linear regression <math display="inline">\Vert X_{t,a}\theta_{t,a}-E_{t,a}\Vert_2^2+\lambda\Vert \theta_{t,a}\Vert_2^2 </math> (for some <math display="inline">\lambda>0</math>) by <math display="inline">\hat{\theta}_{t,a}=\bar{V}_{t,a}^{-1}X_{t,a}^TE_{t,a} </math>, where <math display="inline">\bar{V}_{t,a}=\lambda I + X_{t,a}^TX_{t,a}</math>.<br />
<br />
<br />
According to Theorem 2 in (Abbasi-Yadkori, Pal, and Szepesvari, 2011), <math display="inline">|\hat{\theta}_{t,a}^{T}x(s_t)-\theta_a^{*T}x(s_t)|\leq\sqrt{\beta_t(\delta)x(s_t)^T\bar{V}_{t,a}^{-1}x(s_t)} \forall t>0</math>, where <math display="inline">\sqrt{\beta_t(\delta)}=R\sqrt{2log(det(\bar{V}_{t,a}^{1/2})det(\lambda I)^{-1/2}/\delta)}+\lambda^{1/2}S</math>, with probability of at least <math display="inline">1-\delta</math>. If <math display="inline">\forall s \Vert x(s)\Vert_2 \leq L</math>, then <math display="inline">\beta_t</math> can be bounded by <math display="inline">\sqrt{\beta_t(\delta)} \leq R \sqrt{dlog(1+tL^2/\lambda/\delta)}+\lambda^{1/2}S</math>. Next, define <math display="inline">\tilde{\delta}=\delta/k</math> and bound this probability for all the actions. i.e., <math display="inline">\forall a,t>0</math><br />
<br />
<math display="inline">Pr(|\hat{\theta}_{t,a}^{T}x(s_t)-\theta_a^{*T}x(s_t)|\leq\sqrt{\beta_t(\delta)x(s_t)^T\bar{V}_{t,a}^{-1}x(s_t)}) \leq 1-\delta</math><br />
<br />
Recall that <math display="inline">E[e_t(s,a)]=\theta_a^{*T}x(s_t)\leq l</math> if a is a valid action. Then we can eliminate action a at state <math display="inline">s_t</math> if it satisfies:<br />
<br />
<math display="inline">\hat{\theta}_{t,a}^{T}x(s_t)-\sqrt{\beta_t(\delta)x(s_t)^T\bar{V}_{t,a}^{-1}x(s_t)})>l</math><br />
<br />
with probability <math display="inline">1-\delta</math> that we never eliminate any valid action. Note that <math display="inline">l, u</math> are not known. In practice, choosing <math display="inline">l</math> to be 0.5 should suffice.<br />
<br />
==Concurrent Learning==<br />
In fact, Q-learning and contextual bandit algorithms can learn simultaneously, resulting in the convergence of both algorithms, i.e., finding an optimal policy and a minimal valid action space. <br />
<br />
If the elimination is done based on the concentration bounds of the linear contextual bandits, it can be ensured that Action Elimination Q-learning converges, as shown in Proposition 1.<br />
<br />
'''Proposition 1:'''<br />
<br />
Assume that all state action pairs (s,a) are visited infinitely often, unless eliminated according to <math display="inline">\hat{\theta}_{t-1,a}^Tx(s)-\sqrt{\beta_{t-1}(\tilde{\delta})x(s)^T\bar{V}_{t-1,a}^{-1}x(s))}>l</math>. Then, with a probability of at least <math display="inline">1-\delta</math>, action elimination Q-learning converges to the optimal Q-function for any valid state-action pairs. In addition, actions which should be eliminated are visited at most <math display="inline">T_{s,a}(t)\leq 4\beta_t/(u-l)^2<br />
+1</math> times.<br />
<br />
Notice that when there is no noise in the elimination signal(R=0), we correctly eliminate actions with probability 1. so invalid actions will be sampled a finite number of times.<br />
<br />
=Method=<br />
The assumption that <math display="inline">e_t(s_t,a)=\theta_a^{*T}x(s_t)+\eta_t </math> might not hold when using raw features like word2vec. So the paper proposes to use the neural network's last layer as features. A practical challenge here is that the features must be fixed over time when used by the contextual bandit. So batch-updates framework(Levine et al., 2017;Riquelme, Tucker, and Snoek, 2018) is used, where a new contextual bandit model is learned for every few steps that uses the last layer activation of the AEN as features.<br />
<br />
==Architecture of action elimination framework==<br />
<br />
[[File:lnottol_fig1b.png|300px|center]]<br />
<br />
After taking action <math display="inline">a_t</math>, the agent observes <math display="inline">(r_t,s_{t+1},e_t)</math>. The agent use it to learn two function approximation deep neural networks: A DQN and an AEN. AEN provides an admissible actions set <math display="inline">A'</math> to the DQN, which uses this set to decide how to act and learn. The architecture for both the AEN and DQN is an NLP CNN(100 convolutional filters for AEN and 500 for DQN, with three different 1D kernels of length (1,2,3)), based on(Kim, 2014). The state is represented as a sequence of words, composed of the game descriptor and the player's inventory. These are truncated or zero padded to a length of 50 descriptor + 15 inventory words and each word is embedded into continuous vectors using word2vec in <math display="inline">R^{300}</math>. The features of the last four states are then concatenated together such that the final state representations s are in <math display="inline">R^{78000}</math>. The AEN is trained to minimize the MSE loss, using the elimination signal as a label. The code, the Zork domain, and the implementation of the elimination signal can be found [https://github.com/TomZahavy/CB_AE_DQN here.]<br />
<br />
==Psuedocode of the Algorithm==<br />
<br />
[[File:lnottol_fig2.png|750px|center]]<br />
<br />
AE-DQN trains two networks: a DQN denoted by Q and an AEN denoted by E. The algorithm creates a linear contextual bandit model from it every L iterations with procedure AENUpdate(). This procedure uses the activations of the last hidden layer of E as features, which are then used to create a contextual linear bandit model.AENUpdate() then solved this model and plugin it into the target AEN. The contextual linear bandit model <math display="inline">(E^-,V)</math> is then used to eliminate actions via the ACT() and Target() functions. ACT() follows an <math display="inline">\epsilon</math>-greedy mechanism on the admissible actions set. For exploitation, it selects the action with highest Q-value by taking an argmax on Q-values among <math display="inline">A'</math>. For exploration, it selects an action uniformly from <math display="inline">A'</math>. The targets() procedure is estimating the value function by taking max over Q-values only among admissible actions, hence, reducing function approximation errors.<br />
<br />
=Experiment=<br />
==Zork domain==<br />
The world of Zork presents a rich environment with a large state and action space. <br />
Zork players describe their actions using natural language instructions. For example, "open the mailbox". Then their actions were processed by a sophisticated natural language parser. Based on the results, the game presents the outcome of the action. The goal of Zork is to collect the Twenty Treasures of Zork and install them in the trophy case. Points that are generated from the game's scoring system are given to the agent as the reward. For example, the player gets the points when solving the puzzles. Placing all treasures in the trophy will get 350 points. The elimination signal is given in two forms, "wrong parse" flag, and text feedback "you cannot take that". These two signals are grouped together into a single binary signal which then provided to the algorithm. <br />
<br />
Experiments begin with the two subdomains of Zork domains: Egg Quest and the Troll Quest. For these subdomains, an additional reward signal is provided to guide the agent towards solving specific tasks and make the results more visible. A reward of -1 is applied at every time step to encourage the agent to favor short paths. Each trajectory terminates is upon completing the quest or after T steps are taken. The discounted factor for training is <math display="inline">\gamma=0.8</math> and <math display="inline">\gamma=1</math> during evaluation. Also <math display="inline">\beta=0.5, l=0.6</math> in all experiments. <br />
<br />
===Egg Quest===<br />
The goal for this quest is to find and open the jewel-encrusted egg hidden on a tree in the forest. The agent will get 100 points upon completing this task. For action space, there are 9 fixed actions for navigation, and a second subset which consisting <math display="inline">N_{Take}</math> actions for taking possible objects in the game. <math display="inline">N_{Take}=200 (set A_1), N_{Take}=300 (set A_2)</math> has been tested separately.<br />
AE-DQN (blue) and a vanilla DQN agent (green) has been tested in this quest.<br />
<br />
[[File:AEF_zork_comparison.png|1200px|thumb|center|Performance of agents in the egg quest.]]<br />
<br />
Figure a) corresponds to the set <math display="inline">A_1</math>, with T=100, b) corresponds to the set <math display="inline">A_2</math>, with T=100, and c) corresponds to the set <math display="inline">A_2</math>, with T=200. Both agents has performed well on sets a and c. However the AE-DQN agent has learned much faster than the DQN on set b, which implies that action elimination is more robust to hyperparameter optimization when the action space is large. One important observation to note is that the three figures have different scales for the cumulative reward. While the AE-DQN outperformed the standard DQN in figure b, both models performed significantly better with the hyperparameter configuration in figure c.<br />
<br />
===Troll Quest===<br />
The goal of this quest is to find the troll. To do it the agent need to find the way to the house, use a lantern to expose the hidden entrance to the underworld. It will get 100 points upon achieving the goal. This quest is a larger problem than Egg Quest. The action set <math display="inline">A_1</math> is 200 take actions and 15 necessary actions, 215 in total.<br />
<br />
[[File:AEF_troll_comparison.png|400px|thumb|center|Results in the Troll Quest.]]<br />
<br />
The red line above is an "optimal elimination" baseline which consists of only 35 actions(15 essential, and 20 relevant take actions). We can see that AE-DQN still outperforms DQN and its improvement over DQN is more significant in the Troll Quest than the Egg quest. Also, it achieves compatible performance to the "optimal elimination" baseline.<br />
<br />
===Open Zork===<br />
Lastly, the "Open Zork" domain has been tested which only the environment reward has been used. 1M steps has been trained. Each trajectory terminates after T=200 steps. Two action sets have been used:<math display="inline">A_3</math>, the "Minimal Zork" action set, which is the minimal set of actions (131) that is required to solve the game. <math display="inline">A_4</math>, the "Open Zork" action set (1227) which composed of {Verb, Object} tuples for all the verbs and objects in the game.<br />
<br />
[[]]<br />
<br />
[[File:AEF_open_zork_comparison.png|600px|thumb|center|Results in "Open Zork".]]<br />
<br />
<br />
The above Figure shows the learning curve for both AE-DQN and DQN. We can see that AE-DQN (blue) still outperform the DQN (blue) in terms of speed and cumulative reward.<br />
<br />
=Conclusion=<br />
In this paper, the authors proposed a Deep Reinforcement Learning model for sub-optimal actions while performing Q-learning. Moreover, they showed that by eliminating actions, using linear contextual bandits with theoretical guarantees of convergence, the size of the action space is reduced, exploration is more effective, and learning is improved when tested on Zork, a text-based game.<br />
<br />
For future work the authors aim to investigate more sophisticated architectures and tackle learning shared representations for elimination and control which may boost performance on both tasks.<br />
<br />
They also hope to to investigate other mechanisms for action elimination, such as eliminating actions that result from low Q-values as in Even-Dar, Mannor, and Mansour, 2003.<br />
<br />
The authors also hope to generate elimination signals in real-world domains and achieve the purpose of eliminating the signal implicitly.<br />
<br />
=Critique=<br />
The paper is not a significant algorithmic contribution and it merely adds an extra layer of complexity to the most famous DQN algorithm. All the experimental domains considered in the paper are discrete action problems that have so many actions that it could have been easily extended to a continuous action problem. In continuous action space there are several policy gradient based RL algorithms that have provided stronger performances. The authors should have ideally compared their methods to such algorithms like PPO or DRPO.<br />
Even with the critique above, the paper presents mathematical/theoretical justifications of the methodology. Moreover, since the methodology is built on the standard RL framework, this means that other variant RL algorithms can apply the idea to decrease the complexity and increase the performance. Moreover, the there are some rooms for applying technical variantions for the algorithm.<br />
<br />
=Reference=</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=learn_what_not_to_learn&diff=41905learn what not to learn2018-11-29T22:06:12Z<p>Ka2khan: /* Action Elimination */</p>
<hr />
<div>=Introduction=<br />
In reinforcement learning, it is often difficult for an agent to learn when the action space is large, especially the difficulties from function approximation and exploration. In some cases many actions are irrelevant and it is sometimes easier for the algorithm to learn which action not to take. The paper proposes a new reinforcement learning approach for dealing with large action spaces based on action elimination by restricting the available actions in each state to a subset of the most likely ones. There is a core assumption being made in the proposed method that it is easier to predict which actions in each state are invalid or inferior and use that information for control. More specifically, it proposes a system that learns the approximation of a Q-function and concurrently learns to eliminate actions. The method utilizes an external elimination signal which incorporates domain-specific prior knowledge. For example, in parser-based text games, the parser gives feedback regarding irrelevant actions after the action is played (e.g., Player: "Climb the tree." Parser: "There are no trees to climb"). Then a machine learning model can be trained to generalize to unseen states. <br />
<br />
The paper focuses on tasks where both states and the actions are natural language. It introduces a novel deep reinforcement learning approach which has a Deep Q-Network (DQN) and an Action Elimination Network (AEN), both using the Convolutional Neural Networks (CNN) for Natural Language Processing (NLP) tasks. The AEN is trained to predict invalid actions, supervised by the elimination signal from the environment. The proposed method uses the final layer activations of AEN to build a linear contextual bandit model which allows the elimination of sub-optimal actions with high probability. '''Note that the core assumption is that it is easy to predict which actions are invalid or inferior in each state and leverage that information for control.'''<br />
<br />
The text-based game called "Zork", which lets players to interact with a virtual world through a text based interface, is tested by using the elimination framework. The AEN algorithm has achieved faster learning rate than the baseline agents through eliminating irrelevant actions.<br />
<br />
Below shows an example for the Zork interface:<br />
<br />
[[File:lnottol_fig1.png|500px|center]]<br />
<br />
All states and actions are given in natural language. Input for the game contains more than a thousand possible actions in each state since player can type anything.<br />
<br />
=Related Work=<br />
Text-Based Games(TBG): The state of the environment in TBG is described by simple language. The player interacts with the environment with text command which respects a pre-defined grammar. A popular example is Zork which has been tested in the paper. TBG is a good research intersection of RL and NLP, it requires language understanding, long-term memory, planning, exploration, affordability extraction and common sense. It also often introduce stochastic dynamics to increase randomness.<br />
<br />
Representations for TBG: Good word representation is necessary in order to learn control policies from texts. Previous work on TBG used pre-trained embeddings directly for control. other works combined pre-trained embedding with neural networks.<br />
<br />
DRL with linear function approximation: DRL methods such as the DQN have achieved state-of-the-art results in a variety of challenging, high-dimensional domains. This is mainly because neural networks can learn rich domain representations for value function and policy. On the other hand, linear representation batch reinforcement learning methods are more stable and accurate, while feature engineering is necessary.<br />
<br />
RL in Large Action Spaces: Prior work concentrated on factorizing the action space into binary subspace(Pazis and Parr, 2011; Dulac-Arnold et al., 2012; Lagoudakis and Parr, 2003), other works proposed to embed the discrete actions into a continuous space, then choose the nearest discrete action according to the optimal actions in the continuous space(Dulac-Arnold et al., 2015; Van Hasselt and Wiering, 2009). He et. al. (2015)extended DQN to unbounded(natural language) action spaces.<br />
Learning to eliminate actions was first mentioned by (Even-Dar, Mannor, and Mansour, 2003). They proposed to learn confidence intervals around the value function in each state. Lipton et al.(2016a) proposed to learn a classifier that detects hazardous state and then use it to shape the reward. Fulda et al.(2017) presented a method for affordability extraction via inner products of pre-trained word embedding.<br />
<br />
=Action Elimination=<br />
<br />
The approach in the paper builds on the standard Reinforcement Learning formulation. At each time step <math>t</math>, the agent observes state <math display="inline">s_t </math> and chooses a discrete action <math display="inline">a_t\in\{1,...,|A|\} </math>. Then, after action execution, the agent obtains a reward <math display="inline">r_t(s_t,a_t) </math> and observes next state <math display="inline">s_{t+1} </math> according to a transition kernel <math>P(s_{t+1}|s_t,a_t)</math>. The goal of the algorithm is to learn a policy <math display="inline">\pi(a|s) </math> which maximizes the expected future discounted cumulative return <math display="inline">V^\pi(s)=E^\pi[\sum_{t=0}^{\infty}\gamma^tr(s_t,a_t)|s_0=s]</math>, where <math> 0< \gamma <1 </math>. The Q-function is <math display="inline">Q^\pi(s,a)=E^\pi[\sum_{t=0}^{\infty}\gamma^tr(s_t,a_t)|s_0=s,a_0=a]</math>, and it can be optimized by Q-learning algorithm.<br />
<br />
After executing an action, the agent observes a binary elimination signal <math>e(s, a)</math> to determine which actions not to take. It equals 1 if action <math>a</math> may be eliminated in state <math>s</math> (and 0 otherwise). The signal helps mitigating the problem of large discrete action spaces. We start with the following definitions:<br />
<br />
'''Definition 1:''' <br />
<br />
Valid state-action pairs with respect to an elimination signal are state action pairs which the elimination process should not eliminate. <br />
<br />
The set of valid state-action pairs contains all of the state-action pairs that are a part of some optimal policy, i.e., only strictly suboptimal state-actions can be invalid.<br />
<br />
'''Definition 2:'''<br />
<br />
Admissible state-action pairs with respect to an elimination algorithm are state action pairs which the elimination algorithm does not eliminate.<br />
<br />
'''Definition 3:'''<br />
<br />
Action Elimination Q-learning is a Q-learning algorithm which updates only admissible state-action pairs and chooses the best action in the next state from its admissible actions. We allow the base Q-learning algorithm to be any algorithm that converges to <math display="inline">Q^*</math> with probability 1 after observing each state-action infinitely often.<br />
<br />
==Advantages of Action Elimination==<br />
The main advantages of action elimination is that it allows the agent to overcome some of the main difficulties in large action spaces which are Function Approximation and Sample Complexity. <br />
<br />
Function approximation: Errors in the Q-function estimates may cause the learning algorithm to converge to a suboptimal policy, this phenomenon becomes more noticeable when the action space is large. Action elimination mitigate this effect by taking the max operator only on valid actions, thus, reducing potential overestimation errors. Besides, by ignoring the invalid actions, the function approximation can also learn a simpler mapping (i.e., only the Q-values of the valid state-action pairs) leading to faster convergence and better solution.<br />
<br />
Sample complexity: The sample complexity measures the number of steps during learning, in which the policy is not <math display="inline">\epsilon</math>-optimal. Assume that there are <math>A'</math> actions that should be eliminated and are <math>\epsilon</math>-optimal, i.e. their value is at least <math>V^*(s)-\epsilon</math>. The invalid action often returns no reward and doesn't change the state, (Lattimore and Hutter, 2012)resulting in an action gap of <math display="inline">\epsilon=(1-\gamma)V^*(s)</math>, and this translates to <math display="inline">V^*(s)^{-2}(1-\gamma)^{-5}log(1/\delta)</math> wasted samples for learning each invalid state-action pair. Practically, elimination algorithm can eliminate these invalid actions and therefore speed up the learning process approximately by <math display="inline">A/A'</math>.<br />
<br />
Because it is difficult to embedde the elimination signal into the MDP, the authors use contextual multi-armed bandits to decouple the elimination signal from the MDP, which can correctly eliminate actions when applying standard Q learning into learning process.<br />
<br />
==Action elimination with contextual bandits==<br />
<br />
Let <math display="inline">x(s_t)\in R^d </math> be the feature representation of <math display="inline">s_t </math>. We assume that under this representation there exists a set of parameters <math display="inline">\theta_a^*\in R_d </math> such that the elimination signal in state <math display="inline">s_t </math> is <math display="inline">e_t(s_t,a) = \theta_a^Tx(s_t)+\eta_t </math>, where <math display="inline"> \Vert\theta_a^*\Vert_2\leq S</math>. <math display="inline">\eta_t</math> is an R-subgaussian random variable with zero mean that models additive noise to the elimination signal. When there is no noise in the elimination signal, R=0. Otherwise, <math display="inline">R\leq 1</math> since the elimination signal is bounded in [0,1]. Assume the elimination signal satisfies: <math display="inline">0\leq E[e_t(s_t,a)]\leq l </math> for any valid action and <math display="inline"> u\leq E[e_t(s_t, a)]\leq 1</math> for any invalid action. And <math display="inline"> l\leq u</math>. Denote by <math display="inline">X_{t,a}</math> as the matrix whose rows are the observed state representation vectors in which action a was chosen, up to time t. <math display="inline">E_{t,a}</math> as the vector whose elements are the observed state representation elimination signals in which action a was chosen, up to time t. Denote the solution to the regularized linear regression <math display="inline">\Vert X_{t,a}\theta_{t,a}-E_{t,a}\Vert_2^2+\lambda\Vert \theta_{t,a}\Vert_2^2 </math> (for some <math display="inline">\lambda>0</math>) by <math display="inline">\hat{\theta}_{t,a}=\bar{V}_{t,a}^{-1}X_{t,a}^TE_{t,a} </math>, where <math display="inline">\bar{V}_{t,a}=\lambda I + X_{t,a}^TX_{t,a}</math>.<br />
<br />
<br />
According to Theorem 2 in (Abbasi-Yadkori, Pal, and Szepesvari, 2011), <math display="inline">|\hat{\theta}_{t,a}^{T}x(s_t)-\theta_a^{*T}x(s_t)|\leq\sqrt{\beta_t(\delta)x(s_t)^T\bar{V}_{t,a}^{-1}x(s_t)} \forall t>0</math>, where <math display="inline">\sqrt{\beta_t(\delta)}=R\sqrt{2log(det(\bar{V}_{t,a}^{1/2})det(\lambda I)^{-1/2}/\delta)}+\lambda^{1/2}S</math>, with probability of at least <math display="inline">1-\delta</math>. If <math display="inline">\forall s \Vert x(s)\Vert_2 \leq L</math>, then <math display="inline">\beta_t</math> can be bounded by <math display="inline">\sqrt{\beta_t(\delta)} \leq R \sqrt{dlog(1+tL^2/\lambda/\delta)}+\lambda^{1/2}S</math>. Next, define <math display="inline">\tilde{\delta}=\delta/k</math> and bound this probability for all the actions. i.e., <math display="inline">\forall a,t>0</math><br />
<br />
<math display="inline">Pr(|\hat{\theta}_{t,a}^{T}x(s_t)-\theta_a^{*T}x(s_t)|\leq\sqrt{\beta_t(\delta)x(s_t)^T\bar{V}_{t,a}^{-1}x(s_t)}) \leq 1-\delta</math><br />
<br />
Recall that <math display="inline">E[e_t(s,a)]=\theta_a^{*T}x(s_t)\leq l</math> if a is a valid action. Then we can eliminate action a at state <math display="inline">s_t</math> if it satisfies:<br />
<br />
<math display="inline">\hat{\theta}_{t,a}^{T}x(s_t)-\sqrt{\beta_t(\delta)x(s_t)^T\bar{V}_{t,a}^{-1}x(s_t)})>l</math><br />
<br />
with probability <math display="inline">1-\delta</math> that we never eliminate any valid action. Note that <math display="inline">l, u</math> are not known. In practice, choosing <math display="inline">l</math> to be 0.5 should suffice.<br />
<br />
==Concurrent Learning==<br />
In fact, Q-learning and contextual bandit algorithms can learn simultaneously, resulting in the convergence of both algorithms, i.e., finding an optimal policy and a minimal valid action space. <br />
<br />
If the elimination is done based on the concentration bounds of the linear contextual bandits, it can be ensured that Action Elimination Q-learning converges, as shown in Proposition 1.<br />
<br />
'''Proposition 1:'''<br />
<br />
Assume that all state action pairs (s,a) are visited infinitely often, unless eliminated according to <math display="inline">\hat{\theta}_{t-1,a}^Tx(s)-\sqrt{\beta_{t-1}(\tilde{\delta})x(s)^T\bar{V}_{t-1,a}^{-1}x(s))}>l</math>. Then, with a probability of at least <math display="inline">1-\delta</math>, action elimination Q-learning converges to the optimal Q-function for any valid state-action pairs. In addition, actions which should be eliminated are visited at most <math display="inline">T_{s,a}(t)\leq 4\beta_t/(u-l)^2<br />
+1</math> times.<br />
<br />
Notice that when there is no noise in the elimination signal(R=0), we correctly eliminate actions with probability 1. so invalid actions will be sampled a finite number of times.<br />
<br />
=Method=<br />
The assumption that <math display="inline">e_t(s_t,a)=\theta_a^{*T}x(s_t)+\eta_t </math> might not hold when using raw features like word2vec. So the paper proposes to use the neural network's last layer as features. A practical challenge here is that the features must be fixed over time when used by the contextual bandit. So batch-updates framework(Levine et al., 2017;Riquelme, Tucker, and Snoek, 2018) is used, where a new contextual bandit model is learned for every few steps that uses the last layer activation of the AEN as features.<br />
<br />
==Architecture of action elimination framework==<br />
<br />
[[File:lnottol_fig1b.png|300px|center]]<br />
<br />
After taking action <math display="inline">a_t</math>, the agent observes <math display="inline">(r_t,s_{t+1},e_t)</math>. The agent use it to learn two function approximation deep neural networks: A DQN and an AEN. AEN provides an admissible actions set <math display="inline">A'</math> to the DQN, which uses this set to decide how to act and learn. The architecture for both the AEN and DQN is an NLP CNN(100 convolutional filters for AEN and 500 for DQN, with three different 1D kernels of length (1,2,3)), based on(Kim, 2014). The state is represented as a sequence of words, composed of the game descriptor and the player's inventory. These are truncated or zero padded to a length of 50 descriptor + 15 inventory words and each word is embedded into continuous vectors using word2vec in <math display="inline">R^{300}</math>. The features of the last four states are then concatenated together such that the final state representations s are in <math display="inline">R^{78000}</math>. The AEN is trained to minimize the MSE loss, using the elimination signal as a label. The code, the Zork domain, and the implementation of the elimination signal can be found [https://github.com/TomZahavy/CB_AE_DQN here.]<br />
<br />
==Psuedocode of the Algorithm==<br />
<br />
[[File:lnottol_fig2.png|750px|center]]<br />
<br />
AE-DQN trains two networks: a DQN denoted by Q and an AEN denoted by E. The algorithm creates a linear contextual bandit model from it every L iterations with procedure AENUpdate(). This procedure uses the activations of the last hidden layer of E as features, which are then used to create a contextual linear bandit model.AENUpdate() then solved this model and plugin it into the target AEN. The contextual linear bandit model <math display="inline">(E^-,V)</math> is then used to eliminate actions via the ACT() and Target() functions. ACT() follows an <math display="inline">\epsilon</math>-greedy mechanism on the admissible actions set. For exploitation, it selects the action with highest Q-value by taking an argmax on Q-values among <math display="inline">A'</math>. For exploration, it selects an action uniformly from <math display="inline">A'</math>. The targets() procedure is estimating the value function by taking max over Q-values only among admissible actions, hence, reducing function approximation errors.<br />
<br />
=Experiment=<br />
==Zork domain==<br />
The world of Zork presents a rich environment with a large state and action space. <br />
Zork players describe their actions using natural language instructions. For example, "open the mailbox". Then their actions were processed by a sophisticated natural language parser. Based on the results, the game presents the outcome of the action. The goal of Zork is to collect the Twenty Treasures of Zork and install them in the trophy case. Points that are generated from the game's scoring system are given to the agent as the reward. For example, the player gets the points when solving the puzzles. Placing all treasures in the trophy will get 350 points. The elimination signal is given in two forms, "wrong parse" flag, and text feedback "you cannot take that". These two signals are grouped together into a single binary signal which then provided to the algorithm. <br />
<br />
Experiments begin with the two subdomains of Zork domains: Egg Quest and the Troll Quest. For these subdomains, an additional reward signal is provided to guide the agent towards solving specific tasks and make the results more visible. A reward of -1 is applied at every time step to encourage the agent to favor short paths. Each trajectory terminates is upon completing the quest or after T steps are taken. The discounted factor for training is <math display="inline">\gamma=0.8</math> and <math display="inline">\gamma=1</math> during evaluation. Also <math display="inline">\beta=0.5, l=0.6</math> in all experiments. <br />
<br />
===Egg Quest===<br />
The goal for this quest is to find and open the jewel-encrusted egg hidden on a tree in the forest. The agent will get 100 points upon completing this task. For action space, there are 9 fixed actions for navigation, and a second subset which consisting <math display="inline">N_{Take}</math> actions for taking possible objects in the game. <math display="inline">N_{Take}=200 (set A_1), N_{Take}=300 (set A_2)</math> has been tested separately.<br />
AE-DQN (blue) and a vanilla DQN agent (green) has been tested in this quest.<br />
<br />
[[File:AEF_zork_comparison.png|1200px|thumb|center|Performance of agents in the egg quest.]]<br />
<br />
Figure a) corresponds to the set <math display="inline">A_1</math>, with T=100, b) corresponds to the set <math display="inline">A_2</math>, with T=100, and c) corresponds to the set <math display="inline">A_2</math>, with T=200. Both agents has performed well on sets a and c. However the AE-DQN agent has learned much faster than the DQN on set b, which implies that action elimination is more robust to hyperparameter optimization when the action space is large. One important observation to note is that the three figures have different scales for the cumulative reward. While the AE-DQN outperformed the standard DQN in figure b, both models performed significantly better with the hyperparameter configuration in figure c.<br />
<br />
===Troll Quest===<br />
The goal of this quest is to find the troll. To do it the agent need to find the way to the house, use a lantern to expose the hidden entrance to the underworld. It will get 100 points upon achieving the goal. This quest is a larger problem than Egg Quest. The action set <math display="inline">A_1</math> is 200 take actions and 15 necessary actions, 215 in total.<br />
<br />
[[File:AEF_troll_comparison.png|400px|thumb|center|Results in the Troll Quest.]]<br />
<br />
The red line above is an "optimal elimination" baseline which consists of only 35 actions(15 essential, and 20 relevant take actions). We can see that AE-DQN still outperforms DQN and its improvement over DQN is more significant in the Troll Quest than the Egg quest. Also, it achieves compatible performance to the "optimal elimination" baseline.<br />
<br />
===Open Zork===<br />
Lastly, the "Open Zork" domain has been tested which only the environment reward has been used. 1M steps has been trained. Each trajectory terminates after T=200 steps. Two action sets have been used:<math display="inline">A_3</math>, the "Minimal Zork" action set, which is the minimal set of actions (131) that is required to solve the game. <math display="inline">A_4</math>, the "Open Zork" action set (1227) which composed of {Verb, Object} tuples for all the verbs and objects in the game.<br />
<br />
[[]]<br />
<br />
[[File:AEF_open_zork_comparison.png|600px|thumb|center|Results in "Open Zork".]]<br />
<br />
<br />
The above Figure shows the learning curve for both AE-DQN and DQN. We can see that AE-DQN (blue) still outperform the DQN (blue) in terms of speed and cumulative reward.<br />
<br />
=Conclusion=<br />
In this paper, the authors proposed a Deep Reinforcement Learning model for sub-optimal actions while performing Q-learning. Moreover, they showed that by eliminating actions, using linear contextual bandits with theoretical guarantees of convergence, the size of the action space is reduced, exploration is more effective, and learning is improved when tested on Zork, a text-based game.<br />
<br />
For future work the authors aim to investigate more sophisticated architectures and tackle learning shared representations for elimination and control which may boost performance on both tasks.<br />
<br />
They also hope to to investigate other mechanisms for action elimination, such as eliminating actions that result from low Q-values as in Even-Dar, Mannor, and Mansour, 2003.<br />
<br />
The authors also hope to generate elimination signals in real-world domains and achieve the purpose of eliminating the signal implicitly.<br />
<br />
=Critique=<br />
The paper is not a significant algorithmic contribution and it merely adds an extra layer of complexity to the most famous DQN algorithm. All the experimental domains considered in the paper are discrete action problems that have so many actions that it could have been easily extended to a continuous action problem. In continuous action space there are several policy gradient based RL algorithms that have provided stronger performances. The authors should have ideally compared their methods to such algorithms like PPO or DRPO.<br />
Even with the critique above, the paper presents mathematical/theoretical justifications of the methodology. Moreover, since the methodology is built on the standard RL framework, this means that other variant RL algorithms can apply the idea to decrease the complexity and increase the performance. Moreover, the there are some rooms for applying technical variantions for the algorithm.<br />
<br />
=Reference=</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=learn_what_not_to_learn&diff=41904learn what not to learn2018-11-29T22:00:33Z<p>Ka2khan: /* Introduction */</p>
<hr />
<div>=Introduction=<br />
In reinforcement learning, it is often difficult for an agent to learn when the action space is large, especially the difficulties from function approximation and exploration. In some cases many actions are irrelevant and it is sometimes easier for the algorithm to learn which action not to take. The paper proposes a new reinforcement learning approach for dealing with large action spaces based on action elimination by restricting the available actions in each state to a subset of the most likely ones. There is a core assumption being made in the proposed method that it is easier to predict which actions in each state are invalid or inferior and use that information for control. More specifically, it proposes a system that learns the approximation of a Q-function and concurrently learns to eliminate actions. The method utilizes an external elimination signal which incorporates domain-specific prior knowledge. For example, in parser-based text games, the parser gives feedback regarding irrelevant actions after the action is played (e.g., Player: "Climb the tree." Parser: "There are no trees to climb"). Then a machine learning model can be trained to generalize to unseen states. <br />
<br />
The paper focuses on tasks where both states and the actions are natural language. It introduces a novel deep reinforcement learning approach which has a Deep Q-Network (DQN) and an Action Elimination Network (AEN), both using the Convolutional Neural Networks (CNN) for Natural Language Processing (NLP) tasks. The AEN is trained to predict invalid actions, supervised by the elimination signal from the environment. The proposed method uses the final layer activations of AEN to build a linear contextual bandit model which allows the elimination of sub-optimal actions with high probability. '''Note that the core assumption is that it is easy to predict which actions are invalid or inferior in each state and leverage that information for control.'''<br />
<br />
The text-based game called "Zork", which lets players to interact with a virtual world through a text based interface, is tested by using the elimination framework. The AEN algorithm has achieved faster learning rate than the baseline agents through eliminating irrelevant actions.<br />
<br />
Below shows an example for the Zork interface:<br />
<br />
[[File:lnottol_fig1.png|500px|center]]<br />
<br />
All states and actions are given in natural language. Input for the game contains more than a thousand possible actions in each state since player can type anything.<br />
<br />
=Related Work=<br />
Text-Based Games(TBG): The state of the environment in TBG is described by simple language. The player interacts with the environment with text command which respects a pre-defined grammar. A popular example is Zork which has been tested in the paper. TBG is a good research intersection of RL and NLP, it requires language understanding, long-term memory, planning, exploration, affordability extraction and common sense. It also often introduce stochastic dynamics to increase randomness.<br />
<br />
Representations for TBG: Good word representation is necessary in order to learn control policies from texts. Previous work on TBG used pre-trained embeddings directly for control. other works combined pre-trained embedding with neural networks.<br />
<br />
DRL with linear function approximation: DRL methods such as the DQN have achieved state-of-the-art results in a variety of challenging, high-dimensional domains. This is mainly because neural networks can learn rich domain representations for value function and policy. On the other hand, linear representation batch reinforcement learning methods are more stable and accurate, while feature engineering is necessary.<br />
<br />
RL in Large Action Spaces: Prior work concentrated on factorizing the action space into binary subspace(Pazis and Parr, 2011; Dulac-Arnold et al., 2012; Lagoudakis and Parr, 2003), other works proposed to embed the discrete actions into a continuous space, then choose the nearest discrete action according to the optimal actions in the continuous space(Dulac-Arnold et al., 2015; Van Hasselt and Wiering, 2009). He et. al. (2015)extended DQN to unbounded(natural language) action spaces.<br />
Learning to eliminate actions was first mentioned by (Even-Dar, Mannor, and Mansour, 2003). They proposed to learn confidence intervals around the value function in each state. Lipton et al.(2016a) proposed to learn a classifier that detects hazardous state and then use it to shape the reward. Fulda et al.(2017) presented a method for affordability extraction via inner products of pre-trained word embedding.<br />
<br />
=Action Elimination=<br />
<br />
The approach in the paper builds on the standard Reinforcement Learning formulation. At each time step <math>t</math>, the agent observes state <math display="inline">s_t </math> and chooses a discrete action <math display="inline">a_t\in\{1,...,|A|\} </math>. Then the agent obtains a reward <math display="inline">r_t(s_t,a_t) </math> and observes next state <math display="inline">s_{t+1} </math> according to a transition kernel <math>P(s_{t+1}|s_t,a_t)</math>. The goal of the algorithm is to learn a policy <math display="inline">\pi(a|s) </math> which maximizes the expected future discounted cumulative return <math display="inline">V^\pi(s)=E^\pi[\sum_{t=0}^{\infty}\gamma^tr(s_t,a_t)|s_0=s]</math>, where <math> 0< \gamma <1 </math>. The Q-function is <math display="inline">Q^\pi(s,a)=E^\pi[\sum_{t=0}^{\infty}\gamma^tr(s_t,a_t)|s_0=s,a_0=a]</math>, and it can be optimized by Q-learning algorithm.<br />
<br />
After executing an action, the agent observes a binary elimination signal <math>e(s, a)</math> to determine which actions not to take. It equals 1 if action <math>a</math> may be eliminated in state <math>s</math> (and 0 otherwise). The signal helps mitigating the problem of large discrete action spaces. We start with the following definitions:<br />
<br />
'''Definition 1:''' <br />
<br />
Valid state-action pairs with respect to an elimination signal are state action pairs which the elimination process should not eliminate. <br />
<br />
The set of valid state-action pairs contains all of the state-action pairs that are a part of some optimal policy, i.e., only strictly suboptimal state-actions can be invalid.<br />
<br />
'''Definition 2:'''<br />
<br />
Admissible state-action pairs with respect to an elimination algorithm are state action pairs which the elimination algorithm does not eliminate.<br />
<br />
'''Definition 3:'''<br />
<br />
Action Elimination Q-learning is a Q-learning algorithm which updates only admissible state-action pairs and chooses the best action in the next state from its admissible actions. We allow the base Q-learning algorithm to be any algorithm that converges to <math display="inline">Q^*</math> with probability 1 after observing each state-action infinitely often.<br />
<br />
==Advantages of Action Elimination==<br />
The main advantages of action elimination is that it allows the agent to overcome some of the main difficulties in large action spaces which are Function Approximation and Sample Complexity. <br />
<br />
Function approximation: Errors in the Q-function estimates may cause the learning algorithm to converge to a suboptimal policy, this phenomenon becomes more noticeable when the action space is large. Action elimination mitigate this effect by taking the max operator only on valid actions, thus, reducing potential overestimation errors. Besides, by ignoring the invalid actions, the function approximation can also learn a simpler mapping (i.e., only the Q-values of the valid state-action pairs) leading to faster convergence and better solution.<br />
<br />
Sample complexity: The sample complexity measures the number of steps during learning, in which the policy is not <math display="inline">\epsilon</math>-optimal. Assume that there are <math>A'</math> actions that should be eliminated and are <math>\epsilon</math>-optimal, i.e. their value is at least <math>V^*(s)-\epsilon</math>. The invalid action often returns no reward and doesn't change the state, (Lattimore and Hutter, 2012)resulting in an action gap of <math display="inline">\epsilon=(1-\gamma)V^*(s)</math>, and this translates to <math display="inline">V^*(s)^{-2}(1-\gamma)^{-5}log(1/\delta)</math> wasted samples for learning each invalid state-action pair. Practically, elimination algorithm can eliminate these invalid actions and therefore speed up the learning process approximately by <math display="inline">A/A'</math>.<br />
<br />
Because it is difficult to embedde the elimination signal into the MDP, the authors use contextual multi-armed bandits to decouple the elimination signal from the MDP, which can correctly eliminate actions when applying standard Q learning into learning process.<br />
<br />
==Action elimination with contextual bandits==<br />
<br />
Let <math display="inline">x(s_t)\in R^d </math> be the feature representation of <math display="inline">s_t </math>. We assume that under this representation there exists a set of parameters <math display="inline">\theta_a^*\in R_d </math> such that the elimination signal in state <math display="inline">s_t </math> is <math display="inline">e_t(s_t,a) = \theta_a^Tx(s_t)+\eta_t </math>, where <math display="inline"> \Vert\theta_a^*\Vert_2\leq S</math>. <math display="inline">\eta_t</math> is an R-subgaussian random variable with zero mean that models additive noise to the elimination signal. When there is no noise in the elimination signal, R=0. Otherwise, <math display="inline">R\leq 1</math> since the elimination signal is bounded in [0,1]. Assume the elimination signal satisfies: <math display="inline">0\leq E[e_t(s_t,a)]\leq l </math> for any valid action and <math display="inline"> u\leq E[e_t(s_t, a)]\leq 1</math> for any invalid action. And <math display="inline"> l\leq u</math>. Denote by <math display="inline">X_{t,a}</math> as the matrix whose rows are the observed state representation vectors in which action a was chosen, up to time t. <math display="inline">E_{t,a}</math> as the vector whose elements are the observed state representation elimination signals in which action a was chosen, up to time t. Denote the solution to the regularized linear regression <math display="inline">\Vert X_{t,a}\theta_{t,a}-E_{t,a}\Vert_2^2+\lambda\Vert \theta_{t,a}\Vert_2^2 </math> (for some <math display="inline">\lambda>0</math>) by <math display="inline">\hat{\theta}_{t,a}=\bar{V}_{t,a}^{-1}X_{t,a}^TE_{t,a} </math>, where <math display="inline">\bar{V}_{t,a}=\lambda I + X_{t,a}^TX_{t,a}</math>.<br />
<br />
<br />
According to Theorem 2 in (Abbasi-Yadkori, Pal, and Szepesvari, 2011), <math display="inline">|\hat{\theta}_{t,a}^{T}x(s_t)-\theta_a^{*T}x(s_t)|\leq\sqrt{\beta_t(\delta)x(s_t)^T\bar{V}_{t,a}^{-1}x(s_t)} \forall t>0</math>, where <math display="inline">\sqrt{\beta_t(\delta)}=R\sqrt{2log(det(\bar{V}_{t,a}^{1/2})det(\lambda I)^{-1/2}/\delta)}+\lambda^{1/2}S</math>, with probability of at least <math display="inline">1-\delta</math>. If <math display="inline">\forall s \Vert x(s)\Vert_2 \leq L</math>, then <math display="inline">\beta_t</math> can be bounded by <math display="inline">\sqrt{\beta_t(\delta)} \leq R \sqrt{dlog(1+tL^2/\lambda/\delta)}+\lambda^{1/2}S</math>. Next, define <math display="inline">\tilde{\delta}=\delta/k</math> and bound this probability for all the actions. i.e., <math display="inline">\forall a,t>0</math><br />
<br />
<math display="inline">Pr(|\hat{\theta}_{t,a}^{T}x(s_t)-\theta_a^{*T}x(s_t)|\leq\sqrt{\beta_t(\delta)x(s_t)^T\bar{V}_{t,a}^{-1}x(s_t)}) \leq 1-\delta</math><br />
<br />
Recall that <math display="inline">E[e_t(s,a)]=\theta_a^{*T}x(s_t)\leq l</math> if a is a valid action. Then we can eliminate action a at state <math display="inline">s_t</math> if it satisfies:<br />
<br />
<math display="inline">\hat{\theta}_{t,a}^{T}x(s_t)-\sqrt{\beta_t(\delta)x(s_t)^T\bar{V}_{t,a}^{-1}x(s_t)})>l</math><br />
<br />
with probability <math display="inline">1-\delta</math> that we never eliminate any valid action. Note that <math display="inline">l, u</math> are not known. In practice, choosing <math display="inline">l</math> to be 0.5 should suffice.<br />
<br />
==Concurrent Learning==<br />
In fact, Q-learning and contextual bandit algorithms can learn simultaneously, resulting in the convergence of both algorithms, i.e., finding an optimal policy and a minimal valid action space. <br />
<br />
If the elimination is done based on the concentration bounds of the linear contextual bandits, it can be ensured that Action Elimination Q-learning converges, as shown in Proposition 1.<br />
<br />
'''Proposition 1:'''<br />
<br />
Assume that all state action pairs (s,a) are visited infinitely often, unless eliminated according to <math display="inline">\hat{\theta}_{t-1,a}^Tx(s)-\sqrt{\beta_{t-1}(\tilde{\delta})x(s)^T\bar{V}_{t-1,a}^{-1}x(s))}>l</math>. Then, with a probability of at least <math display="inline">1-\delta</math>, action elimination Q-learning converges to the optimal Q-function for any valid state-action pairs. In addition, actions which should be eliminated are visited at most <math display="inline">T_{s,a}(t)\leq 4\beta_t/(u-l)^2<br />
+1</math> times.<br />
<br />
Notice that when there is no noise in the elimination signal(R=0), we correctly eliminate actions with probability 1. so invalid actions will be sampled a finite number of times.<br />
<br />
=Method=<br />
The assumption that <math display="inline">e_t(s_t,a)=\theta_a^{*T}x(s_t)+\eta_t </math> might not hold when using raw features like word2vec. So the paper proposes to use the neural network's last layer as features. A practical challenge here is that the features must be fixed over time when used by the contextual bandit. So batch-updates framework(Levine et al., 2017;Riquelme, Tucker, and Snoek, 2018) is used, where a new contextual bandit model is learned for every few steps that uses the last layer activation of the AEN as features.<br />
<br />
==Architecture of action elimination framework==<br />
<br />
[[File:lnottol_fig1b.png|300px|center]]<br />
<br />
After taking action <math display="inline">a_t</math>, the agent observes <math display="inline">(r_t,s_{t+1},e_t)</math>. The agent use it to learn two function approximation deep neural networks: A DQN and an AEN. AEN provides an admissible actions set <math display="inline">A'</math> to the DQN, which uses this set to decide how to act and learn. The architecture for both the AEN and DQN is an NLP CNN(100 convolutional filters for AEN and 500 for DQN, with three different 1D kernels of length (1,2,3)), based on(Kim, 2014). The state is represented as a sequence of words, composed of the game descriptor and the player's inventory. These are truncated or zero padded to a length of 50 descriptor + 15 inventory words and each word is embedded into continuous vectors using word2vec in <math display="inline">R^{300}</math>. The features of the last four states are then concatenated together such that the final state representations s are in <math display="inline">R^{78000}</math>. The AEN is trained to minimize the MSE loss, using the elimination signal as a label. The code, the Zork domain, and the implementation of the elimination signal can be found [https://github.com/TomZahavy/CB_AE_DQN here.]<br />
<br />
==Psuedocode of the Algorithm==<br />
<br />
[[File:lnottol_fig2.png|750px|center]]<br />
<br />
AE-DQN trains two networks: a DQN denoted by Q and an AEN denoted by E. The algorithm creates a linear contextual bandit model from it every L iterations with procedure AENUpdate(). This procedure uses the activations of the last hidden layer of E as features, which are then used to create a contextual linear bandit model.AENUpdate() then solved this model and plugin it into the target AEN. The contextual linear bandit model <math display="inline">(E^-,V)</math> is then used to eliminate actions via the ACT() and Target() functions. ACT() follows an <math display="inline">\epsilon</math>-greedy mechanism on the admissible actions set. For exploitation, it selects the action with highest Q-value by taking an argmax on Q-values among <math display="inline">A'</math>. For exploration, it selects an action uniformly from <math display="inline">A'</math>. The targets() procedure is estimating the value function by taking max over Q-values only among admissible actions, hence, reducing function approximation errors.<br />
<br />
=Experiment=<br />
==Zork domain==<br />
The world of Zork presents a rich environment with a large state and action space. <br />
Zork players describe their actions using natural language instructions. For example, "open the mailbox". Then their actions were processed by a sophisticated natural language parser. Based on the results, the game presents the outcome of the action. The goal of Zork is to collect the Twenty Treasures of Zork and install them in the trophy case. Points that are generated from the game's scoring system are given to the agent as the reward. For example, the player gets the points when solving the puzzles. Placing all treasures in the trophy will get 350 points. The elimination signal is given in two forms, "wrong parse" flag, and text feedback "you cannot take that". These two signals are grouped together into a single binary signal which then provided to the algorithm. <br />
<br />
Experiments begin with the two subdomains of Zork domains: Egg Quest and the Troll Quest. For these subdomains, an additional reward signal is provided to guide the agent towards solving specific tasks and make the results more visible. A reward of -1 is applied at every time step to encourage the agent to favor short paths. Each trajectory terminates is upon completing the quest or after T steps are taken. The discounted factor for training is <math display="inline">\gamma=0.8</math> and <math display="inline">\gamma=1</math> during evaluation. Also <math display="inline">\beta=0.5, l=0.6</math> in all experiments. <br />
<br />
===Egg Quest===<br />
The goal for this quest is to find and open the jewel-encrusted egg hidden on a tree in the forest. The agent will get 100 points upon completing this task. For action space, there are 9 fixed actions for navigation, and a second subset which consisting <math display="inline">N_{Take}</math> actions for taking possible objects in the game. <math display="inline">N_{Take}=200 (set A_1), N_{Take}=300 (set A_2)</math> has been tested separately.<br />
AE-DQN (blue) and a vanilla DQN agent (green) has been tested in this quest.<br />
<br />
[[File:AEF_zork_comparison.png|1200px|thumb|center|Performance of agents in the egg quest.]]<br />
<br />
Figure a) corresponds to the set <math display="inline">A_1</math>, with T=100, b) corresponds to the set <math display="inline">A_2</math>, with T=100, and c) corresponds to the set <math display="inline">A_2</math>, with T=200. Both agents has performed well on sets a and c. However the AE-DQN agent has learned much faster than the DQN on set b, which implies that action elimination is more robust to hyperparameter optimization when the action space is large. One important observation to note is that the three figures have different scales for the cumulative reward. While the AE-DQN outperformed the standard DQN in figure b, both models performed significantly better with the hyperparameter configuration in figure c.<br />
<br />
===Troll Quest===<br />
The goal of this quest is to find the troll. To do it the agent need to find the way to the house, use a lantern to expose the hidden entrance to the underworld. It will get 100 points upon achieving the goal. This quest is a larger problem than Egg Quest. The action set <math display="inline">A_1</math> is 200 take actions and 15 necessary actions, 215 in total.<br />
<br />
[[File:AEF_troll_comparison.png|400px|thumb|center|Results in the Troll Quest.]]<br />
<br />
The red line above is an "optimal elimination" baseline which consists of only 35 actions(15 essential, and 20 relevant take actions). We can see that AE-DQN still outperforms DQN and its improvement over DQN is more significant in the Troll Quest than the Egg quest. Also, it achieves compatible performance to the "optimal elimination" baseline.<br />
<br />
===Open Zork===<br />
Lastly, the "Open Zork" domain has been tested which only the environment reward has been used. 1M steps has been trained. Each trajectory terminates after T=200 steps. Two action sets have been used:<math display="inline">A_3</math>, the "Minimal Zork" action set, which is the minimal set of actions (131) that is required to solve the game. <math display="inline">A_4</math>, the "Open Zork" action set (1227) which composed of {Verb, Object} tuples for all the verbs and objects in the game.<br />
<br />
[[]]<br />
<br />
[[File:AEF_open_zork_comparison.png|600px|thumb|center|Results in "Open Zork".]]<br />
<br />
<br />
The above Figure shows the learning curve for both AE-DQN and DQN. We can see that AE-DQN (blue) still outperform the DQN (blue) in terms of speed and cumulative reward.<br />
<br />
=Conclusion=<br />
In this paper, the authors proposed a Deep Reinforcement Learning model for sub-optimal actions while performing Q-learning. Moreover, they showed that by eliminating actions, using linear contextual bandits with theoretical guarantees of convergence, the size of the action space is reduced, exploration is more effective, and learning is improved when tested on Zork, a text-based game.<br />
<br />
For future work the authors aim to investigate more sophisticated architectures and tackle learning shared representations for elimination and control which may boost performance on both tasks.<br />
<br />
They also hope to to investigate other mechanisms for action elimination, such as eliminating actions that result from low Q-values as in Even-Dar, Mannor, and Mansour, 2003.<br />
<br />
The authors also hope to generate elimination signals in real-world domains and achieve the purpose of eliminating the signal implicitly.<br />
<br />
=Critique=<br />
The paper is not a significant algorithmic contribution and it merely adds an extra layer of complexity to the most famous DQN algorithm. All the experimental domains considered in the paper are discrete action problems that have so many actions that it could have been easily extended to a continuous action problem. In continuous action space there are several policy gradient based RL algorithms that have provided stronger performances. The authors should have ideally compared their methods to such algorithms like PPO or DRPO.<br />
Even with the critique above, the paper presents mathematical/theoretical justifications of the methodology. Moreover, since the methodology is built on the standard RL framework, this means that other variant RL algorithms can apply the idea to decrease the complexity and increase the performance. Moreover, the there are some rooms for applying technical variantions for the algorithm.<br />
<br />
=Reference=</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=learn_what_not_to_learn&diff=41902learn what not to learn2018-11-29T21:54:48Z<p>Ka2khan: /* Introduction */</p>
<hr />
<div>=Introduction=<br />
In reinforcement learning, it is often difficult for an agent to learn when the action space is large, especially the difficulties from function approximation and exploration. In some cases many actions are irrelevant and it is sometimes easier for the algorithm to learn which action not to take. The paper proposes a new reinforcement learning approach for dealing with large action spaces based on action elimination by restricting the available actions in each state to a subset of the most likely ones. There is a core assumption being made in the proposed method that it is easier to predict which actions in each state are invalid or inferior and use that information for control. More specifically, it proposes a system that learns the approximation of a Q-function and concurrently learns to eliminate actions. The method utilizes an external elimination signal which incorporates domain-specific prior knowledge. For example, in parser-based text games, the parser gives feedback regarding irrelevant actions after the action is played (e.g., Player: "Climb the tree." Parser: "There are no trees to climb"). Then a machine learning model can be trained to generalize to unseen states. <br />
<br />
The paper focuses on tasks where both states and the actions are natural language. It introduces a novel deep reinforcement learning approach which has a Deep Q-Network (DQN) and an Action Elimination Network (AEN), both using the Convolutional Neural Networks (CNN) for Natural Language Processing (NLP) tasks. The AEN is trained to predict invalid actions, supervised by the elimination signal from the environment. '''Note that the core assumption is that it is easy to predict which actions are invalid or inferior in each state and leverage that information for control.'''<br />
<br />
The text-based game called "Zork", which lets players to interact with a virtual world through a text based interface, is tested by using the elimination framework. The AEN algorithm has achieved faster learning rate than the baseline agents through eliminating irrelevant actions.<br />
<br />
Below shows an example for the Zork interface:<br />
<br />
[[File:lnottol_fig1.png|500px|center]]<br />
<br />
All states and actions are given in natural language. Input for the game contains more than a thousand possible actions in each state since player can type anything.<br />
<br />
=Related Work=<br />
Text-Based Games(TBG): The state of the environment in TBG is described by simple language. The player interacts with the environment with text command which respects a pre-defined grammar. A popular example is Zork which has been tested in the paper. TBG is a good research intersection of RL and NLP, it requires language understanding, long-term memory, planning, exploration, affordability extraction and common sense. It also often introduce stochastic dynamics to increase randomness.<br />
<br />
Representations for TBG: Good word representation is necessary in order to learn control policies from texts. Previous work on TBG used pre-trained embeddings directly for control. other works combined pre-trained embedding with neural networks.<br />
<br />
DRL with linear function approximation: DRL methods such as the DQN have achieved state-of-the-art results in a variety of challenging, high-dimensional domains. This is mainly because neural networks can learn rich domain representations for value function and policy. On the other hand, linear representation batch reinforcement learning methods are more stable and accurate, while feature engineering is necessary.<br />
<br />
RL in Large Action Spaces: Prior work concentrated on factorizing the action space into binary subspace(Pazis and Parr, 2011; Dulac-Arnold et al., 2012; Lagoudakis and Parr, 2003), other works proposed to embed the discrete actions into a continuous space, then choose the nearest discrete action according to the optimal actions in the continuous space(Dulac-Arnold et al., 2015; Van Hasselt and Wiering, 2009). He et. al. (2015)extended DQN to unbounded(natural language) action spaces.<br />
Learning to eliminate actions was first mentioned by (Even-Dar, Mannor, and Mansour, 2003). They proposed to learn confidence intervals around the value function in each state. Lipton et al.(2016a) proposed to learn a classifier that detects hazardous state and then use it to shape the reward. Fulda et al.(2017) presented a method for affordability extraction via inner products of pre-trained word embedding.<br />
<br />
=Action Elimination=<br />
<br />
The approach in the paper builds on the standard Reinforcement Learning formulation. At each time step <math>t</math>, the agent observes state <math display="inline">s_t </math> and chooses a discrete action <math display="inline">a_t\in\{1,...,|A|\} </math>. Then the agent obtains a reward <math display="inline">r_t(s_t,a_t) </math> and observes next state <math display="inline">s_{t+1} </math> according to a transition kernel <math>P(s_{t+1}|s_t,a_t)</math>. The goal of the algorithm is to learn a policy <math display="inline">\pi(a|s) </math> which maximizes the expected future discounted cumulative return <math display="inline">V^\pi(s)=E^\pi[\sum_{t=0}^{\infty}\gamma^tr(s_t,a_t)|s_0=s]</math>, where <math> 0< \gamma <1 </math>. The Q-function is <math display="inline">Q^\pi(s,a)=E^\pi[\sum_{t=0}^{\infty}\gamma^tr(s_t,a_t)|s_0=s,a_0=a]</math>, and it can be optimized by Q-learning algorithm.<br />
<br />
After executing an action, the agent observes a binary elimination signal <math>e(s, a)</math> to determine which actions not to take. It equals 1 if action <math>a</math> may be eliminated in state <math>s</math> (and 0 otherwise). The signal helps mitigating the problem of large discrete action spaces. We start with the following definitions:<br />
<br />
'''Definition 1:''' <br />
<br />
Valid state-action pairs with respect to an elimination signal are state action pairs which the elimination process should not eliminate. <br />
<br />
The set of valid state-action pairs contains all of the state-action pairs that are a part of some optimal policy, i.e., only strictly suboptimal state-actions can be invalid.<br />
<br />
'''Definition 2:'''<br />
<br />
Admissible state-action pairs with respect to an elimination algorithm are state action pairs which the elimination algorithm does not eliminate.<br />
<br />
'''Definition 3:'''<br />
<br />
Action Elimination Q-learning is a Q-learning algorithm which updates only admissible state-action pairs and chooses the best action in the next state from its admissible actions. We allow the base Q-learning algorithm to be any algorithm that converges to <math display="inline">Q^*</math> with probability 1 after observing each state-action infinitely often.<br />
<br />
==Advantages of Action Elimination==<br />
The main advantages of action elimination is that it allows the agent to overcome some of the main difficulties in large action spaces which are Function Approximation and Sample Complexity. <br />
<br />
Function approximation: Errors in the Q-function estimates may cause the learning algorithm to converge to a suboptimal policy, this phenomenon becomes more noticeable when the action space is large. Action elimination mitigate this effect by taking the max operator only on valid actions, thus, reducing potential overestimation errors. Besides, by ignoring the invalid actions, the function approximation can also learn a simpler mapping (i.e., only the Q-values of the valid state-action pairs) leading to faster convergence and better solution.<br />
<br />
Sample complexity: The sample complexity measures the number of steps during learning, in which the policy is not <math display="inline">\epsilon</math>-optimal. Assume that there are <math>A'</math> actions that should be eliminated and are <math>\epsilon</math>-optimal, i.e. their value is at least <math>V^*(s)-\epsilon</math>. The invalid action often returns no reward and doesn't change the state, (Lattimore and Hutter, 2012)resulting in an action gap of <math display="inline">\epsilon=(1-\gamma)V^*(s)</math>, and this translates to <math display="inline">V^*(s)^{-2}(1-\gamma)^{-5}log(1/\delta)</math> wasted samples for learning each invalid state-action pair. Practically, elimination algorithm can eliminate these invalid actions and therefore speed up the learning process approximately by <math display="inline">A/A'</math>.<br />
<br />
Because it is difficult to embedde the elimination signal into the MDP, the authors use contextual multi-armed bandits to decouple the elimination signal from the MDP, which can correctly eliminate actions when applying standard Q learning into learning process.<br />
<br />
==Action elimination with contextual bandits==<br />
<br />
Let <math display="inline">x(s_t)\in R^d </math> be the feature representation of <math display="inline">s_t </math>. We assume that under this representation there exists a set of parameters <math display="inline">\theta_a^*\in R_d </math> such that the elimination signal in state <math display="inline">s_t </math> is <math display="inline">e_t(s_t,a) = \theta_a^Tx(s_t)+\eta_t </math>, where <math display="inline"> \Vert\theta_a^*\Vert_2\leq S</math>. <math display="inline">\eta_t</math> is an R-subgaussian random variable with zero mean that models additive noise to the elimination signal. When there is no noise in the elimination signal, R=0. Otherwise, <math display="inline">R\leq 1</math> since the elimination signal is bounded in [0,1]. Assume the elimination signal satisfies: <math display="inline">0\leq E[e_t(s_t,a)]\leq l </math> for any valid action and <math display="inline"> u\leq E[e_t(s_t, a)]\leq 1</math> for any invalid action. And <math display="inline"> l\leq u</math>. Denote by <math display="inline">X_{t,a}</math> as the matrix whose rows are the observed state representation vectors in which action a was chosen, up to time t. <math display="inline">E_{t,a}</math> as the vector whose elements are the observed state representation elimination signals in which action a was chosen, up to time t. Denote the solution to the regularized linear regression <math display="inline">\Vert X_{t,a}\theta_{t,a}-E_{t,a}\Vert_2^2+\lambda\Vert \theta_{t,a}\Vert_2^2 </math> (for some <math display="inline">\lambda>0</math>) by <math display="inline">\hat{\theta}_{t,a}=\bar{V}_{t,a}^{-1}X_{t,a}^TE_{t,a} </math>, where <math display="inline">\bar{V}_{t,a}=\lambda I + X_{t,a}^TX_{t,a}</math>.<br />
<br />
<br />
According to Theorem 2 in (Abbasi-Yadkori, Pal, and Szepesvari, 2011), <math display="inline">|\hat{\theta}_{t,a}^{T}x(s_t)-\theta_a^{*T}x(s_t)|\leq\sqrt{\beta_t(\delta)x(s_t)^T\bar{V}_{t,a}^{-1}x(s_t)} \forall t>0</math>, where <math display="inline">\sqrt{\beta_t(\delta)}=R\sqrt{2log(det(\bar{V}_{t,a}^{1/2})det(\lambda I)^{-1/2}/\delta)}+\lambda^{1/2}S</math>, with probability of at least <math display="inline">1-\delta</math>. If <math display="inline">\forall s \Vert x(s)\Vert_2 \leq L</math>, then <math display="inline">\beta_t</math> can be bounded by <math display="inline">\sqrt{\beta_t(\delta)} \leq R \sqrt{dlog(1+tL^2/\lambda/\delta)}+\lambda^{1/2}S</math>. Next, define <math display="inline">\tilde{\delta}=\delta/k</math> and bound this probability for all the actions. i.e., <math display="inline">\forall a,t>0</math><br />
<br />
<math display="inline">Pr(|\hat{\theta}_{t,a}^{T}x(s_t)-\theta_a^{*T}x(s_t)|\leq\sqrt{\beta_t(\delta)x(s_t)^T\bar{V}_{t,a}^{-1}x(s_t)}) \leq 1-\delta</math><br />
<br />
Recall that <math display="inline">E[e_t(s,a)]=\theta_a^{*T}x(s_t)\leq l</math> if a is a valid action. Then we can eliminate action a at state <math display="inline">s_t</math> if it satisfies:<br />
<br />
<math display="inline">\hat{\theta}_{t,a}^{T}x(s_t)-\sqrt{\beta_t(\delta)x(s_t)^T\bar{V}_{t,a}^{-1}x(s_t)})>l</math><br />
<br />
with probability <math display="inline">1-\delta</math> that we never eliminate any valid action. Note that <math display="inline">l, u</math> are not known. In practice, choosing <math display="inline">l</math> to be 0.5 should suffice.<br />
<br />
==Concurrent Learning==<br />
In fact, Q-learning and contextual bandit algorithms can learn simultaneously, resulting in the convergence of both algorithms, i.e., finding an optimal policy and a minimal valid action space. <br />
<br />
If the elimination is done based on the concentration bounds of the linear contextual bandits, it can be ensured that Action Elimination Q-learning converges, as shown in Proposition 1.<br />
<br />
'''Proposition 1:'''<br />
<br />
Assume that all state action pairs (s,a) are visited infinitely often, unless eliminated according to <math display="inline">\hat{\theta}_{t-1,a}^Tx(s)-\sqrt{\beta_{t-1}(\tilde{\delta})x(s)^T\bar{V}_{t-1,a}^{-1}x(s))}>l</math>. Then, with a probability of at least <math display="inline">1-\delta</math>, action elimination Q-learning converges to the optimal Q-function for any valid state-action pairs. In addition, actions which should be eliminated are visited at most <math display="inline">T_{s,a}(t)\leq 4\beta_t/(u-l)^2<br />
+1</math> times.<br />
<br />
Notice that when there is no noise in the elimination signal(R=0), we correctly eliminate actions with probability 1. so invalid actions will be sampled a finite number of times.<br />
<br />
=Method=<br />
The assumption that <math display="inline">e_t(s_t,a)=\theta_a^{*T}x(s_t)+\eta_t </math> might not hold when using raw features like word2vec. So the paper proposes to use the neural network's last layer as features. A practical challenge here is that the features must be fixed over time when used by the contextual bandit. So batch-updates framework(Levine et al., 2017;Riquelme, Tucker, and Snoek, 2018) is used, where a new contextual bandit model is learned for every few steps that uses the last layer activation of the AEN as features.<br />
<br />
==Architecture of action elimination framework==<br />
<br />
[[File:lnottol_fig1b.png|300px|center]]<br />
<br />
After taking action <math display="inline">a_t</math>, the agent observes <math display="inline">(r_t,s_{t+1},e_t)</math>. The agent use it to learn two function approximation deep neural networks: A DQN and an AEN. AEN provides an admissible actions set <math display="inline">A'</math> to the DQN, which uses this set to decide how to act and learn. The architecture for both the AEN and DQN is an NLP CNN(100 convolutional filters for AEN and 500 for DQN, with three different 1D kernels of length (1,2,3)), based on(Kim, 2014). The state is represented as a sequence of words, composed of the game descriptor and the player's inventory. These are truncated or zero padded to a length of 50 descriptor + 15 inventory words and each word is embedded into continuous vectors using word2vec in <math display="inline">R^{300}</math>. The features of the last four states are then concatenated together such that the final state representations s are in <math display="inline">R^{78000}</math>. The AEN is trained to minimize the MSE loss, using the elimination signal as a label. The code, the Zork domain, and the implementation of the elimination signal can be found [https://github.com/TomZahavy/CB_AE_DQN here.]<br />
<br />
==Psuedocode of the Algorithm==<br />
<br />
[[File:lnottol_fig2.png|750px|center]]<br />
<br />
AE-DQN trains two networks: a DQN denoted by Q and an AEN denoted by E. The algorithm creates a linear contextual bandit model from it every L iterations with procedure AENUpdate(). This procedure uses the activations of the last hidden layer of E as features, which are then used to create a contextual linear bandit model.AENUpdate() then solved this model and plugin it into the target AEN. The contextual linear bandit model <math display="inline">(E^-,V)</math> is then used to eliminate actions via the ACT() and Target() functions. ACT() follows an <math display="inline">\epsilon</math>-greedy mechanism on the admissible actions set. For exploitation, it selects the action with highest Q-value by taking an argmax on Q-values among <math display="inline">A'</math>. For exploration, it selects an action uniformly from <math display="inline">A'</math>. The targets() procedure is estimating the value function by taking max over Q-values only among admissible actions, hence, reducing function approximation errors.<br />
<br />
=Experiment=<br />
==Zork domain==<br />
The world of Zork presents a rich environment with a large state and action space. <br />
Zork players describe their actions using natural language instructions. For example, "open the mailbox". Then their actions were processed by a sophisticated natural language parser. Based on the results, the game presents the outcome of the action. The goal of Zork is to collect the Twenty Treasures of Zork and install them in the trophy case. Points that are generated from the game's scoring system are given to the agent as the reward. For example, the player gets the points when solving the puzzles. Placing all treasures in the trophy will get 350 points. The elimination signal is given in two forms, "wrong parse" flag, and text feedback "you cannot take that". These two signals are grouped together into a single binary signal which then provided to the algorithm. <br />
<br />
Experiments begin with the two subdomains of Zork domains: Egg Quest and the Troll Quest. For these subdomains, an additional reward signal is provided to guide the agent towards solving specific tasks and make the results more visible. A reward of -1 is applied at every time step to encourage the agent to favor short paths. Each trajectory terminates is upon completing the quest or after T steps are taken. The discounted factor for training is <math display="inline">\gamma=0.8</math> and <math display="inline">\gamma=1</math> during evaluation. Also <math display="inline">\beta=0.5, l=0.6</math> in all experiments. <br />
<br />
===Egg Quest===<br />
The goal for this quest is to find and open the jewel-encrusted egg hidden on a tree in the forest. The agent will get 100 points upon completing this task. For action space, there are 9 fixed actions for navigation, and a second subset which consisting <math display="inline">N_{Take}</math> actions for taking possible objects in the game. <math display="inline">N_{Take}=200 (set A_1), N_{Take}=300 (set A_2)</math> has been tested separately.<br />
AE-DQN (blue) and a vanilla DQN agent (green) has been tested in this quest.<br />
<br />
[[File:AEF_zork_comparison.png|1200px|thumb|center|Performance of agents in the egg quest.]]<br />
<br />
Figure a) corresponds to the set <math display="inline">A_1</math>, with T=100, b) corresponds to the set <math display="inline">A_2</math>, with T=100, and c) corresponds to the set <math display="inline">A_2</math>, with T=200. Both agents has performed well on sets a and c. However the AE-DQN agent has learned much faster than the DQN on set b, which implies that action elimination is more robust to hyperparameter optimization when the action space is large. One important observation to note is that the three figures have different scales for the cumulative reward. While the AE-DQN outperformed the standard DQN in figure b, both models performed significantly better with the hyperparameter configuration in figure c.<br />
<br />
===Troll Quest===<br />
The goal of this quest is to find the troll. To do it the agent need to find the way to the house, use a lantern to expose the hidden entrance to the underworld. It will get 100 points upon achieving the goal. This quest is a larger problem than Egg Quest. The action set <math display="inline">A_1</math> is 200 take actions and 15 necessary actions, 215 in total.<br />
<br />
[[File:AEF_troll_comparison.png|400px|thumb|center|Results in the Troll Quest.]]<br />
<br />
The red line above is an "optimal elimination" baseline which consists of only 35 actions(15 essential, and 20 relevant take actions). We can see that AE-DQN still outperforms DQN and its improvement over DQN is more significant in the Troll Quest than the Egg quest. Also, it achieves compatible performance to the "optimal elimination" baseline.<br />
<br />
===Open Zork===<br />
Lastly, the "Open Zork" domain has been tested which only the environment reward has been used. 1M steps has been trained. Each trajectory terminates after T=200 steps. Two action sets have been used:<math display="inline">A_3</math>, the "Minimal Zork" action set, which is the minimal set of actions (131) that is required to solve the game. <math display="inline">A_4</math>, the "Open Zork" action set (1227) which composed of {Verb, Object} tuples for all the verbs and objects in the game.<br />
<br />
[[]]<br />
<br />
[[File:AEF_open_zork_comparison.png|600px|thumb|center|Results in "Open Zork".]]<br />
<br />
<br />
The above Figure shows the learning curve for both AE-DQN and DQN. We can see that AE-DQN (blue) still outperform the DQN (blue) in terms of speed and cumulative reward.<br />
<br />
=Conclusion=<br />
In this paper, the authors proposed a Deep Reinforcement Learning model for sub-optimal actions while performing Q-learning. Moreover, they showed that by eliminating actions, using linear contextual bandits with theoretical guarantees of convergence, the size of the action space is reduced, exploration is more effective, and learning is improved when tested on Zork, a text-based game.<br />
<br />
For future work the authors aim to investigate more sophisticated architectures and tackle learning shared representations for elimination and control which may boost performance on both tasks.<br />
<br />
They also hope to to investigate other mechanisms for action elimination, such as eliminating actions that result from low Q-values as in Even-Dar, Mannor, and Mansour, 2003.<br />
<br />
The authors also hope to generate elimination signals in real-world domains and achieve the purpose of eliminating the signal implicitly.<br />
<br />
=Critique=<br />
The paper is not a significant algorithmic contribution and it merely adds an extra layer of complexity to the most famous DQN algorithm. All the experimental domains considered in the paper are discrete action problems that have so many actions that it could have been easily extended to a continuous action problem. In continuous action space there are several policy gradient based RL algorithms that have provided stronger performances. The authors should have ideally compared their methods to such algorithms like PPO or DRPO.<br />
Even with the critique above, the paper presents mathematical/theoretical justifications of the methodology. Moreover, since the methodology is built on the standard RL framework, this means that other variant RL algorithms can apply the idea to decrease the complexity and increase the performance. Moreover, the there are some rooms for applying technical variantions for the algorithm.<br />
<br />
=Reference=</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=CapsuleNets&diff=41792CapsuleNets2018-11-29T04:15:07Z<p>Ka2khan: /* Dynamic Routing */</p>
<hr />
<div>The paper "Dynamic Routing Between Capsules" was written by three researchers at Google Brain: Sara Sabour, Nicholas Frosst, and Geoffrey E. Hinton. This paper was published and presented at the 31st Conference on Neural Information Processing Systems (NIPS 2017) in Long Beach, California. The same three researchers recently published a highly related paper "Matrix Capsules with EM Routing" for ICLR 2018.<br />
<br />
=Motivation=<br />
<br />
Ever since AlexNet eclipsed the performance of competing architectures in the 2012 ImageNet challenge, convolutional neural networks have maintained their dominance in computer vision applications. Despite the recent successes and innovations brought about by convolutional neural networks, some assumptions made in these networks are perhaps unwarranted and deficient. Using a novel neural network architecture, the authors create CapsuleNets, a network that they claim is able to learn image representations in a more robust, human-like manner. With only a 3 layer capsule network, they achieved near state-of-the-art results on MNIST.<br />
==Adversarial Examples==<br />
<br />
First discussed by Christian Szegedy et. al. in late 2013, adversarial examples have been heavily discussed by the deep learning community as a potential security threat to AI learning. Adversarial examples are defined as inputs that an attacker creates intentionally fool a machine learning model. An example of an adversarial example is shown below: <br />
<br />
[[File:adversarial_img_1.png |center]]<br />
To the human eye, the image appears to be a panda both before and after noise is injected into the image, whereas the trained ConvNet model discerns the noisy image as a Gibbon with almost 100% certainty. The fact that the network is unable to classify the above image as a panda after the epsilon perturbation leads to many potential security risks in AI dependent systems such as self-driving vehicles. Although various methods have been suggested to combat adversarial examples, robust defences are hard to construct due to the inherent difficulties in constructing theoretical models for the adversarial example crafting process. However, beyond the fact that these examples may serve as a security threat, it emphasizes that these convolutional neural networks do not learn image classification/object detection patterns the same way that a human would. Rather than identifying the core features of a panda such as: its eyes, mouth, nose, and the gradient changes in its black/white fur, the convolutional neural network seems to be learning image representations in a completely different manner. Deep learning researchers often attempt to model neural networks after human learning, and it is clear that further steps must be taken to robustify ConvNets against targeted noise perturbations.<br />
<br />
==Drawbacks of CNNs==<br />
Hinton claims that the key fault with traditional CNNs lies within the pooling function. Although pooling builds translational invariance into the network, it fails to preserve spatial relationships between objects. When we pool, we effectively reduce a kxk kernel of convolved cells into a scalar input. This results in a desired local invariance without inhibiting the network's ability to detect features, but causes valuable spatial information to be lost.<br />
<br />
In the example below, the network is able to detect the similar features (eyes, mouth, nose, etc) within both images, but fails to recognize that one image is a human face, while the other is a Picasso-esque due to the CNN's inability to encode spatial relationships after multiple pooling layers.<br />
<br />
<br />
[[File:Equivariance Face.png |center]]<br />
<br />
Conversely, we hope that a CNN can recognize that both of the following pictures contain a kitten. Unfortunately, when we feed the two images into a ResNet50 architecture, only the first image is correctly classified, while the second image is predicted to be a guinea pig.<br />
<br />
<br />
[[File:kitten.jpeg |center]]<br />
<br />
<br />
[[File:kitten-rotated-180.jpg |center]]<br />
<br />
For a more in depth discussion on the problems with ConvNets, please listen to Geoffrey Hinton's talk "What is wrong with convolutional neural nets?" given at MIT during the Brain & Cognitive Sciences - Fall Colloquium Series (December 4, 2014).<br />
<br />
==Intuition for Capsules==<br />
Human vision ignores irrelevant details by using a carefully determined sequence of fixation points to ensure that only a tiny fraction of the optic array is ever processed at the highest resolution. Hinton argues that our brains reason visual information by deconstructing it into a hierarchical representation which we then match to familiar patterns and relationships from memory. The key difference between this understanding and the functionality of CNNs is that recognition of an object should not depend on the angle from which it is viewed. <br />
<br />
To enforce rotational and translational equivariance, Capsule Networks store and preserve hierarchical pose relationships between objects. The core idea behind capsule theory is the explicit numerical representations of relative relationships between different objects within an image. Building these relationships into the Capsule Networks model, the network is able to recognize newly seen objects as a rotated view of a previously seen object. For example, the below image shows the Statue of Liberty under five different angles. If a person had only seen the Statue of Liberty from one angle, they would be able to ascertain that all five pictures below contain the same object (just from a different angle).<br />
<br />
[[File:Rotational Invariance.jpeg |center]]<br />
<br />
Building on this idea of hierarchical representation of spatial relationships between key entities within an image, the authors introduce Capsule Networks. Unlike traditional CNNs, Capsule Networks are better equipped to classify correctly under rotational invariance. Furthermore, the authors managed to achieve state of the art results on MNIST using a fraction of the training samples that alternative state of the art networks require.<br />
<br />
<br />
=Background, Notation, and Definitions=<br />
<br />
==What is a Capsule==<br />
"Each capsule learns to recognize an implicitly defined visual entity over a limited domain of viewing conditions and deformations and it outputs both the probability that the entity is present within its limited domain and a set of “instantiation parameters” that may include the precise pose, lighting and deformation of the visual entity relative to an implicitly defined canonical version of that entity. When the capsule is working properly, the probability of the visual entity being present is locally invariant — it does not change as the entity moves over the manifold of possible appearances within the limited domain covered by the capsule. The instantiation parameters, however, are “equivariant” — as the viewing conditions change and the entity moves over the appearance manifold, the instantiation parameters change by a corresponding amount because they are representing the intrinsic coordinates of the entity on the appearance manifold."<br />
<br />
In essence, capsules store object properties in a vector form; probability of detection is encoded as the vector's length, while spatial properties are encoded as the individual vector components. Thus, when a feature is present but the image captures it under a different angle, the probability of detection remains unchanged.<br />
<br />
A brief overview/understanding of capsules can be found in other papers from the author. To quote from [https://openreview.net/pdf?id=HJWLfGWRb this paper]:<br />
<br />
<blockquote><br />
A capsule network consists of several layers of capsules. The set of capsules in layer L is denoted<br />
as <math>\Omega_L</math>. Each capsule has a 4x4 pose matrix, <math>M</math>, and an activation probability, <math>a</math>. These are like the<br />
activities in a standard neural net: they depend on the current input and are not stored. In between<br />
each capsule i in layer L and each capsule j in layer L + 1 is a 4x4 trainable transformation matrix,<br />
<math>W_{ij}</math> . These <math>W_{ij}</math>'s (and two learned biases per capsule) are the only stored parameters and they<br />
are learned discriminatively. The pose matrix of capsule i is transformed by <math>W_{ij}</math> to cast a vote<br />
<math>V_{ij} = M_iW_{ij}</math> for the pose matrix of capsule j. The poses and activations of all the capsules in layer<br />
L + 1 are calculated by using a non-linear routing procedure which gets as input <math>V_{ij}</math> and <math>a_i</math> for all<br />
<math>i \in \Omega_L, j \in \Omega_{L+1}</math><br />
</blockquote><br />
<math></math><br />
<br />
==Notation==<br />
<br />
We want the length of the output vector of a capsule to represent the probability that the entity represented by the capsule is present in the current input. The paper performs a non-linear squashing operation to ensure that vector length falls between 0 and 1, with shorter vectors (less likely to exist entities) being shrunk towards 0. <br />
<br />
\begin{align} \mathbf{v}_j &= \frac{||\mathbf{s}_j||^2}{1+ ||\mathbf{s}_j||^2} \frac{\mathbf{s}_j}{||\mathbf{s}_j||} \end{align}<br />
<br />
where <math>\mathbf{v}_j</math> is the vector output of capsule <math>j</math> and <math>s_j</math> is its total input.<br />
<br />
For all but the first layer of capsules, the total input to a capsule <math>s_j</math> is a weighted sum over all “prediction vectors” <math>\hat{\mathbf{u}}_{j|i}</math> from the capsules in the layer below and is produced by multiplying the output <math>\mathbf{u}i</math> of a capsule in the layer below by a weight matrix <math>\mathbf{W}ij</math><br />
<br />
\begin{align}<br />
\mathbf{s}_j = \sum_i c_{ij}\hat{\mathbf{u}}_{j|i}, \hat{\mathbf{u}}_{j|i}= \mathbf{W}_{ij}\mathbf{u}_i<br />
\end{align}<br />
where the <math>c_{ij}</math> are coupling coefficients that are determined by the iterative dynamic routing process.<br />
<br />
The coupling coefficients between capsule <math>i</math> and all the capsules in the layer above sum to 1 and are determined by a “routing softmax” whose initial logits <math>b_{ij}</math> are the log prior probabilities that capsule <math>i</math> should be coupled to capsule <math>j</math>.<br />
<br />
\begin{align}<br />
c_{ij} = \frac{\exp(b_{ij})}{\sum_k \exp(b_{ik})}<br />
\end{align}<br />
<br />
=Network Training and Dynamic Routing=<br />
<br />
==Understanding Capsules==<br />
The notation can get somewhat confusing, so I will provide intuition behind the computational steps within a capsule. The following image is taken from naturomic's talk on Capsule Networks.<br />
<br />
[[File:CapsuleNets.jpeg|center|800px]]<br />
<br />
The above image illustrates the key mathematical operations happening within a capsule (and compares them to the structure of a neuron). Although the operations are rather straightforward, it's crucial to note that the capsule computes an affine transformation onto each input vector. The length of the input vectors <math>\mathbf{u}_{i}</math> represent the probability of entity <math>i</math> existing in a lower level. This vector is then reoriented with an affine transform using <math>\mathbf{W}_{ij}</math> matrices that encode spatial relationships between entity <math>\mathbf{u}_{i}</math> and other lower level features.<br />
<br />
We illustrate the intuition behind vector-to-vector matrix multiplication within capsules using the following example: if vectors <math>\mathbf{u}_{1}</math>, <math>\mathbf{u}_{2}</math>, and <math>\mathbf{u}_{3}</math> represent detection of eyes, nose, and mouth respectively, then after multiplication with trained weight matrices <math>\mathbf{W}_{ij}</math> (where j denotes existence of a face), we should get a general idea of the general location of the higher level feature (face), similar to the image below.<br />
<br />
[[File:Predictions.jpeg |center]]<br />
<br />
==Dynamic Routing==<br />
A capsule <math>i</math> in a lower-level layer needs to decide how to send its output vector to higher-level capsules <math>j</math>. This decision is made with probability proportional to <math>c_{ij}</math>. If there are <math>K</math> capsules in the level that capsule <math>i</math> routes to, then we know the following properties about <math>c_{ij}</math>: <math>\sum_{j=1}^M c_{ij} = 1, c_{ij} \geq 0</math><br />
<br />
In essence, the <math>\{c_{ij}\}_{j=1}^M</math> denotes a discrete probability distribution with respect to capsule <math>i</math>'s output location. Lower level capsules decide which higher level capsules to send vectors into by adjusting the corresponding routing weights <math>\{c_{ij}\}_{j=1}^M</math>. After a few iterations in training, numerous vectors will have already been sent to all higher level capsules. Based on the similarity between the current vector being routed and all vectors already sent into the higher level capsules, we decide which capsule to send the current vector into.<br />
[[File:Dynamic Routing.png|center|900px]]<br />
<br />
In the image above, we notice that a cluster of points similar to the current vector has already been routed into capsule K, while most points in capsule J are high dissimilar. It thus makes more sense to route the current observation into capsule K; we adjust the corresponding weight upwards during training.<br />
<br />
These weights are determined through the dynamic routing procedure:<br />
[[File:Routing Algo.png|900px]]<br />
<br />
<br />
Although dynamic routing is not the only manner in which we can encode relationships between capsules, the premise of the paper is to demonstrate the capabilities of capsules under a simple implementation. Since the paper's release in 2017, numerous alternative routing implementations have been released including an EM matrix routing algorithm by the same authors (ICLR 2018).<br />
<br />
=Architecture=<br />
The capsule network architecture given by the authors has 11.36 million trainable parameters. The paper itself is not very detailed on exact implementation of each architectural layer, and hence it leaves some degree of ambiguity on coding various aspects of the original network. The capsule network has 6 overall layers, with the first three layers denoting components of the encoder, and the last 3 denoting components of the decoder.<br />
<br />
==Loss Function==<br />
[[File:Loss Function.png|900px]]<br />
<br />
The cost function looks very complicated, but can be broken down into intuitive components. Before diving into the equation, remember that the length of the vector denotes the probability of object existence. The left side of the equation denotes loss when the network classifies an observation correctly; the term becomes zero when classification is incorrect. To compute loss when the network correctly classifies the label, we subtract the vector norm from a fixed quantity <math>m^+ := 0.9</math>. On the other hand, when the network classifies a label incorrectly, we penalize the loss based on the network's confidence in the incorrect label; we compute the loss by subtracting <math>m^- := 0.1</math> from the vector norm.<br />
<br />
A graphical representation of loss function values under varying vector norms is given below.<br />
[[File:Loss function chart.png|900px]]<br />
<br />
==Encoder Layers==<br />
All experiments within this paper were conducted on the MNIST dataset, and thus the architecture is built to classify the corresponding dataset. For more complex datasets, the experiments were less promising. <br />
<br />
[[File:Architecture.png|center|900px]]<br />
<br />
The encoder layer takes in a 28x28 MNIST image, and learns a 16 dimensional representation of instantiation parameters.<br />
<br />
'''Layer 1: Convolution''': <br />
This layer is a standard convolution layer. Using kernels with size 9x9x1, a stride of 1, and a ReLU activation function, we detect the 2D features within the network.<br />
<br />
'''Layer 2: PrimaryCaps''': <br />
We represent the low level features detected during convolution as 32 primary capsules. Each capsule applies eight convolutional kernels with stride 2 to the output of the convolution layer, and feeds the corresponding transformed tensors into the DigiCaps layer.<br />
<br />
'''Layer 3: DigiCaps''': <br />
This layer contains 10 digit capsules, one for each digit. As explained in the dynamic routing procedure, each input vector from the PrimaryCaps layer has its own corresponding weight matrix <math>W_{ij}</math>. Using the routing coefficients <math>c_{ij}</math> and temporary coefficients <math>b_{ij}</math>, we train the DigiCaps layer to output a ten 16 dimensional vectors. The length of the <math>i^{th}</math> vector in this layer corresponds to the probability of detection of digit <math>i</math>.<br />
<br />
==Decoder Layers==<br />
The decoder layer aims to train the capsules to extract meaningful features for image detection/classification. During training, it takes the 16 layer instantiation vector of the correct (not predicted) DigiCaps layer, and attempts to recreate the 28x28 MNIST image as best as possible. Setting the loss function as reconstruction error (Euclidean distance between reconstructed image and original image), we tune the capsules to encode features that are meaningful within the actual image.<br />
<br />
[[File:Decoder.png|center|900px]]<br />
<br />
The layer consists of three fully connected layers, and transforms a 16x1 vector from the encoder layer into a 28x28 image.<br />
<br />
In addition to the digicaps loss function, we add reconstruction error as a form of regularization. We minimize the Euclidean distance between the outputs of the logistic units and the pixel intensities of the original and reconstructed images. We scale down this reconstruction loss by 0.0005 so that it does not dominate the margin loss during training. As illustrated below, reconstructions from the 16D output of the CapsNet are robust while keeping only important details.<br />
<br />
[[File:Reconstruction.png|center|900px]]<br />
<br />
=MNIST Experimental Results=<br />
<br />
==Accuracy==<br />
The paper tests on the MNIST dataset with 60K training examples, and 10K testing. Wan et al. [2013] achieves 0.21% test error with ensembling and augmenting the data with rotation and scaling. They achieve 0.39% without them. As shown in Table 1, the authors manage to achieve 0.25% test error with only a 3 layer network; the previous state of the art only beat this number with very deep networks. This example shows the importance of routing and reconstruction regularizer, which boosts the performance. On the other hand, while the accuracies are very high, the number of parameters is much smaller compared to the baseline model.<br />
<br />
[[File:Accuracies.png|center|900px]]<br />
<br />
==What Capsules Represent for MNIST==<br />
The following figure shows the digit representation under capsules. Each row shows the reconstruction when one of the 16 dimensions in the DigitCaps representation is tweaked by intervals of 0.05 in the range [−0.25, 0.25]. By tweaking the values, we notice how the reconstruction changes, and thus get a sense for what each dimension is representing. The authors found that some dimensions represent global properties of the digits, while other represent localized properties. <br />
[[File:CapsuleReps.png|center|900px]]<br />
<br />
One example the authors provide is: different dimensions are used for the length of the ascender of a 6 and the size of the loop. The variations include stroke thickness, skew and width, as well as digit-specific variations. The authors are able to show dimension representations using a decoder network by feeding a perturbed vector.<br />
<br />
==Robustness of CapsNet==<br />
The authors conclude that DigitCaps capsules learn more robust representations for each digit class than traditional CNNs. The trained CapsNet becomes moderately robust to small affine transformations in the test data.<br />
<br />
To compare the robustness of CapsNet to affine transformations against traditional CNNs, both models (CapsNet and a traditional CNN with MaxPooling and DropOut) were trained on a padded and translated MNIST training set, in which each example is an MNIST digit placed randomly on a black background of 40 × 40 pixels. The networks were then tested on the [http://www.cs.toronto.edu/~tijmen/affNIST/ affNIST] dataset (MNIST digits with random affine transformation). An under-trained CapsNet which achieved 99.23% accuracy on the MNIST test set achieved a corresponding 79% accuracy on the affnist test set. A traditional CNN achieved similar accuracy (99.22%) on the mnist test set, but only 66% on the affnist test set.<br />
<br />
=MultiMNIST & Other Experiments=<br />
<br />
==MultiMNIST==<br />
To evaluate the performance of the model on highly overlapping digits, the authors generate a 'MultiMNIST' dataset. In MultiMNIST, images are two overlaid MNIST digits of the same set(train or test) but different classes. The results indicate a classification error rate of 5%. Additionally, CapsNet can be used to segment the image into the two digits that compose it. Moreover, the model is able to deal with the overlaps and reconstruct digits correctly since each digit capsule can learn the style from the votes of PrimaryCapsules layer (Figure 5).<br />
<br />
There are some additional steps to generating the MultiMNIST dataset.<br />
<br />
1. Both images are shifted by up to 4 pixels in each direction resulting in a 36 × 36 image. Bounding boxes of digits in MNIST overlap by approximately 80%, so this is used to make both digits identifiable (since there is no RGB difference learnable by the network to separate the digits)<br />
<br />
2. The label becomes a vector of two numbers, representing the original digit and the randomly generated (and overlaid) digit.<br />
<br />
<br />
<br />
[[File:CapsuleNets MultiMNIST.PNG|600px|thumb|center|Figure 5: Sample reconstructions of a CapsNet with 3 routing iterations on MultiMNIST test dataset.<br />
The two reconstructed digits are overlayed in green and red as the lower image. The upper image<br />
shows the input image. L:(l1; l2) represents the label for the two digits in the image and R:(r1; r2)<br />
represents the two digits used for reconstruction. The two right most columns show two examples<br />
with wrong classification reconstructed from the label and from the prediction (P). In the (2; 8)<br />
example the model confuses 8 with a 7 and in (4; 9) it confuses 9 with 0. The other columns have<br />
correct classifications and show that the model accounts for all the pixels while being able to assign<br />
one pixel to two digits in extremely difficult scenarios (column 1 − 4). Note that in dataset generation<br />
the pixel values are clipped at 1. The two columns with the (*) mark show reconstructions from a<br />
digit that is neither the label nor the prediction. These columns suggests that the model is not just<br />
finding the best fit for all the digits in the image including the ones that do not exist. Therefore in case<br />
of (5; 0) it cannot reconstruct a 7 because it knows that there is a 5 and 0 that fit best and account for<br />
all the pixels. Also, in case of (8; 1) the loop of 8 has not triggered 0 because it is already accounted<br />
for by 8. Therefore it will not assign one pixel to two digits if one of them does not have any other<br />
support.]]<br />
<br />
==Other datasets==<br />
The authors also tested the proposed capsule model on CIFAR10 dataset and achieved an error rate of 10.6%. The model tested was an ensemble of 7 models. Each of the model in the ensemble had the same architecture as the model used for MNIST (apart from 3 additional channels and 64 different types of primary capsules being used). These 7 models were trained on 24x24 patches of the training images for 3 iterations. During experimentation, the authors also found out that adding an additional none-of-the-above category helped improved the overall performance. The error rate achieved is comparable to the error rate achieved by a standard CNN model. According to the authors, one of the reason for low performance is the fact that background in CIFAR-10 images are too varied for it to be adequately modeled by reasonably sized capsule net.<br />
<br />
The proposed model was also evaluated using a small subset of SVHN dataset. The network trained was much smaller and trained using only 73257 training images. The network still managed to achieve an error rate of 4.3% on the test set.<br />
<br />
=Critique=<br />
Although the network performs incredibly favourably in the author's experiments, it has a long way to go on more complex datasets. On CIFAR 10, the network achieved subpar results, and the experimental results seem to be worse when the problem becomes more complex. This is anticipated, since these networks are still in their early stage; later innovations might come in the upcoming decades/years.<br />
<br />
Hinton talks about CapsuleNets revolutionizing areas such as self-driving, but such groundbreaking innovations are far away from CIFAR10, and even further from MNIST. Only time can tell if CapsNets will live up to their hype.<br />
<br />
Capsules inherently segment images, and learn a lower dimensional embedding in a new manner, which makes them likely to perform well on segmentation and computer vision tasks once further research is done. <br />
<br />
Additionally these networks are more interpretable than CNNs, and have strong theoretical reasoning for why they could work. Naturally, it would be hard for a new architecture to beat the heavily researched/modified CNNs.<br />
<br />
* ([https://openreview.net/forum?id=HJWLfGWRb]) it's not fully clear how effective it can be performed / how scalable it is. Evaluation is performed on a small dataset for shape recognition. The approach will need to be tested on larger, more challenging datasets.<br />
<br />
=Future Work=<br />
The same authors [N. F. Geoffrey E Hinton, Sara Sabour] presented another paper "MATRIX CAPSULES WITH EM ROUTING" in ICLR 2018, which achieved better results than the work presented in this paper. They presented a novel capsule type, where each capsule has a logistic unit and a 4x4 pose matrix. This new type reduced number of errors by 45%, and performed better than standard CNN on white box adversarial attacks. <br />
Moreover, we may try to change the curvature and sensitivities to various factors by introducing new form of loss function. It may improve the performance of the model for more complicated data set which is one of the model's drawback.<br />
<br />
=References=<br />
#N. F. Geoffrey E Hinton, Sara Sabour. Matrix capsules with em routing. In International Conference on Learning Representations, 2018.<br />
#S. Sabour, N. Frosst, and G. E. Hinton, “Dynamic routing between capsules,” arXiv preprint arXiv:1710.09829v2, 2017<br />
# Hinton, G. E., Krizhevsky, A. and Wang, S. D. (2011), Transforming Auto-encoders <br />
#Geoffrey Hinton's talk: What is wrong with convolutional neural nets? - Talk given at MIT. Brain & Cognitive Sciences - Fall Colloquium Series. [https://www.youtube.com/watch?v=rTawFwUvnLE ]<br />
#Understanding Hinton’s Capsule Networks - Max Pechyonkin's series [https://medium.com/ai%C2%B3-theory-practice-business/understanding-hintons-capsule-networks-part-i-intuition-b4b559d1159b]</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=CapsuleNets&diff=41791CapsuleNets2018-11-29T04:14:31Z<p>Ka2khan: /* MultiMNIST & Other Experiments */</p>
<hr />
<div>The paper "Dynamic Routing Between Capsules" was written by three researchers at Google Brain: Sara Sabour, Nicholas Frosst, and Geoffrey E. Hinton. This paper was published and presented at the 31st Conference on Neural Information Processing Systems (NIPS 2017) in Long Beach, California. The same three researchers recently published a highly related paper "Matrix Capsules with EM Routing" for ICLR 2018.<br />
<br />
=Motivation=<br />
<br />
Ever since AlexNet eclipsed the performance of competing architectures in the 2012 ImageNet challenge, convolutional neural networks have maintained their dominance in computer vision applications. Despite the recent successes and innovations brought about by convolutional neural networks, some assumptions made in these networks are perhaps unwarranted and deficient. Using a novel neural network architecture, the authors create CapsuleNets, a network that they claim is able to learn image representations in a more robust, human-like manner. With only a 3 layer capsule network, they achieved near state-of-the-art results on MNIST.<br />
==Adversarial Examples==<br />
<br />
First discussed by Christian Szegedy et. al. in late 2013, adversarial examples have been heavily discussed by the deep learning community as a potential security threat to AI learning. Adversarial examples are defined as inputs that an attacker creates intentionally fool a machine learning model. An example of an adversarial example is shown below: <br />
<br />
[[File:adversarial_img_1.png |center]]<br />
To the human eye, the image appears to be a panda both before and after noise is injected into the image, whereas the trained ConvNet model discerns the noisy image as a Gibbon with almost 100% certainty. The fact that the network is unable to classify the above image as a panda after the epsilon perturbation leads to many potential security risks in AI dependent systems such as self-driving vehicles. Although various methods have been suggested to combat adversarial examples, robust defences are hard to construct due to the inherent difficulties in constructing theoretical models for the adversarial example crafting process. However, beyond the fact that these examples may serve as a security threat, it emphasizes that these convolutional neural networks do not learn image classification/object detection patterns the same way that a human would. Rather than identifying the core features of a panda such as: its eyes, mouth, nose, and the gradient changes in its black/white fur, the convolutional neural network seems to be learning image representations in a completely different manner. Deep learning researchers often attempt to model neural networks after human learning, and it is clear that further steps must be taken to robustify ConvNets against targeted noise perturbations.<br />
<br />
==Drawbacks of CNNs==<br />
Hinton claims that the key fault with traditional CNNs lies within the pooling function. Although pooling builds translational invariance into the network, it fails to preserve spatial relationships between objects. When we pool, we effectively reduce a kxk kernel of convolved cells into a scalar input. This results in a desired local invariance without inhibiting the network's ability to detect features, but causes valuable spatial information to be lost.<br />
<br />
In the example below, the network is able to detect the similar features (eyes, mouth, nose, etc) within both images, but fails to recognize that one image is a human face, while the other is a Picasso-esque due to the CNN's inability to encode spatial relationships after multiple pooling layers.<br />
<br />
<br />
[[File:Equivariance Face.png |center]]<br />
<br />
Conversely, we hope that a CNN can recognize that both of the following pictures contain a kitten. Unfortunately, when we feed the two images into a ResNet50 architecture, only the first image is correctly classified, while the second image is predicted to be a guinea pig.<br />
<br />
<br />
[[File:kitten.jpeg |center]]<br />
<br />
<br />
[[File:kitten-rotated-180.jpg |center]]<br />
<br />
For a more in depth discussion on the problems with ConvNets, please listen to Geoffrey Hinton's talk "What is wrong with convolutional neural nets?" given at MIT during the Brain & Cognitive Sciences - Fall Colloquium Series (December 4, 2014).<br />
<br />
==Intuition for Capsules==<br />
Human vision ignores irrelevant details by using a carefully determined sequence of fixation points to ensure that only a tiny fraction of the optic array is ever processed at the highest resolution. Hinton argues that our brains reason visual information by deconstructing it into a hierarchical representation which we then match to familiar patterns and relationships from memory. The key difference between this understanding and the functionality of CNNs is that recognition of an object should not depend on the angle from which it is viewed. <br />
<br />
To enforce rotational and translational equivariance, Capsule Networks store and preserve hierarchical pose relationships between objects. The core idea behind capsule theory is the explicit numerical representations of relative relationships between different objects within an image. Building these relationships into the Capsule Networks model, the network is able to recognize newly seen objects as a rotated view of a previously seen object. For example, the below image shows the Statue of Liberty under five different angles. If a person had only seen the Statue of Liberty from one angle, they would be able to ascertain that all five pictures below contain the same object (just from a different angle).<br />
<br />
[[File:Rotational Invariance.jpeg |center]]<br />
<br />
Building on this idea of hierarchical representation of spatial relationships between key entities within an image, the authors introduce Capsule Networks. Unlike traditional CNNs, Capsule Networks are better equipped to classify correctly under rotational invariance. Furthermore, the authors managed to achieve state of the art results on MNIST using a fraction of the training samples that alternative state of the art networks require.<br />
<br />
<br />
=Background, Notation, and Definitions=<br />
<br />
==What is a Capsule==<br />
"Each capsule learns to recognize an implicitly defined visual entity over a limited domain of viewing conditions and deformations and it outputs both the probability that the entity is present within its limited domain and a set of “instantiation parameters” that may include the precise pose, lighting and deformation of the visual entity relative to an implicitly defined canonical version of that entity. When the capsule is working properly, the probability of the visual entity being present is locally invariant — it does not change as the entity moves over the manifold of possible appearances within the limited domain covered by the capsule. The instantiation parameters, however, are “equivariant” — as the viewing conditions change and the entity moves over the appearance manifold, the instantiation parameters change by a corresponding amount because they are representing the intrinsic coordinates of the entity on the appearance manifold."<br />
<br />
In essence, capsules store object properties in a vector form; probability of detection is encoded as the vector's length, while spatial properties are encoded as the individual vector components. Thus, when a feature is present but the image captures it under a different angle, the probability of detection remains unchanged.<br />
<br />
A brief overview/understanding of capsules can be found in other papers from the author. To quote from [https://openreview.net/pdf?id=HJWLfGWRb this paper]:<br />
<br />
<blockquote><br />
A capsule network consists of several layers of capsules. The set of capsules in layer L is denoted<br />
as <math>\Omega_L</math>. Each capsule has a 4x4 pose matrix, <math>M</math>, and an activation probability, <math>a</math>. These are like the<br />
activities in a standard neural net: they depend on the current input and are not stored. In between<br />
each capsule i in layer L and each capsule j in layer L + 1 is a 4x4 trainable transformation matrix,<br />
<math>W_{ij}</math> . These <math>W_{ij}</math>'s (and two learned biases per capsule) are the only stored parameters and they<br />
are learned discriminatively. The pose matrix of capsule i is transformed by <math>W_{ij}</math> to cast a vote<br />
<math>V_{ij} = M_iW_{ij}</math> for the pose matrix of capsule j. The poses and activations of all the capsules in layer<br />
L + 1 are calculated by using a non-linear routing procedure which gets as input <math>V_{ij}</math> and <math>a_i</math> for all<br />
<math>i \in \Omega_L, j \in \Omega_{L+1}</math><br />
</blockquote><br />
<math></math><br />
<br />
==Notation==<br />
<br />
We want the length of the output vector of a capsule to represent the probability that the entity represented by the capsule is present in the current input. The paper performs a non-linear squashing operation to ensure that vector length falls between 0 and 1, with shorter vectors (less likely to exist entities) being shrunk towards 0. <br />
<br />
\begin{align} \mathbf{v}_j &= \frac{||\mathbf{s}_j||^2}{1+ ||\mathbf{s}_j||^2} \frac{\mathbf{s}_j}{||\mathbf{s}_j||} \end{align}<br />
<br />
where <math>\mathbf{v}_j</math> is the vector output of capsule <math>j</math> and <math>s_j</math> is its total input.<br />
<br />
For all but the first layer of capsules, the total input to a capsule <math>s_j</math> is a weighted sum over all “prediction vectors” <math>\hat{\mathbf{u}}_{j|i}</math> from the capsules in the layer below and is produced by multiplying the output <math>\mathbf{u}i</math> of a capsule in the layer below by a weight matrix <math>\mathbf{W}ij</math><br />
<br />
\begin{align}<br />
\mathbf{s}_j = \sum_i c_{ij}\hat{\mathbf{u}}_{j|i}, \hat{\mathbf{u}}_{j|i}= \mathbf{W}_{ij}\mathbf{u}_i<br />
\end{align}<br />
where the <math>c_{ij}</math> are coupling coefficients that are determined by the iterative dynamic routing process.<br />
<br />
The coupling coefficients between capsule <math>i</math> and all the capsules in the layer above sum to 1 and are determined by a “routing softmax” whose initial logits <math>b_{ij}</math> are the log prior probabilities that capsule <math>i</math> should be coupled to capsule <math>j</math>.<br />
<br />
\begin{align}<br />
c_{ij} = \frac{\exp(b_{ij})}{\sum_k \exp(b_{ik})}<br />
\end{align}<br />
<br />
=Network Training and Dynamic Routing=<br />
<br />
==Understanding Capsules==<br />
The notation can get somewhat confusing, so I will provide intuition behind the computational steps within a capsule. The following image is taken from naturomic's talk on Capsule Networks.<br />
<br />
[[File:CapsuleNets.jpeg|center|800px]]<br />
<br />
The above image illustrates the key mathematical operations happening within a capsule (and compares them to the structure of a neuron). Although the operations are rather straightforward, it's crucial to note that the capsule computes an affine transformation onto each input vector. The length of the input vectors <math>\mathbf{u}_{i}</math> represent the probability of entity <math>i</math> existing in a lower level. This vector is then reoriented with an affine transform using <math>\mathbf{W}_{ij}</math> matrices that encode spatial relationships between entity <math>\mathbf{u}_{i}</math> and other lower level features.<br />
<br />
We illustrate the intuition behind vector-to-vector matrix multiplication within capsules using the following example: if vectors <math>\mathbf{u}_{1}</math>, <math>\mathbf{u}_{2}</math>, and <math>\mathbf{u}_{3}</math> represent detection of eyes, nose, and mouth respectively, then after multiplication with trained weight matrices <math>\mathbf{W}_{ij}</math> (where j denotes existence of a face), we should get a general idea of the general location of the higher level feature (face), similar to the image below.<br />
<br />
[[File:Predictions.jpeg |center]]<br />
<br />
==Dynamic Routing==<br />
A capsule <math>i</math> in a lower-level layer needs to decide how to send its output vector to higher-level capsules <math>j</math>. This decision is made with probability proportional to <math>c_{ij}</math>. If there are <math>K</math> capsules in the level that capsule <math>i</math> routes to, then we know the following properties about <math>c_{ij}</math>: <math>\sum_{j=1}^M c_{ij} = 1, c_{ij} \geq 0</math><br />
<br />
In essence, the <math>\{c_{ij}\}_{j=1}^M</math> denotes a discrete probability distribution with respect to capsule <math>i</math>'s output location. Lower level capsules decide which higher level capsules to send vectors into by adjusting the corresponding routing weights <math>\{c_{ij}\}_{j=1}^M</math>. After a few iterations in training, numerous vectors will have already been sent to all higher level capsules. Based on the similarity between the current vector being routed and all vectors already sent into the higher level capsules, we decide which capsule to send the current vector into.<br />
[[File:Dynamic Routing.png|center|900px]]<br />
<br />
In the image above, we notice that a cluster of points similar to the current vector has already been routed into capsule K, while most points in capsule J are high dissimilar. It thus makes more sense to route the current observation into capsule K; we adjust the corresponding weight upwards during training.<br />
<br />
These weights are determined through the dynamic routing procedure:<br />
[[File:Routing Algo.png|900px]]<br />
<br />
<br />
Although dynamic routing is not the only manner in which we can encode relationships between capsules, the premise of the paper is to demonstrate the capabilities of capsules under a simple implementation. Since the paper's release in 2017, numerous alternative routing implementations have been released including an EM matrix routing algorithm by the same authors (ICLR 208).<br />
<br />
=Architecture=<br />
The capsule network architecture given by the authors has 11.36 million trainable parameters. The paper itself is not very detailed on exact implementation of each architectural layer, and hence it leaves some degree of ambiguity on coding various aspects of the original network. The capsule network has 6 overall layers, with the first three layers denoting components of the encoder, and the last 3 denoting components of the decoder.<br />
<br />
==Loss Function==<br />
[[File:Loss Function.png|900px]]<br />
<br />
The cost function looks very complicated, but can be broken down into intuitive components. Before diving into the equation, remember that the length of the vector denotes the probability of object existence. The left side of the equation denotes loss when the network classifies an observation correctly; the term becomes zero when classification is incorrect. To compute loss when the network correctly classifies the label, we subtract the vector norm from a fixed quantity <math>m^+ := 0.9</math>. On the other hand, when the network classifies a label incorrectly, we penalize the loss based on the network's confidence in the incorrect label; we compute the loss by subtracting <math>m^- := 0.1</math> from the vector norm.<br />
<br />
A graphical representation of loss function values under varying vector norms is given below.<br />
[[File:Loss function chart.png|900px]]<br />
<br />
==Encoder Layers==<br />
All experiments within this paper were conducted on the MNIST dataset, and thus the architecture is built to classify the corresponding dataset. For more complex datasets, the experiments were less promising. <br />
<br />
[[File:Architecture.png|center|900px]]<br />
<br />
The encoder layer takes in a 28x28 MNIST image, and learns a 16 dimensional representation of instantiation parameters.<br />
<br />
'''Layer 1: Convolution''': <br />
This layer is a standard convolution layer. Using kernels with size 9x9x1, a stride of 1, and a ReLU activation function, we detect the 2D features within the network.<br />
<br />
'''Layer 2: PrimaryCaps''': <br />
We represent the low level features detected during convolution as 32 primary capsules. Each capsule applies eight convolutional kernels with stride 2 to the output of the convolution layer, and feeds the corresponding transformed tensors into the DigiCaps layer.<br />
<br />
'''Layer 3: DigiCaps''': <br />
This layer contains 10 digit capsules, one for each digit. As explained in the dynamic routing procedure, each input vector from the PrimaryCaps layer has its own corresponding weight matrix <math>W_{ij}</math>. Using the routing coefficients <math>c_{ij}</math> and temporary coefficients <math>b_{ij}</math>, we train the DigiCaps layer to output a ten 16 dimensional vectors. The length of the <math>i^{th}</math> vector in this layer corresponds to the probability of detection of digit <math>i</math>.<br />
<br />
==Decoder Layers==<br />
The decoder layer aims to train the capsules to extract meaningful features for image detection/classification. During training, it takes the 16 layer instantiation vector of the correct (not predicted) DigiCaps layer, and attempts to recreate the 28x28 MNIST image as best as possible. Setting the loss function as reconstruction error (Euclidean distance between reconstructed image and original image), we tune the capsules to encode features that are meaningful within the actual image.<br />
<br />
[[File:Decoder.png|center|900px]]<br />
<br />
The layer consists of three fully connected layers, and transforms a 16x1 vector from the encoder layer into a 28x28 image.<br />
<br />
In addition to the digicaps loss function, we add reconstruction error as a form of regularization. We minimize the Euclidean distance between the outputs of the logistic units and the pixel intensities of the original and reconstructed images. We scale down this reconstruction loss by 0.0005 so that it does not dominate the margin loss during training. As illustrated below, reconstructions from the 16D output of the CapsNet are robust while keeping only important details.<br />
<br />
[[File:Reconstruction.png|center|900px]]<br />
<br />
=MNIST Experimental Results=<br />
<br />
==Accuracy==<br />
The paper tests on the MNIST dataset with 60K training examples, and 10K testing. Wan et al. [2013] achieves 0.21% test error with ensembling and augmenting the data with rotation and scaling. They achieve 0.39% without them. As shown in Table 1, the authors manage to achieve 0.25% test error with only a 3 layer network; the previous state of the art only beat this number with very deep networks. This example shows the importance of routing and reconstruction regularizer, which boosts the performance. On the other hand, while the accuracies are very high, the number of parameters is much smaller compared to the baseline model.<br />
<br />
[[File:Accuracies.png|center|900px]]<br />
<br />
==What Capsules Represent for MNIST==<br />
The following figure shows the digit representation under capsules. Each row shows the reconstruction when one of the 16 dimensions in the DigitCaps representation is tweaked by intervals of 0.05 in the range [−0.25, 0.25]. By tweaking the values, we notice how the reconstruction changes, and thus get a sense for what each dimension is representing. The authors found that some dimensions represent global properties of the digits, while other represent localized properties. <br />
[[File:CapsuleReps.png|center|900px]]<br />
<br />
One example the authors provide is: different dimensions are used for the length of the ascender of a 6 and the size of the loop. The variations include stroke thickness, skew and width, as well as digit-specific variations. The authors are able to show dimension representations using a decoder network by feeding a perturbed vector.<br />
<br />
==Robustness of CapsNet==<br />
The authors conclude that DigitCaps capsules learn more robust representations for each digit class than traditional CNNs. The trained CapsNet becomes moderately robust to small affine transformations in the test data.<br />
<br />
To compare the robustness of CapsNet to affine transformations against traditional CNNs, both models (CapsNet and a traditional CNN with MaxPooling and DropOut) were trained on a padded and translated MNIST training set, in which each example is an MNIST digit placed randomly on a black background of 40 × 40 pixels. The networks were then tested on the [http://www.cs.toronto.edu/~tijmen/affNIST/ affNIST] dataset (MNIST digits with random affine transformation). An under-trained CapsNet which achieved 99.23% accuracy on the MNIST test set achieved a corresponding 79% accuracy on the affnist test set. A traditional CNN achieved similar accuracy (99.22%) on the mnist test set, but only 66% on the affnist test set.<br />
<br />
=MultiMNIST & Other Experiments=<br />
<br />
==MultiMNIST==<br />
To evaluate the performance of the model on highly overlapping digits, the authors generate a 'MultiMNIST' dataset. In MultiMNIST, images are two overlaid MNIST digits of the same set(train or test) but different classes. The results indicate a classification error rate of 5%. Additionally, CapsNet can be used to segment the image into the two digits that compose it. Moreover, the model is able to deal with the overlaps and reconstruct digits correctly since each digit capsule can learn the style from the votes of PrimaryCapsules layer (Figure 5).<br />
<br />
There are some additional steps to generating the MultiMNIST dataset.<br />
<br />
1. Both images are shifted by up to 4 pixels in each direction resulting in a 36 × 36 image. Bounding boxes of digits in MNIST overlap by approximately 80%, so this is used to make both digits identifiable (since there is no RGB difference learnable by the network to separate the digits)<br />
<br />
2. The label becomes a vector of two numbers, representing the original digit and the randomly generated (and overlaid) digit.<br />
<br />
<br />
<br />
[[File:CapsuleNets MultiMNIST.PNG|600px|thumb|center|Figure 5: Sample reconstructions of a CapsNet with 3 routing iterations on MultiMNIST test dataset.<br />
The two reconstructed digits are overlayed in green and red as the lower image. The upper image<br />
shows the input image. L:(l1; l2) represents the label for the two digits in the image and R:(r1; r2)<br />
represents the two digits used for reconstruction. The two right most columns show two examples<br />
with wrong classification reconstructed from the label and from the prediction (P). In the (2; 8)<br />
example the model confuses 8 with a 7 and in (4; 9) it confuses 9 with 0. The other columns have<br />
correct classifications and show that the model accounts for all the pixels while being able to assign<br />
one pixel to two digits in extremely difficult scenarios (column 1 − 4). Note that in dataset generation<br />
the pixel values are clipped at 1. The two columns with the (*) mark show reconstructions from a<br />
digit that is neither the label nor the prediction. These columns suggests that the model is not just<br />
finding the best fit for all the digits in the image including the ones that do not exist. Therefore in case<br />
of (5; 0) it cannot reconstruct a 7 because it knows that there is a 5 and 0 that fit best and account for<br />
all the pixels. Also, in case of (8; 1) the loop of 8 has not triggered 0 because it is already accounted<br />
for by 8. Therefore it will not assign one pixel to two digits if one of them does not have any other<br />
support.]]<br />
<br />
==Other datasets==<br />
The authors also tested the proposed capsule model on CIFAR10 dataset and achieved an error rate of 10.6%. The model tested was an ensemble of 7 models. Each of the model in the ensemble had the same architecture as the model used for MNIST (apart from 3 additional channels and 64 different types of primary capsules being used). These 7 models were trained on 24x24 patches of the training images for 3 iterations. During experimentation, the authors also found out that adding an additional none-of-the-above category helped improved the overall performance. The error rate achieved is comparable to the error rate achieved by a standard CNN model. According to the authors, one of the reason for low performance is the fact that background in CIFAR-10 images are too varied for it to be adequately modeled by reasonably sized capsule net.<br />
<br />
The proposed model was also evaluated using a small subset of SVHN dataset. The network trained was much smaller and trained using only 73257 training images. The network still managed to achieve an error rate of 4.3% on the test set.<br />
<br />
=Critique=<br />
Although the network performs incredibly favourably in the author's experiments, it has a long way to go on more complex datasets. On CIFAR 10, the network achieved subpar results, and the experimental results seem to be worse when the problem becomes more complex. This is anticipated, since these networks are still in their early stage; later innovations might come in the upcoming decades/years.<br />
<br />
Hinton talks about CapsuleNets revolutionizing areas such as self-driving, but such groundbreaking innovations are far away from CIFAR10, and even further from MNIST. Only time can tell if CapsNets will live up to their hype.<br />
<br />
Capsules inherently segment images, and learn a lower dimensional embedding in a new manner, which makes them likely to perform well on segmentation and computer vision tasks once further research is done. <br />
<br />
Additionally these networks are more interpretable than CNNs, and have strong theoretical reasoning for why they could work. Naturally, it would be hard for a new architecture to beat the heavily researched/modified CNNs.<br />
<br />
* ([https://openreview.net/forum?id=HJWLfGWRb]) it's not fully clear how effective it can be performed / how scalable it is. Evaluation is performed on a small dataset for shape recognition. The approach will need to be tested on larger, more challenging datasets.<br />
<br />
=Future Work=<br />
The same authors [N. F. Geoffrey E Hinton, Sara Sabour] presented another paper "MATRIX CAPSULES WITH EM ROUTING" in ICLR 2018, which achieved better results than the work presented in this paper. They presented a novel capsule type, where each capsule has a logistic unit and a 4x4 pose matrix. This new type reduced number of errors by 45%, and performed better than standard CNN on white box adversarial attacks. <br />
Moreover, we may try to change the curvature and sensitivities to various factors by introducing new form of loss function. It may improve the performance of the model for more complicated data set which is one of the model's drawback.<br />
<br />
=References=<br />
#N. F. Geoffrey E Hinton, Sara Sabour. Matrix capsules with em routing. In International Conference on Learning Representations, 2018.<br />
#S. Sabour, N. Frosst, and G. E. Hinton, “Dynamic routing between capsules,” arXiv preprint arXiv:1710.09829v2, 2017<br />
# Hinton, G. E., Krizhevsky, A. and Wang, S. D. (2011), Transforming Auto-encoders <br />
#Geoffrey Hinton's talk: What is wrong with convolutional neural nets? - Talk given at MIT. Brain & Cognitive Sciences - Fall Colloquium Series. [https://www.youtube.com/watch?v=rTawFwUvnLE ]<br />
#Understanding Hinton’s Capsule Networks - Max Pechyonkin's series [https://medium.com/ai%C2%B3-theory-practice-business/understanding-hintons-capsule-networks-part-i-intuition-b4b559d1159b]</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=ShakeDrop_Regularization&diff=41776ShakeDrop Regularization2018-11-29T00:33:37Z<p>Ka2khan: /* Experiments */</p>
<hr />
<div>=Introduction=<br />
Current state of the art techniques for object classification are deep neural networks based on the residual block, first published by (He et al., 2016). This technique has been the foundation of several improved networks, including Wide ResNet (Zagoruyko & Komodakis, 2016), PyramdNet (Han et al., 2017) and ResNeXt (Xie et al., 2017). They have been further improved by regularization, such as Stochastic Depth (ResDrop) (Huang et al., 2016) and Shake-Shake (Gastaldi, 2017). Shake-Shake applied to ResNeXt has achieved one of the lowest error rates on the CIFAR-10 and CIFAR-100 datasets. However, it is only applicable to multi-branch architectures and is not memory efficient since it requires two branches of residual blocks to apply. This paper seeks to formulate a general expansion of Shake-Shake that can be applied to any residual block based network.<br />
<br />
=Existing Methods=<br />
<br />
'''Deep Approaches'''<br />
<br />
'''ResNet''', was the first use of residual blocks, a foundational feature in many modern state of the art convolution neural networks. They can be formulated as <math>G(x) = x + F(x)</math> where <math>x</math> and <math>G(x)</math> are the input and output of the residual block, and <math>F(x)</math> is the output of the residual branch on the residual block. A residual block typically performs a convolution operation and then passes the result plus its input onto the next block.<br />
<br />
[[File:ResidualBlock.png|600px|centre|thumb|An example of a simple residual block from Deep Residual Learning for Image Recognition by He et al., 2016]]<br />
<br />
ResNet is constructed out of a large number of these residual blocks sequentially stacked. It is interesting to note that having too many layers can cause overfitting, as pointed out by He et al. (2016) with the high error rates for the 1,202-layer ResNet on CIFAR datasets. Another paper (Veit et al., 2016) empirically showed that the cause of the high error rates can be mostly attributed to specific residual blocks whose channels increase greatly.<br />
<br />
'''PyramidNet''' is an important iteration that built on ResNet and WideResNet by gradually increasing channels on each residual block. The residual block is similar to those used in ResNet. It has been used to generate some of the first successful convolution neural networks with very large depth, at 272 layers. Amongst unmodified residual network architectures, it performs the best on the CIFAR datasets.<br />
<br />
[[File:ResidualBlockComparison.png|900px|centre|thumb|A simple illustration of different residual blocks from Deep Pyramidal Residual Networks by Han et al., 2017]]<br />
<br />
<br />
'''Non-Deep Approaches'''<br />
<br />
'''Wide ResNet''' modified ResNet by increasing channels in each layer, having a wider and shallower structure. Similarly to PyramidNet, this architecture avoids some of the pitfalls in the original formulation of ResNet.<br />
<br />
'''ResNeXt''' achieved performance beyond that of Wide ResNet with only a small increase in the number of parameters. It can be formulated as <math>G(x) = x + F_1(x)+F_2(x)</math>. In this case, <math>F_1(x)</math> and <math>F_2(x)</math> are the outputs of two paired convolution operations in a single residual block. The number of branches is not limited to 2, and will control the result of this network.<br />
<br />
<br />
[[File:SimplifiedResNeXt.png|600px|centre|thumb|Simplified ResNeXt Convolution Block. Yamada et al., 2018]]<br />
<br />
<br />
'''Regularization Methods'''<br />
<br />
'''Stochastic Depth''' helped address the issue of vanishing gradients in ResNet. It works by randomly dropping residual blocks. On the <math>l^{th}</math> residual block the Stochastic Depth process is given as <math>G(x)=x+b_lF(x)</math> where <math>b_l \in \{0,1\}</math> is a Bernoulli random variable with probability <math>p_l</math>. Using a constant value for <math>p_l</math> didn't work well, so instead a linear decay rule <math>p_l = 1 - \frac{l}{L}(1-p_L)</math> was used. In this equation, <math>L</math> is the number of layers, and <math>p_L</math> is the initial parameter. <br />
<br />
'''Shake-Shake''' is a regularization method that specifically improves the ResNeXt architecture. It can be given as <math>G(x)=x+\alpha F_1(x)+(1-\alpha)F_2(x)</math>, where <math>\alpha \in [0,1]</math> is a random coefficient. <math>\alpha</math> is used during the forward pass, and another identically distributed random parameter <math>\beta</math> is used in the backward pass. This caused one of the two paired convolution operations to be dropped, and further improved ResNeXt.<br />
<br />
[[File:Paper 32.jpg|600px|centre|thumb| Shake-Shake (ResNeXt + Shake-Shake) (Gastaldi, 2017), in which some processing layers omitted for conciseness.]]<br />
<br />
=Proposed Method=<br />
This paper seeks to generalize the method proposed in Shake-Shake to be applied to any residual structure network. Shake-Shake. The initial formulation of 1-branch shake is <math>G(x) = x + \alpha F(x)</math>. In this case, <math>\alpha</math> is a coefficient that disturbs the forward pass, but is not necessarily constrained to be [0,1]. Another corresponding coefficient <math>\beta</math> is used in the backwards pass. Applying this simple adaptation of Shake-Shake on a 110-layer version of PyramidNet with <math>\alpha \in [0,1]</math> and <math>\beta \in [0,1]</math> performs abysmally, with an error rate of 77.99%.<br />
<br />
This failure is a result of the setup causing too much perturbation. A trick is needed to promote learning with large perturbations, to preserve the regularization effect. The idea of the authors is to borrow from ResDrop and combine that with Shake-Shake. This works by randomly deciding whether to apply 1-branch shake. This creates in effect two networks, the original network without a regularization component, and a regularized network. When the non regularized network is selected, learning is promoted, when the perturbed network is selected, learning is disturbed. Achieving good performance requires a balance between the two. <br />
<br />
'''ShakeDrop''' is given as <br />
<br />
<div align="center"><br />
<math>G(x) = x + (b_l + \alpha - b_l \alpha)F(x)</math>,<br />
</div><br />
<br />
where <math>b_l</math> is a Bernoulli random variable following the linear decay rule used in Stochastic Depth. An alternative presentation is <br />
<br />
<div align="center"><br />
<math><br />
G(x) = \begin{cases}<br />
x + F(x) ~~ \text{if } b_l = 1 \\<br />
x + \alpha F(x) ~~ \text{otherwise}<br />
\end{cases}<br />
</math><br />
</div><br />
<br />
If <math>b_l = 1</math> then ShakeDrop is equivalent to the original network, otherwise it is the network + 1-branch Shake. Regardless of the value of <math>\beta</math> on the backwards pass, network weights will be updated.<br />
<br />
=Experiments=<br />
<br />
'''Parameter Search'''<br />
<br />
The authors experiments began with a hyperparameter search utilizing ShakeDrop on pyramidal networks. The PyramidNet used was made up of a total of 110 layers which included a convolutional layer and a final fully connected layer. It had 54 additive pyramidal residual blocks and the final residual block had 286 channels. The results of this search are presented below. <br />
<br />
[[File:ShakeDropHyperParameterSearch.png|600px|centre|thumb|Average Top-1 errors (%) of “PyramidNet + ShakeDrop” with several ranges of parameters of 4 runs at the final (300th) epoch on CIFAR-100 dataset in the “Batch” level. In some settings, it is equivalent to PyramidNet and PyramidDrop. Borrowed from ShakeDrop Regularization by Yamada et al., 2018.]]<br />
<br />
The setting that are used throughout the rest of the experiments are then <math>\alpha \in [-1,1]</math> and <math>\beta \in [0,1]</math>. Cases H and F outperform PyramidNet, suggesting that the strong perturbations imposed by ShakeDrop are functioning as intended. However, fully applying the perturbations in the backwards pass appears to destabilize the network, resulting in performance that is worse than standard PyramidNet.<br />
<br />
[[File:ParameterUpdateShakeDrop.png|400px|centre]]<br />
<br />
Following this initial parameter decision, the authors tested 4 different strategies for parameter update among "Batch" (same coefficients for all images in minibatch for each residual block), "Image" (same scaling coefficients for each image for each residual block), "Channel" (same scaling coefficients for each element for each residual block), and "Pixel" (same scaling coefficients for each element for each residual block). While Pixel was the best in terms of error rate, it is not very memory efficient, so Image was selected as it had the second best performance without the memory drawback.<br />
<br />
'''Comparison with Regularization Methods'''<br />
<br />
For these experiments, there are a few modifications that were made to assist with training. For ResNeXt, the EraseRelu formulation has each residual block ends in batch normalization. The Wide ResNet also is compared between vanilla with batch normalization and without. Batch normalization keeps the outputs of residual blocks in a certain range, as otherwise <math>\alpha</math> and <math>\beta</math> could cause perturbations that are too large, causing divergent learning. There is also a comparison of ResDrop/ShakeDrop Type A (where the regularization unit is inserted before the add unit for a residual branch) and after (where the regularization unit is inserted after the add unit for a residual branch). <br />
<br />
These experiments are performed on the CIFAR-100 dataset.<br />
<br />
[[File:ShakeDropArchitectureComparison1.png|800px|centre|thumb|]]<br />
<br />
[[File:ShakeDropArchitectureComparison2.png|800px|centre|thumb|]]<br />
<br />
[[File:ShakeDropArchitectureComparison3.png|800px|centre|thumb|]]<br />
<br />
For a final round of testing, the training setup was modified to incorporate other techniques used in state of the art methods. For most of the tests, the learning rate for the 300 epoch version started at 0.1 and decayed by a factor of 0.1 1/2 & 3/4 of the way through training. The alternative was cosine annealing, based on the presentation by Loshchilov and Hutter in their paper SGDR: Stochastic Gradient Descent with Warm Restarts. This is indicated in the Cos column, with a check indicating cosine annealing. <br />
<br />
[[File:CosineAnnealing.png|400px|centre|thumb|]]<br />
<br />
The Reg column indicates the regularization method used, either none, ResDrop (RD), Shake-Shake (SS), or ShakeDrop (SD). Fianlly, the Fil Column determines the type of data augmentation used, either none, cutout (CO) (DeVries & Taylor, 2017b), or Random Erasing (RE) (Zhong et al., 2017). <br />
<br />
[[File:ShakeDropComparison.png|800px|centre|thumb|Top-1 Errors (%) at final epoch on CIFAR-10/100 datasets]]<br />
<br />
'''State-of-the-Art Comparisons'''<br />
<br />
A direct comparison with state of the art methods is favorable for this new method. <br />
<br />
# Fair comparison of ResNeXt + Shake-Shake with PyramidNet + ShakeDrop gives an improvement of 0.19% on CIFAR-10 and 1.86% on CIFAR-100. Under these conditions, the final error rate is then 2.67% for CIFAR-10 and 13.99% for CIFAR-100.<br />
# Fair comparison of ResNeXt + Shake-Shake + Cutout with PyramidNet + ShakeDrop + Random Erasing gives an improvement of 0.25% on CIFAR-10 and 3.01% on CIFAR 100. Under these conditions, the final error rate is then 2.31% for CIFAR-10 and 12.19% for CIFAR-100.<br />
<br />
=Conclusion=<br />
<br />
This paper proposed a new stochastic regularization method, ShakeDrop, which outperforms previous state of the art methods while maintaining similar memory efficiency. It demonstrates that heavily perturbing a network can help to overcome issues with overfitting. It is also an effective way to regularize residual networks for image classification. The method was tested by CIFAR-10/100 and Tiny ImageNet datasets and showed great performance.<br />
<br />
=References=<br />
[Yamada et al., 2018] Yamada Y, Iwamura M, Kise K. ShakeDrop regularization. arXiv preprint arXiv:1802.02375. 2018 Feb 7.<br />
<br />
[He et al., 2016] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. CVPR, 2016.<br />
<br />
[Zagoruyko & Komodakis, 2016] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Proc. BMVC, 2016.<br />
<br />
[Han et al., 2017] Dongyoon Han, Jiwhan Kim, and Junmo Kim. Deep pyramidal residual networks. In Proc. CVPR, 2017a.<br />
<br />
[Xie et al., 2017] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proc. CVPR, 2017.<br />
<br />
[Huang et al., 2016] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Weinberger. Deep networks with stochastic depth. arXiv preprint arXiv:1603.09382v3, 2016.<br />
<br />
[Gastaldi, 2017] Xavier Gastaldi. Shake-shake regularization. arXiv preprint arXiv:1705.07485v2, 2017.<br />
<br />
[Loshilov & Hutter, 2016] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016.<br />
<br />
[DeVries & Taylor, 2017b] Terrance DeVries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017b.<br />
<br />
[Zhong et al., 2017] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. arXiv preprint arXiv:1708.04896, 2017.<br />
<br />
[Dutt et al., 2017] Anuvabh Dutt, Denis Pellerin, and Georges Qunot. Coupled ensembles of neural networks. arXiv preprint 1709.06053v1, 2017.<br />
<br />
[Veit et al., 2016] Andreas Veit, Michael J Wilber, and Serge Belongie. Residual networks behave like ensembles of relatively shallow networks. Advances in Neural Information Processing Systems 29, 2016.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=ShakeDrop_Regularization&diff=41775ShakeDrop Regularization2018-11-29T00:27:48Z<p>Ka2khan: /* Proposed Method */</p>
<hr />
<div>=Introduction=<br />
Current state of the art techniques for object classification are deep neural networks based on the residual block, first published by (He et al., 2016). This technique has been the foundation of several improved networks, including Wide ResNet (Zagoruyko & Komodakis, 2016), PyramdNet (Han et al., 2017) and ResNeXt (Xie et al., 2017). They have been further improved by regularization, such as Stochastic Depth (ResDrop) (Huang et al., 2016) and Shake-Shake (Gastaldi, 2017). Shake-Shake applied to ResNeXt has achieved one of the lowest error rates on the CIFAR-10 and CIFAR-100 datasets. However, it is only applicable to multi-branch architectures and is not memory efficient since it requires two branches of residual blocks to apply. This paper seeks to formulate a general expansion of Shake-Shake that can be applied to any residual block based network.<br />
<br />
=Existing Methods=<br />
<br />
'''Deep Approaches'''<br />
<br />
'''ResNet''', was the first use of residual blocks, a foundational feature in many modern state of the art convolution neural networks. They can be formulated as <math>G(x) = x + F(x)</math> where <math>x</math> and <math>G(x)</math> are the input and output of the residual block, and <math>F(x)</math> is the output of the residual branch on the residual block. A residual block typically performs a convolution operation and then passes the result plus its input onto the next block.<br />
<br />
[[File:ResidualBlock.png|600px|centre|thumb|An example of a simple residual block from Deep Residual Learning for Image Recognition by He et al., 2016]]<br />
<br />
ResNet is constructed out of a large number of these residual blocks sequentially stacked. It is interesting to note that having too many layers can cause overfitting, as pointed out by He et al. (2016) with the high error rates for the 1,202-layer ResNet on CIFAR datasets. Another paper (Veit et al., 2016) empirically showed that the cause of the high error rates can be mostly attributed to specific residual blocks whose channels increase greatly.<br />
<br />
'''PyramidNet''' is an important iteration that built on ResNet and WideResNet by gradually increasing channels on each residual block. The residual block is similar to those used in ResNet. It has been used to generate some of the first successful convolution neural networks with very large depth, at 272 layers. Amongst unmodified residual network architectures, it performs the best on the CIFAR datasets.<br />
<br />
[[File:ResidualBlockComparison.png|900px|centre|thumb|A simple illustration of different residual blocks from Deep Pyramidal Residual Networks by Han et al., 2017]]<br />
<br />
<br />
'''Non-Deep Approaches'''<br />
<br />
'''Wide ResNet''' modified ResNet by increasing channels in each layer, having a wider and shallower structure. Similarly to PyramidNet, this architecture avoids some of the pitfalls in the original formulation of ResNet.<br />
<br />
'''ResNeXt''' achieved performance beyond that of Wide ResNet with only a small increase in the number of parameters. It can be formulated as <math>G(x) = x + F_1(x)+F_2(x)</math>. In this case, <math>F_1(x)</math> and <math>F_2(x)</math> are the outputs of two paired convolution operations in a single residual block. The number of branches is not limited to 2, and will control the result of this network.<br />
<br />
<br />
[[File:SimplifiedResNeXt.png|600px|centre|thumb|Simplified ResNeXt Convolution Block. Yamada et al., 2018]]<br />
<br />
<br />
'''Regularization Methods'''<br />
<br />
'''Stochastic Depth''' helped address the issue of vanishing gradients in ResNet. It works by randomly dropping residual blocks. On the <math>l^{th}</math> residual block the Stochastic Depth process is given as <math>G(x)=x+b_lF(x)</math> where <math>b_l \in \{0,1\}</math> is a Bernoulli random variable with probability <math>p_l</math>. Using a constant value for <math>p_l</math> didn't work well, so instead a linear decay rule <math>p_l = 1 - \frac{l}{L}(1-p_L)</math> was used. In this equation, <math>L</math> is the number of layers, and <math>p_L</math> is the initial parameter. <br />
<br />
'''Shake-Shake''' is a regularization method that specifically improves the ResNeXt architecture. It can be given as <math>G(x)=x+\alpha F_1(x)+(1-\alpha)F_2(x)</math>, where <math>\alpha \in [0,1]</math> is a random coefficient. <math>\alpha</math> is used during the forward pass, and another identically distributed random parameter <math>\beta</math> is used in the backward pass. This caused one of the two paired convolution operations to be dropped, and further improved ResNeXt.<br />
<br />
[[File:Paper 32.jpg|600px|centre|thumb| Shake-Shake (ResNeXt + Shake-Shake) (Gastaldi, 2017), in which some processing layers omitted for conciseness.]]<br />
<br />
=Proposed Method=<br />
This paper seeks to generalize the method proposed in Shake-Shake to be applied to any residual structure network. Shake-Shake. The initial formulation of 1-branch shake is <math>G(x) = x + \alpha F(x)</math>. In this case, <math>\alpha</math> is a coefficient that disturbs the forward pass, but is not necessarily constrained to be [0,1]. Another corresponding coefficient <math>\beta</math> is used in the backwards pass. Applying this simple adaptation of Shake-Shake on a 110-layer version of PyramidNet with <math>\alpha \in [0,1]</math> and <math>\beta \in [0,1]</math> performs abysmally, with an error rate of 77.99%.<br />
<br />
This failure is a result of the setup causing too much perturbation. A trick is needed to promote learning with large perturbations, to preserve the regularization effect. The idea of the authors is to borrow from ResDrop and combine that with Shake-Shake. This works by randomly deciding whether to apply 1-branch shake. This creates in effect two networks, the original network without a regularization component, and a regularized network. When the non regularized network is selected, learning is promoted, when the perturbed network is selected, learning is disturbed. Achieving good performance requires a balance between the two. <br />
<br />
'''ShakeDrop''' is given as <br />
<br />
<div align="center"><br />
<math>G(x) = x + (b_l + \alpha - b_l \alpha)F(x)</math>,<br />
</div><br />
<br />
where <math>b_l</math> is a Bernoulli random variable following the linear decay rule used in Stochastic Depth. An alternative presentation is <br />
<br />
<div align="center"><br />
<math><br />
G(x) = \begin{cases}<br />
x + F(x) ~~ \text{if } b_l = 1 \\<br />
x + \alpha F(x) ~~ \text{otherwise}<br />
\end{cases}<br />
</math><br />
</div><br />
<br />
If <math>b_l = 1</math> then ShakeDrop is equivalent to the original network, otherwise it is the network + 1-branch Shake. Regardless of the value of <math>\beta</math> on the backwards pass, network weights will be updated.<br />
<br />
=Experiments=<br />
<br />
'''Parameter Search'''<br />
<br />
The authors experiments began with a hyperparameter search utilizing ShakeDrop on pyramidal networks. The results of this search are presented below. <br />
<br />
[[File:ShakeDropHyperParameterSearch.png|600px|centre|thumb|Average Top-1 errors (%) of “PyramidNet + ShakeDrop” with several ranges of parameters of 4 runs at the final (300th) epoch on CIFAR-100 dataset in the “Batch” level. In some settings, it is equivalent to PyramidNet and PyramidDrop. Borrowed from ShakeDrop Regularization by Yamada et al., 2018.]]<br />
<br />
The setting that are used throughout the rest of the experiments are then <math>\alpha \in [-1,1]</math> and <math>\beta \in [0,1]</math>. Cases H and F outperform PyramidNet, suggesting that the strong perturbations imposed by ShakeDrop are functioning as intended. However, fully applying the perturbations in the backwards pass appears to destabilize the network, resulting in performance that is worse than standard PyramidNet.<br />
<br />
[[File:ParameterUpdateShakeDrop.png|400px|centre]]<br />
<br />
Following this initial parameter decision, the authors tested 4 different strategies for parameter update among "Batch" (same coefficients for all images in minibatch for each residual block), "Image" (same scaling coefficients for each image for each residual block), "Channel" (same scaling coefficients for each element for each residual block), and "Pixel" (same scaling coefficients for each element for each residual block). While Pixel was the best in terms of error rate, it is not very memory efficient, so Image was selected as it had the second best performance without the memory drawback.<br />
<br />
'''Comparison with Regularization Methods'''<br />
<br />
For these experiments, there are a few modifications that were made to assist with training. For ResNeXt, the EraseRelu formulation has each residual block ends in batch normalization. The Wide ResNet also is compared between vanilla with batch normalization and without. Batch normalization keeps the outputs of residual blocks in a certain range, as otherwise <math>\alpha</math> and <math>\beta</math> could cause perturbations that are too large, causing divergent learning. There is also a comparison of ResDrop/ShakeDrop Type A (where the regularization unit is inserted before the add unit for a residual branch) and after (where the regularization unit is inserted after the add unit for a residual branch). <br />
<br />
These experiments are performed on the CIFAR-100 dataset.<br />
<br />
[[File:ShakeDropArchitectureComparison1.png|800px|centre|thumb|]]<br />
<br />
[[File:ShakeDropArchitectureComparison2.png|800px|centre|thumb|]]<br />
<br />
[[File:ShakeDropArchitectureComparison3.png|800px|centre|thumb|]]<br />
<br />
For a final round of testing, the training setup was modified to incorporate other techniques used in state of the art methods. For most of the tests, the learning rate for the 300 epoch version started at 0.1 and decayed by a factor of 0.1 1/2 & 3/4 of the way through training. The alternative was cosine annealing, based on the presentation by Loshchilov and Hutter in their paper SGDR: Stochastic Gradient Descent with Warm Restarts. This is indicated in the Cos column, with a check indicating cosine annealing. <br />
<br />
[[File:CosineAnnealing.png|400px|centre|thumb|]]<br />
<br />
The Reg column indicates the regularization method used, either none, ResDrop (RD), Shake-Shake (SS), or ShakeDrop (SD). Fianlly, the Fil Column determines the type of data augmentation used, either none, cutout (CO) (DeVries & Taylor, 2017b), or Random Erasing (RE) (Zhong et al., 2017). <br />
<br />
[[File:ShakeDropComparison.png|800px|centre|thumb|Top-1 Errors (%) at final epoch on CIFAR-10/100 datasets]]<br />
<br />
'''State-of-the-Art Comparisons'''<br />
<br />
A direct comparison with state of the art methods is favorable for this new method. <br />
<br />
# Fair comparison of ResNeXt + Shake-Shake with PyramidNet + ShakeDrop gives an improvement of 0.19% on CIFAR-10 and 1.86% on CIFAR-100. Under these conditions, the final error rate is then 2.67% for CIFAR-10 and 13.99% for CIFAR-100.<br />
# Fair comparison of ResNeXt + Shake-Shake + Cutout with PyramidNet + ShakeDrop + Random Erasing gives an improvement of 0.25% on CIFAR-10 and 3.01% on CIFAR 100. Under these conditions, the final error rate is then 2.31% for CIFAR-10 and 12.19% for CIFAR-100.<br />
<br />
=Conclusion=<br />
<br />
This paper proposed a new stochastic regularization method, ShakeDrop, which outperforms previous state of the art methods while maintaining similar memory efficiency. It demonstrates that heavily perturbing a network can help to overcome issues with overfitting. It is also an effective way to regularize residual networks for image classification. The method was tested by CIFAR-10/100 and Tiny ImageNet datasets and showed great performance.<br />
<br />
=References=<br />
[Yamada et al., 2018] Yamada Y, Iwamura M, Kise K. ShakeDrop regularization. arXiv preprint arXiv:1802.02375. 2018 Feb 7.<br />
<br />
[He et al., 2016] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. CVPR, 2016.<br />
<br />
[Zagoruyko & Komodakis, 2016] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Proc. BMVC, 2016.<br />
<br />
[Han et al., 2017] Dongyoon Han, Jiwhan Kim, and Junmo Kim. Deep pyramidal residual networks. In Proc. CVPR, 2017a.<br />
<br />
[Xie et al., 2017] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proc. CVPR, 2017.<br />
<br />
[Huang et al., 2016] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Weinberger. Deep networks with stochastic depth. arXiv preprint arXiv:1603.09382v3, 2016.<br />
<br />
[Gastaldi, 2017] Xavier Gastaldi. Shake-shake regularization. arXiv preprint arXiv:1705.07485v2, 2017.<br />
<br />
[Loshilov & Hutter, 2016] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016.<br />
<br />
[DeVries & Taylor, 2017b] Terrance DeVries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017b.<br />
<br />
[Zhong et al., 2017] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. arXiv preprint arXiv:1708.04896, 2017.<br />
<br />
[Dutt et al., 2017] Anuvabh Dutt, Denis Pellerin, and Georges Qunot. Coupled ensembles of neural networks. arXiv preprint 1709.06053v1, 2017.<br />
<br />
[Veit et al., 2016] Andreas Veit, Michael J Wilber, and Serge Belongie. Residual networks behave like ensembles of relatively shallow networks. Advances in Neural Information Processing Systems 29, 2016.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=ShakeDrop_Regularization&diff=41774ShakeDrop Regularization2018-11-29T00:20:51Z<p>Ka2khan: /* Existing Methods */</p>
<hr />
<div>=Introduction=<br />
Current state of the art techniques for object classification are deep neural networks based on the residual block, first published by (He et al., 2016). This technique has been the foundation of several improved networks, including Wide ResNet (Zagoruyko & Komodakis, 2016), PyramdNet (Han et al., 2017) and ResNeXt (Xie et al., 2017). They have been further improved by regularization, such as Stochastic Depth (ResDrop) (Huang et al., 2016) and Shake-Shake (Gastaldi, 2017). Shake-Shake applied to ResNeXt has achieved one of the lowest error rates on the CIFAR-10 and CIFAR-100 datasets. However, it is only applicable to multi-branch architectures and is not memory efficient since it requires two branches of residual blocks to apply. This paper seeks to formulate a general expansion of Shake-Shake that can be applied to any residual block based network.<br />
<br />
=Existing Methods=<br />
<br />
'''Deep Approaches'''<br />
<br />
'''ResNet''', was the first use of residual blocks, a foundational feature in many modern state of the art convolution neural networks. They can be formulated as <math>G(x) = x + F(x)</math> where <math>x</math> and <math>G(x)</math> are the input and output of the residual block, and <math>F(x)</math> is the output of the residual branch on the residual block. A residual block typically performs a convolution operation and then passes the result plus its input onto the next block.<br />
<br />
[[File:ResidualBlock.png|600px|centre|thumb|An example of a simple residual block from Deep Residual Learning for Image Recognition by He et al., 2016]]<br />
<br />
ResNet is constructed out of a large number of these residual blocks sequentially stacked. It is interesting to note that having too many layers can cause overfitting, as pointed out by He et al. (2016) with the high error rates for the 1,202-layer ResNet on CIFAR datasets. Another paper (Veit et al., 2016) empirically showed that the cause of the high error rates can be mostly attributed to specific residual blocks whose channels increase greatly.<br />
<br />
'''PyramidNet''' is an important iteration that built on ResNet and WideResNet by gradually increasing channels on each residual block. The residual block is similar to those used in ResNet. It has been used to generate some of the first successful convolution neural networks with very large depth, at 272 layers. Amongst unmodified residual network architectures, it performs the best on the CIFAR datasets.<br />
<br />
[[File:ResidualBlockComparison.png|900px|centre|thumb|A simple illustration of different residual blocks from Deep Pyramidal Residual Networks by Han et al., 2017]]<br />
<br />
<br />
'''Non-Deep Approaches'''<br />
<br />
'''Wide ResNet''' modified ResNet by increasing channels in each layer, having a wider and shallower structure. Similarly to PyramidNet, this architecture avoids some of the pitfalls in the original formulation of ResNet.<br />
<br />
'''ResNeXt''' achieved performance beyond that of Wide ResNet with only a small increase in the number of parameters. It can be formulated as <math>G(x) = x + F_1(x)+F_2(x)</math>. In this case, <math>F_1(x)</math> and <math>F_2(x)</math> are the outputs of two paired convolution operations in a single residual block. The number of branches is not limited to 2, and will control the result of this network.<br />
<br />
<br />
[[File:SimplifiedResNeXt.png|600px|centre|thumb|Simplified ResNeXt Convolution Block. Yamada et al., 2018]]<br />
<br />
<br />
'''Regularization Methods'''<br />
<br />
'''Stochastic Depth''' helped address the issue of vanishing gradients in ResNet. It works by randomly dropping residual blocks. On the <math>l^{th}</math> residual block the Stochastic Depth process is given as <math>G(x)=x+b_lF(x)</math> where <math>b_l \in \{0,1\}</math> is a Bernoulli random variable with probability <math>p_l</math>. Using a constant value for <math>p_l</math> didn't work well, so instead a linear decay rule <math>p_l = 1 - \frac{l}{L}(1-p_L)</math> was used. In this equation, <math>L</math> is the number of layers, and <math>p_L</math> is the initial parameter. <br />
<br />
'''Shake-Shake''' is a regularization method that specifically improves the ResNeXt architecture. It can be given as <math>G(x)=x+\alpha F_1(x)+(1-\alpha)F_2(x)</math>, where <math>\alpha \in [0,1]</math> is a random coefficient. <math>\alpha</math> is used during the forward pass, and another identically distributed random parameter <math>\beta</math> is used in the backward pass. This caused one of the two paired convolution operations to be dropped, and further improved ResNeXt.<br />
<br />
[[File:Paper 32.jpg|600px|centre|thumb| Shake-Shake (ResNeXt + Shake-Shake) (Gastaldi, 2017), in which some processing layers omitted for conciseness.]]<br />
<br />
=Proposed Method=<br />
This paper seeks to generalize the method proposed in Shake-Shake to be applied to any residual structure network. Shake-Shake. The initial formulation of 1-branch shake is <math>G(x) = x + \alpha F(x)</math>. In this case, <math>\alpha</math> is a coefficient that disturbs the forward pass, but is not necessarily constrained to be [0,1]. Another corresponding coefficient <math>\beta</math> is used in the backwards pass. Applying this simple adaptation of Shake-Shake on a 110-layer version of PyramidNet with <math>\alpha \in [0,1]</math> and <math>\beta \in [0,1]</math> performs abysmally, with an error rate of 77.99%.<br />
<br />
This failure is a result of the setup causing too much perturbation. A trick is needed to promote learning with large perturbations, to preserve the regularization effect. The idea of the authors is to borrow from ResDrop and combine that with Shake-Shake. This works by randomly deciding whether to apply 1-branch shake. This in creates in effect two networks, the original network without a regularization component, and a regularized network. When the non regularized network is selected, learning is promoted, when the perturbed network is selected, learning is disturbed. Achieving good performance requires a balance between the two. <br />
<br />
'''ShakeDrop''' is given as <br />
<br />
<div align="center"><br />
<math>G(x) = x + (b_l + \alpha - b_l \alpha)F(x)</math>,<br />
</div><br />
<br />
where <math>b_l</math> is a Bernoulli random variable following the linear decay rule used in Stochastic Depth. An alternative presentation is <br />
<br />
<div align="center"><br />
<math><br />
G(x) = \begin{cases}<br />
x + F(x) ~~ \text{if } b_l = 1 \\<br />
x + \alpha F(x) ~~ \text{otherwise}<br />
\end{cases}<br />
</math><br />
</div><br />
<br />
If <math>b_l = 1</math> then ShakeDrop is equivalent to the original network, otherwise it is the network + 1-branch Shake. Regardless of the value of <math>\beta</math> on the backwards pass, network weights will be updated.<br />
<br />
=Experiments=<br />
<br />
'''Parameter Search'''<br />
<br />
The authors experiments began with a hyperparameter search utilizing ShakeDrop on pyramidal networks. The results of this search are presented below. <br />
<br />
[[File:ShakeDropHyperParameterSearch.png|600px|centre|thumb|Average Top-1 errors (%) of “PyramidNet + ShakeDrop” with several ranges of parameters of 4 runs at the final (300th) epoch on CIFAR-100 dataset in the “Batch” level. In some settings, it is equivalent to PyramidNet and PyramidDrop. Borrowed from ShakeDrop Regularization by Yamada et al., 2018.]]<br />
<br />
The setting that are used throughout the rest of the experiments are then <math>\alpha \in [-1,1]</math> and <math>\beta \in [0,1]</math>. Cases H and F outperform PyramidNet, suggesting that the strong perturbations imposed by ShakeDrop are functioning as intended. However, fully applying the perturbations in the backwards pass appears to destabilize the network, resulting in performance that is worse than standard PyramidNet.<br />
<br />
[[File:ParameterUpdateShakeDrop.png|400px|centre]]<br />
<br />
Following this initial parameter decision, the authors tested 4 different strategies for parameter update among "Batch" (same coefficients for all images in minibatch for each residual block), "Image" (same scaling coefficients for each image for each residual block), "Channel" (same scaling coefficients for each element for each residual block), and "Pixel" (same scaling coefficients for each element for each residual block). While Pixel was the best in terms of error rate, it is not very memory efficient, so Image was selected as it had the second best performance without the memory drawback.<br />
<br />
'''Comparison with Regularization Methods'''<br />
<br />
For these experiments, there are a few modifications that were made to assist with training. For ResNeXt, the EraseRelu formulation has each residual block ends in batch normalization. The Wide ResNet also is compared between vanilla with batch normalization and without. Batch normalization keeps the outputs of residual blocks in a certain range, as otherwise <math>\alpha</math> and <math>\beta</math> could cause perturbations that are too large, causing divergent learning. There is also a comparison of ResDrop/ShakeDrop Type A (where the regularization unit is inserted before the add unit for a residual branch) and after (where the regularization unit is inserted after the add unit for a residual branch). <br />
<br />
These experiments are performed on the CIFAR-100 dataset.<br />
<br />
[[File:ShakeDropArchitectureComparison1.png|800px|centre|thumb|]]<br />
<br />
[[File:ShakeDropArchitectureComparison2.png|800px|centre|thumb|]]<br />
<br />
[[File:ShakeDropArchitectureComparison3.png|800px|centre|thumb|]]<br />
<br />
For a final round of testing, the training setup was modified to incorporate other techniques used in state of the art methods. For most of the tests, the learning rate for the 300 epoch version started at 0.1 and decayed by a factor of 0.1 1/2 & 3/4 of the way through training. The alternative was cosine annealing, based on the presentation by Loshchilov and Hutter in their paper SGDR: Stochastic Gradient Descent with Warm Restarts. This is indicated in the Cos column, with a check indicating cosine annealing. <br />
<br />
[[File:CosineAnnealing.png|400px|centre|thumb|]]<br />
<br />
The Reg column indicates the regularization method used, either none, ResDrop (RD), Shake-Shake (SS), or ShakeDrop (SD). Fianlly, the Fil Column determines the type of data augmentation used, either none, cutout (CO) (DeVries & Taylor, 2017b), or Random Erasing (RE) (Zhong et al., 2017). <br />
<br />
[[File:ShakeDropComparison.png|800px|centre|thumb|Top-1 Errors (%) at final epoch on CIFAR-10/100 datasets]]<br />
<br />
'''State-of-the-Art Comparisons'''<br />
<br />
A direct comparison with state of the art methods is favorable for this new method. <br />
<br />
# Fair comparison of ResNeXt + Shake-Shake with PyramidNet + ShakeDrop gives an improvement of 0.19% on CIFAR-10 and 1.86% on CIFAR-100. Under these conditions, the final error rate is then 2.67% for CIFAR-10 and 13.99% for CIFAR-100.<br />
# Fair comparison of ResNeXt + Shake-Shake + Cutout with PyramidNet + ShakeDrop + Random Erasing gives an improvement of 0.25% on CIFAR-10 and 3.01% on CIFAR 100. Under these conditions, the final error rate is then 2.31% for CIFAR-10 and 12.19% for CIFAR-100.<br />
<br />
=Conclusion=<br />
<br />
This paper proposed a new stochastic regularization method, ShakeDrop, which outperforms previous state of the art methods while maintaining similar memory efficiency. It demonstrates that heavily perturbing a network can help to overcome issues with overfitting. It is also an effective way to regularize residual networks for image classification. The method was tested by CIFAR-10/100 and Tiny ImageNet datasets and showed great performance.<br />
<br />
=References=<br />
[Yamada et al., 2018] Yamada Y, Iwamura M, Kise K. ShakeDrop regularization. arXiv preprint arXiv:1802.02375. 2018 Feb 7.<br />
<br />
[He et al., 2016] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. CVPR, 2016.<br />
<br />
[Zagoruyko & Komodakis, 2016] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Proc. BMVC, 2016.<br />
<br />
[Han et al., 2017] Dongyoon Han, Jiwhan Kim, and Junmo Kim. Deep pyramidal residual networks. In Proc. CVPR, 2017a.<br />
<br />
[Xie et al., 2017] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proc. CVPR, 2017.<br />
<br />
[Huang et al., 2016] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Weinberger. Deep networks with stochastic depth. arXiv preprint arXiv:1603.09382v3, 2016.<br />
<br />
[Gastaldi, 2017] Xavier Gastaldi. Shake-shake regularization. arXiv preprint arXiv:1705.07485v2, 2017.<br />
<br />
[Loshilov & Hutter, 2016] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016.<br />
<br />
[DeVries & Taylor, 2017b] Terrance DeVries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017b.<br />
<br />
[Zhong et al., 2017] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. arXiv preprint arXiv:1708.04896, 2017.<br />
<br />
[Dutt et al., 2017] Anuvabh Dutt, Denis Pellerin, and Georges Qunot. Coupled ensembles of neural networks. arXiv preprint 1709.06053v1, 2017.<br />
<br />
[Veit et al., 2016] Andreas Veit, Michael J Wilber, and Serge Belongie. Residual networks behave like ensembles of relatively shallow networks. Advances in Neural Information Processing Systems 29, 2016.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=ShakeDrop_Regularization&diff=41772ShakeDrop Regularization2018-11-29T00:12:55Z<p>Ka2khan: /* Existing Methods */</p>
<hr />
<div>=Introduction=<br />
Current state of the art techniques for object classification are deep neural networks based on the residual block, first published by (He et al., 2016). This technique has been the foundation of several improved networks, including Wide ResNet (Zagoruyko & Komodakis, 2016), PyramdNet (Han et al., 2017) and ResNeXt (Xie et al., 2017). They have been further improved by regularization, such as Stochastic Depth (ResDrop) (Huang et al., 2016) and Shake-Shake (Gastaldi, 2017). Shake-Shake applied to ResNeXt has achieved one of the lowest error rates on the CIFAR-10 and CIFAR-100 datasets. However, it is only applicable to multi-branch architectures and is not memory efficient. This paper seeks to formulate a general expansion of Shake-Shake that can be applied to any residual block based network.<br />
<br />
=Existing Methods=<br />
<br />
'''Deep Approaches'''<br />
<br />
'''ResNet''', was the first use of residual blocks, a foundational feature in many modern state of the art convolution neural networks. They can be formulated as <math>G(x) = x + F(x)</math> where <math>x</math> and <math>G(x)</math> are the input and output of the residual block, and <math>F(x)</math> is the output of the residual branch on the residual block. A residual block typically performs a convolution operation and then passes the result plus its input onto the next block.<br />
<br />
[[File:ResidualBlock.png|600px|centre|thumb|An example of a simple residual block from Deep Residual Learning for Image Recognition by He et al., 2016]]<br />
<br />
ResNet is constructed out of a large number of these residual blocks sequentially stacked. It is interesting to note that having too many layers can cause overfitting, as pointed out by He et al. (2016) with the high error rates for the 1,202-layer ResNet on CIFAR datasets. Another paper (Veit et al., 2016) empirically showed that the cause of the high error rates can be mostly attributed to specific residual blocks whose channels increase greatly.<br />
<br />
'''PyramidNet''' is an important iteration that built on ResNet and WideResNet by gradually increasing channels on each residual block. The residual block is similar to those used in ResNet. It has been use to generate some of the first successful convolution neural networks with very large depth, at 272 layers. Amongst unmodified network architectures, it performs the best on the CIFAR datasets.<br />
<br />
[[File:ResidualBlockComparison.png|900px|centre|thumb|A simple illustration of different residual blocks from Deep Pyramidal Residual Networks by Han et al., 2017]]<br />
<br />
<br />
'''Non-Deep Approaches'''<br />
<br />
'''Wide ResNet''' modified ResNet by increasing channels in each layer, having a wider and shallower structure. Similarly to PyramidNet, this architecture avoids some of the pitfalls in the orginal formulation of ResNet.<br />
<br />
'''ResNeXt''' achieved performance beyond that of Wide ResNet with only a small increase in the number of parameters. It can be formulated as <math>G(x) = x + F_1(x)+F_2(x)</math>. In this case, <math>F_1(x)</math> and <math>F_2(x)</math> are the outputs of two paired convolution operations in a single residual block. The number of branches is not limited to 2, and will control the result of this network.<br />
<br />
<br />
[[File:SimplifiedResNeXt.png|600px|centre|thumb|Simplified ResNeXt Convolution Block. Yamada et al., 2018]]<br />
<br />
<br />
'''Regularization Methods'''<br />
<br />
'''Stochastic Depth''' helped address the issue of vanishing gradients in ResNet. It works by randomly dropping residual blocks. On the <math>l^{th}</math> residual block the Stochastic Depth process is given as <math>G(x)=x+b_lF(x)</math> where <math>b_l \in \{0,1\}</math> is a Bernoulli random variable with probability <math>p_l</math>. Using a constant value for <math>p_l</math> didn't work well, so instead a linear decay rule <math>p_l = 1 - \frac{l}{L}(1-p_L)</math> was used. In this equation, <math>L</math> is the number of layers, and <math>p_L</math> is the initial parameter. <br />
<br />
'''Shake-Shake''' is a regularization method that specifically improves the ResNeXt architecture. It can be given as <math>G(x)=x+\alpha F_1(x)+(1-\alpha)F_2(x)</math>, where <math>\alpha \in [0,1]</math> is a random coefficient. <math>\alpha</math> is used during the forward pass, and another identically distributed random parameter <math>\beta</math> is used in the backward pass. This caused one of the two paired convolution operations to be dropped, and further improved ResNeXt.<br />
<br />
[[File:Paper 32.jpg|600px|centre|thumb| Shake-Shake (ResNeXt + Shake-Shake) (Gastaldi, 2017), in which some processing layers omitted for conciseness.]]<br />
<br />
=Proposed Method=<br />
This paper seeks to generalize the method proposed in Shake-Shake to be applied to any residual structure network. Shake-Shake. The initial formulation of 1-branch shake is <math>G(x) = x + \alpha F(x)</math>. In this case, <math>\alpha</math> is a coefficient that disturbs the forward pass, but is not necessarily constrained to be [0,1]. Another corresponding coefficient <math>\beta</math> is used in the backwards pass. Applying this simple adaptation of Shake-Shake on a 110-layer version of PyramidNet with <math>\alpha \in [0,1]</math> and <math>\beta \in [0,1]</math> performs abysmally, with an error rate of 77.99%.<br />
<br />
This failure is a result of the setup causing too much perturbation. A trick is needed to promote learning with large perturbations, to preserve the regularization effect. The idea of the authors is to borrow from ResDrop and combine that with Shake-Shake. This works by randomly deciding whether to apply 1-branch shake. This in creates in effect two networks, the original network without a regularization component, and a regularized network. When the non regularized network is selected, learning is promoted, when the perturbed network is selected, learning is disturbed. Achieving good performance requires a balance between the two. <br />
<br />
'''ShakeDrop''' is given as <br />
<br />
<div align="center"><br />
<math>G(x) = x + (b_l + \alpha - b_l \alpha)F(x)</math>,<br />
</div><br />
<br />
where <math>b_l</math> is a Bernoulli random variable following the linear decay rule used in Stochastic Depth. An alternative presentation is <br />
<br />
<div align="center"><br />
<math><br />
G(x) = \begin{cases}<br />
x + F(x) ~~ \text{if } b_l = 1 \\<br />
x + \alpha F(x) ~~ \text{otherwise}<br />
\end{cases}<br />
</math><br />
</div><br />
<br />
If <math>b_l = 1</math> then ShakeDrop is equivalent to the original network, otherwise it is the network + 1-branch Shake. Regardless of the value of <math>\beta</math> on the backwards pass, network weights will be updated.<br />
<br />
=Experiments=<br />
<br />
'''Parameter Search'''<br />
<br />
The authors experiments began with a hyperparameter search utilizing ShakeDrop on pyramidal networks. The results of this search are presented below. <br />
<br />
[[File:ShakeDropHyperParameterSearch.png|600px|centre|thumb|Average Top-1 errors (%) of “PyramidNet + ShakeDrop” with several ranges of parameters of 4 runs at the final (300th) epoch on CIFAR-100 dataset in the “Batch” level. In some settings, it is equivalent to PyramidNet and PyramidDrop. Borrowed from ShakeDrop Regularization by Yamada et al., 2018.]]<br />
<br />
The setting that are used throughout the rest of the experiments are then <math>\alpha \in [-1,1]</math> and <math>\beta \in [0,1]</math>. Cases H and F outperform PyramidNet, suggesting that the strong perturbations imposed by ShakeDrop are functioning as intended. However, fully applying the perturbations in the backwards pass appears to destabilize the network, resulting in performance that is worse than standard PyramidNet.<br />
<br />
[[File:ParameterUpdateShakeDrop.png|400px|centre]]<br />
<br />
Following this initial parameter decision, the authors tested 4 different strategies for parameter update among "Batch" (same coefficients for all images in minibatch for each residual block), "Image" (same scaling coefficients for each image for each residual block), "Channel" (same scaling coefficients for each element for each residual block), and "Pixel" (same scaling coefficients for each element for each residual block). While Pixel was the best in terms of error rate, it is not very memory efficient, so Image was selected as it had the second best performance without the memory drawback.<br />
<br />
'''Comparison with Regularization Methods'''<br />
<br />
For these experiments, there are a few modifications that were made to assist with training. For ResNeXt, the EraseRelu formulation has each residual block ends in batch normalization. The Wide ResNet also is compared between vanilla with batch normalization and without. Batch normalization keeps the outputs of residual blocks in a certain range, as otherwise <math>\alpha</math> and <math>\beta</math> could cause perturbations that are too large, causing divergent learning. There is also a comparison of ResDrop/ShakeDrop Type A (where the regularization unit is inserted before the add unit for a residual branch) and after (where the regularization unit is inserted after the add unit for a residual branch). <br />
<br />
These experiments are performed on the CIFAR-100 dataset.<br />
<br />
[[File:ShakeDropArchitectureComparison1.png|800px|centre|thumb|]]<br />
<br />
[[File:ShakeDropArchitectureComparison2.png|800px|centre|thumb|]]<br />
<br />
[[File:ShakeDropArchitectureComparison3.png|800px|centre|thumb|]]<br />
<br />
For a final round of testing, the training setup was modified to incorporate other techniques used in state of the art methods. For most of the tests, the learning rate for the 300 epoch version started at 0.1 and decayed by a factor of 0.1 1/2 & 3/4 of the way through training. The alternative was cosine annealing, based on the presentation by Loshchilov and Hutter in their paper SGDR: Stochastic Gradient Descent with Warm Restarts. This is indicated in the Cos column, with a check indicating cosine annealing. <br />
<br />
[[File:CosineAnnealing.png|400px|centre|thumb|]]<br />
<br />
The Reg column indicates the regularization method used, either none, ResDrop (RD), Shake-Shake (SS), or ShakeDrop (SD). Fianlly, the Fil Column determines the type of data augmentation used, either none, cutout (CO) (DeVries & Taylor, 2017b), or Random Erasing (RE) (Zhong et al., 2017). <br />
<br />
[[File:ShakeDropComparison.png|800px|centre|thumb|Top-1 Errors (%) at final epoch on CIFAR-10/100 datasets]]<br />
<br />
'''State-of-the-Art Comparisons'''<br />
<br />
A direct comparison with state of the art methods is favorable for this new method. <br />
<br />
# Fair comparison of ResNeXt + Shake-Shake with PyramidNet + ShakeDrop gives an improvement of 0.19% on CIFAR-10 and 1.86% on CIFAR-100. Under these conditions, the final error rate is then 2.67% for CIFAR-10 and 13.99% for CIFAR-100.<br />
# Fair comparison of ResNeXt + Shake-Shake + Cutout with PyramidNet + ShakeDrop + Random Erasing gives an improvement of 0.25% on CIFAR-10 and 3.01% on CIFAR 100. Under these conditions, the final error rate is then 2.31% for CIFAR-10 and 12.19% for CIFAR-100.<br />
<br />
=Conclusion=<br />
<br />
This paper proposed a new stochastic regularization method, ShakeDrop, which outperforms previous state of the art methods while maintaining similar memory efficiency. It demonstrates that heavily perturbing a network can help to overcome issues with overfitting. It is also an effective way to regularize residual networks for image classification. The method was tested by CIFAR-10/100 and Tiny ImageNet datasets and showed great performance.<br />
<br />
=References=<br />
[Yamada et al., 2018] Yamada Y, Iwamura M, Kise K. ShakeDrop regularization. arXiv preprint arXiv:1802.02375. 2018 Feb 7.<br />
<br />
[He et al., 2016] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. CVPR, 2016.<br />
<br />
[Zagoruyko & Komodakis, 2016] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Proc. BMVC, 2016.<br />
<br />
[Han et al., 2017] Dongyoon Han, Jiwhan Kim, and Junmo Kim. Deep pyramidal residual networks. In Proc. CVPR, 2017a.<br />
<br />
[Xie et al., 2017] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proc. CVPR, 2017.<br />
<br />
[Huang et al., 2016] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Weinberger. Deep networks with stochastic depth. arXiv preprint arXiv:1603.09382v3, 2016.<br />
<br />
[Gastaldi, 2017] Xavier Gastaldi. Shake-shake regularization. arXiv preprint arXiv:1705.07485v2, 2017.<br />
<br />
[Loshilov & Hutter, 2016] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016.<br />
<br />
[DeVries & Taylor, 2017b] Terrance DeVries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017b.<br />
<br />
[Zhong et al., 2017] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. arXiv preprint arXiv:1708.04896, 2017.<br />
<br />
[Dutt et al., 2017] Anuvabh Dutt, Denis Pellerin, and Georges Qunot. Coupled ensembles of neural networks. arXiv preprint 1709.06053v1, 2017.<br />
<br />
[Veit et al., 2016] Andreas Veit, Michael J Wilber, and Serge Belongie. Residual networks behave like ensembles of relatively shallow networks. Advances in Neural Information Processing Systems 29, 2016.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=ShakeDrop_Regularization&diff=41771ShakeDrop Regularization2018-11-29T00:08:35Z<p>Ka2khan: /* References */</p>
<hr />
<div>=Introduction=<br />
Current state of the art techniques for object classification are deep neural networks based on the residual block, first published by (He et al., 2016). This technique has been the foundation of several improved networks, including Wide ResNet (Zagoruyko & Komodakis, 2016), PyramdNet (Han et al., 2017) and ResNeXt (Xie et al., 2017). They have been further improved by regularization, such as Stochastic Depth (ResDrop) (Huang et al., 2016) and Shake-Shake (Gastaldi, 2017). Shake-Shake applied to ResNeXt has achieved one of the lowest error rates on the CIFAR-10 and CIFAR-100 datasets. However, it is only applicable to multi-branch architectures and is not memory efficient. This paper seeks to formulate a general expansion of Shake-Shake that can be applied to any residual block based network.<br />
<br />
=Existing Methods=<br />
<br />
'''Deep Approaches'''<br />
<br />
'''ResNet''', was the first use of residual blocks, a foundational feature in many modern state of the art convolution neural networks. They can be formulated as <math>G(x) = x + F(x)</math> where <math>x</math> and <math>G(x)</math> are the input and output of the residual block, and <math>F(x)</math> is the output of the residual block. A residual block typically performs a convolution operation and then passes the result plus its input onto the next block.<br />
<br />
[[File:ResidualBlock.png|600px|centre|thumb|An example of a simple residual block from Deep Residual Learning for Image Recognition by He et al., 2016]]<br />
<br />
ResNet is constructed out of a large number of these residual blocks sequentially stacked. It is interesting to note that having too many layers can cause overfitting, as pointed out by He et al. (2016) with the high error rates for the 1,202-layer ResNet on CIFAR datasets.<br />
<br />
'''PyramidNet''' is an important iteration that built on ResNet and WideResNet by gradually increasing channels on each residual block. The residual block is similar to those used in ResNet. It has been use to generate some of the first successful convolution neural networks with very large depth, at 272 layers. Amongst unmodified network architectures, it performs the best on the CIFAR datasets.<br />
<br />
[[File:ResidualBlockComparison.png|900px|centre|thumb|A simple illustration of different residual blocks from Deep Pyramidal Residual Networks by Han et al., 2017]]<br />
<br />
<br />
'''Non-Deep Approaches'''<br />
<br />
'''Wide ResNet''' modified ResNet by increasing channels in each layer, having a wider and shallower structure. Similarly to PyramidNet, this architecture avoids some of the pitfalls in the orginal formulation of ResNet.<br />
<br />
'''ResNeXt''' achieved performance beyond that of Wide ResNet with only a small increase in the number of parameters. It can be formulated as <math>G(x) = x + F_1(x)+F_2(x)</math>. In this case, <math>F_1(x)</math> and <math>F_2(x)</math> are the outputs of two paired convolution operations in a single residual block. The number of branches is not limited to 2, and will control the result of this network.<br />
<br />
<br />
[[File:SimplifiedResNeXt.png|600px|centre|thumb|Simplified ResNeXt Convolution Block. Yamada et al., 2018]]<br />
<br />
<br />
'''Regularization Methods'''<br />
<br />
'''Stochastic Depth''' helped address the issue of vanishing gradients in ResNet. It works by randomly dropping residual blocks. On the <math>l^{th}</math> residual block the Stochastic Depth process is given as <math>G(x)=x+b_lF(x)</math> where <math>b_l \in \{0,1\}</math> is a Bernoulli random variable with probability <math>p_l</math>. Using a constant value for <math>p_l</math> didn't work well, so instead a linear decay rule <math>p_l = 1 - \frac{l}{L}(1-p_L)</math> was used. In this equation, <math>L</math> is the number of layers, and <math>p_L</math> is the initial parameter. <br />
<br />
'''Shake-Shake''' is a regularization method that specifically improves the ResNeXt architecture. It can be given as <math>G(x)=x+\alpha F_1(x)+(1-\alpha)F_2(x)</math>, where <math>\alpha \in [0,1]</math> is a random coefficient. <math>\alpha</math> is used during the forward pass, and another identically distributed random parameter <math>\beta</math> is used in the backward pass. This caused one of the two paired convolution operations to be dropped, and further improved ResNeXt.<br />
<br />
[[File:Paper 32.jpg|600px|centre|thumb| Shake-Shake (ResNeXt + Shake-Shake) (Gastaldi, 2017), in which some processing layers omitted for conciseness.]]<br />
<br />
=Proposed Method=<br />
This paper seeks to generalize the method proposed in Shake-Shake to be applied to any residual structure network. Shake-Shake. The initial formulation of 1-branch shake is <math>G(x) = x + \alpha F(x)</math>. In this case, <math>\alpha</math> is a coefficient that disturbs the forward pass, but is not necessarily constrained to be [0,1]. Another corresponding coefficient <math>\beta</math> is used in the backwards pass. Applying this simple adaptation of Shake-Shake on a 110-layer version of PyramidNet with <math>\alpha \in [0,1]</math> and <math>\beta \in [0,1]</math> performs abysmally, with an error rate of 77.99%.<br />
<br />
This failure is a result of the setup causing too much perturbation. A trick is needed to promote learning with large perturbations, to preserve the regularization effect. The idea of the authors is to borrow from ResDrop and combine that with Shake-Shake. This works by randomly deciding whether to apply 1-branch shake. This in creates in effect two networks, the original network without a regularization component, and a regularized network. When the non regularized network is selected, learning is promoted, when the perturbed network is selected, learning is disturbed. Achieving good performance requires a balance between the two. <br />
<br />
'''ShakeDrop''' is given as <br />
<br />
<div align="center"><br />
<math>G(x) = x + (b_l + \alpha - b_l \alpha)F(x)</math>,<br />
</div><br />
<br />
where <math>b_l</math> is a Bernoulli random variable following the linear decay rule used in Stochastic Depth. An alternative presentation is <br />
<br />
<div align="center"><br />
<math><br />
G(x) = \begin{cases}<br />
x + F(x) ~~ \text{if } b_l = 1 \\<br />
x + \alpha F(x) ~~ \text{otherwise}<br />
\end{cases}<br />
</math><br />
</div><br />
<br />
If <math>b_l = 1</math> then ShakeDrop is equivalent to the original network, otherwise it is the network + 1-branch Shake. Regardless of the value of <math>\beta</math> on the backwards pass, network weights will be updated.<br />
<br />
=Experiments=<br />
<br />
'''Parameter Search'''<br />
<br />
The authors experiments began with a hyperparameter search utilizing ShakeDrop on pyramidal networks. The results of this search are presented below. <br />
<br />
[[File:ShakeDropHyperParameterSearch.png|600px|centre|thumb|Average Top-1 errors (%) of “PyramidNet + ShakeDrop” with several ranges of parameters of 4 runs at the final (300th) epoch on CIFAR-100 dataset in the “Batch” level. In some settings, it is equivalent to PyramidNet and PyramidDrop. Borrowed from ShakeDrop Regularization by Yamada et al., 2018.]]<br />
<br />
The setting that are used throughout the rest of the experiments are then <math>\alpha \in [-1,1]</math> and <math>\beta \in [0,1]</math>. Cases H and F outperform PyramidNet, suggesting that the strong perturbations imposed by ShakeDrop are functioning as intended. However, fully applying the perturbations in the backwards pass appears to destabilize the network, resulting in performance that is worse than standard PyramidNet.<br />
<br />
[[File:ParameterUpdateShakeDrop.png|400px|centre]]<br />
<br />
Following this initial parameter decision, the authors tested 4 different strategies for parameter update among "Batch" (same coefficients for all images in minibatch for each residual block), "Image" (same scaling coefficients for each image for each residual block), "Channel" (same scaling coefficients for each element for each residual block), and "Pixel" (same scaling coefficients for each element for each residual block). While Pixel was the best in terms of error rate, it is not very memory efficient, so Image was selected as it had the second best performance without the memory drawback.<br />
<br />
'''Comparison with Regularization Methods'''<br />
<br />
For these experiments, there are a few modifications that were made to assist with training. For ResNeXt, the EraseRelu formulation has each residual block ends in batch normalization. The Wide ResNet also is compared between vanilla with batch normalization and without. Batch normalization keeps the outputs of residual blocks in a certain range, as otherwise <math>\alpha</math> and <math>\beta</math> could cause perturbations that are too large, causing divergent learning. There is also a comparison of ResDrop/ShakeDrop Type A (where the regularization unit is inserted before the add unit for a residual branch) and after (where the regularization unit is inserted after the add unit for a residual branch). <br />
<br />
These experiments are performed on the CIFAR-100 dataset.<br />
<br />
[[File:ShakeDropArchitectureComparison1.png|800px|centre|thumb|]]<br />
<br />
[[File:ShakeDropArchitectureComparison2.png|800px|centre|thumb|]]<br />
<br />
[[File:ShakeDropArchitectureComparison3.png|800px|centre|thumb|]]<br />
<br />
For a final round of testing, the training setup was modified to incorporate other techniques used in state of the art methods. For most of the tests, the learning rate for the 300 epoch version started at 0.1 and decayed by a factor of 0.1 1/2 & 3/4 of the way through training. The alternative was cosine annealing, based on the presentation by Loshchilov and Hutter in their paper SGDR: Stochastic Gradient Descent with Warm Restarts. This is indicated in the Cos column, with a check indicating cosine annealing. <br />
<br />
[[File:CosineAnnealing.png|400px|centre|thumb|]]<br />
<br />
The Reg column indicates the regularization method used, either none, ResDrop (RD), Shake-Shake (SS), or ShakeDrop (SD). Fianlly, the Fil Column determines the type of data augmentation used, either none, cutout (CO) (DeVries & Taylor, 2017b), or Random Erasing (RE) (Zhong et al., 2017). <br />
<br />
[[File:ShakeDropComparison.png|800px|centre|thumb|Top-1 Errors (%) at final epoch on CIFAR-10/100 datasets]]<br />
<br />
'''State-of-the-Art Comparisons'''<br />
<br />
A direct comparison with state of the art methods is favorable for this new method. <br />
<br />
# Fair comparison of ResNeXt + Shake-Shake with PyramidNet + ShakeDrop gives an improvement of 0.19% on CIFAR-10 and 1.86% on CIFAR-100. Under these conditions, the final error rate is then 2.67% for CIFAR-10 and 13.99% for CIFAR-100.<br />
# Fair comparison of ResNeXt + Shake-Shake + Cutout with PyramidNet + ShakeDrop + Random Erasing gives an improvement of 0.25% on CIFAR-10 and 3.01% on CIFAR 100. Under these conditions, the final error rate is then 2.31% for CIFAR-10 and 12.19% for CIFAR-100.<br />
<br />
=Conclusion=<br />
<br />
This paper proposed a new stochastic regularization method, ShakeDrop, which outperforms previous state of the art methods while maintaining similar memory efficiency. It demonstrates that heavily perturbing a network can help to overcome issues with overfitting. It is also an effective way to regularize residual networks for image classification. The method was tested by CIFAR-10/100 and Tiny ImageNet datasets and showed great performance.<br />
<br />
=References=<br />
[Yamada et al., 2018] Yamada Y, Iwamura M, Kise K. ShakeDrop regularization. arXiv preprint arXiv:1802.02375. 2018 Feb 7.<br />
<br />
[He et al., 2016] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. CVPR, 2016.<br />
<br />
[Zagoruyko & Komodakis, 2016] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Proc. BMVC, 2016.<br />
<br />
[Han et al., 2017] Dongyoon Han, Jiwhan Kim, and Junmo Kim. Deep pyramidal residual networks. In Proc. CVPR, 2017a.<br />
<br />
[Xie et al., 2017] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proc. CVPR, 2017.<br />
<br />
[Huang et al., 2016] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Weinberger. Deep networks with stochastic depth. arXiv preprint arXiv:1603.09382v3, 2016.<br />
<br />
[Gastaldi, 2017] Xavier Gastaldi. Shake-shake regularization. arXiv preprint arXiv:1705.07485v2, 2017.<br />
<br />
[Loshilov & Hutter, 2016] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016.<br />
<br />
[DeVries & Taylor, 2017b] Terrance DeVries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017b.<br />
<br />
[Zhong et al., 2017] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. arXiv preprint arXiv:1708.04896, 2017.<br />
<br />
[Dutt et al., 2017] Anuvabh Dutt, Denis Pellerin, and Georges Qunot. Coupled ensembles of neural networks. arXiv preprint 1709.06053v1, 2017.<br />
<br />
[Veit et al., 2016] Andreas Veit, Michael J Wilber, and Serge Belongie. Residual networks behave like ensembles of relatively shallow networks. Advances in Neural Information Processing Systems 29, 2016.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=stat946F18/Autoregressive_Convolutional_Neural_Networks_for_Asynchronous_Time_Series&diff=41758stat946F18/Autoregressive Convolutional Neural Networks for Asynchronous Time Series2018-11-28T20:43:08Z<p>Ka2khan: /* Model Architecture */</p>
<hr />
<div>This page is a summary of the paper "[http://proceedings.mlr.press/v80/binkowski18a/binkowski18a.pdf Autoregressive Convolutional Neural Networks for Asynchronous Time Series]" by Mikołaj Binkowski, Gautier Marti, Philippe Donnat. It was published at ICML in 2018. The code for this paper is provided [https://github.com/mbinkowski/nntimeseries here].<br />
<br />
=Introduction=<br />
In this paper, the authors propose a deep convolutional network architecture called Significance-Offset Convolutional Neural Network for regression of multivariate asynchronous time series. The model is inspired by standard autoregressive(AR) models and gating systems used in recurrent neural networks. The model is evaluated on various time series data including:<br />
# Hedge fund proprietary dataset of over 2 million quotes for a credit derivative index, <br />
# An artificially generated noisy auto-regressive series, <br />
# A UCI household electricity consumption dataset. <br />
<br />
This paper focuses on time series with multi-variate and noisy signals, especially financial data. Financial time series is challenging to predict due to their low signal-to-noise ratio and heavy-tailed distributions. For example, the same signal (e.g. price of a stock) is obtained from different sources (e.g. financial news, an investment bank, financial analyst etc.) asynchronously. Each source may have a different bias or noise. (Figure 1) The investment bank with more clients can update their information more precisely than the investment bank with fewer clients, then the significance of each past observations may depend on other factors that change in time. Therefore, the traditional econometric models such as AR, VAR, VARMA[1] might not be sufficient. However, their relatively good performance could allow us to combine such linear econometric models with deep neural networks that can learn highly nonlinear relationships. This model is inspired by the gating mechanism which is successful in RNNs and Highway Networks.<br />
<br />
The time series forecasting problem can be expressed as a conditional probability distribution below,<br />
<div style="text-align: center;"><math>p(X_{t+d}|X_t,X_{t-1},...) = f(X_t,X_{t-1},...)</math></div><br />
Thus, we focus on modeling the predictors of future values of time series given their past values. <br />
The predictability of financial dataset still remains an open problem and is discussed in various publications [2].<br />
<br />
[[File:Junyi1.png | 500px|thumb|center|Figure 1: Quotes from four different market participants (sources) for the same credit default swaps (CDS) throughout one day. Each trader displays from time to time the prices for which he offers to buy (bid) and sell (ask) the underlying CDS. The filled area marks the difference between the best sell and buy offers (spread) at each time.]]<br />
<br />
The paper also provides empirical evidence that their model which combines linear models with deep learning models could perform better than just DL models like CNN, LSTMs and Phased LSTMs.<br />
<br />
=Related Work=<br />
===Time series forecasting===<br />
From recent proceedings in main machine learning venues i.e. ICML, NIPS, AISTATS, UAI, we can notice that time series are often forecast using Gaussian processes[3,4], especially for irregularly sampled time series[5]. Though still largely independent, combined models have started to appear, for example, the Gaussian Copula Process Volatility model[6]. For this paper, the authors use coupling AR models and neural networks to achieve such combined models.<br />
<br />
Although deep neural networks have been applied into many fields and produced satisfactory results, there still is little literature on deep learning for time series forecasting. More recently, the papers include Sirignano (2016)[7] that used 4-layer perceptrons in modeling price change distributions in Limit Order Books, and Borovykh et al. (2017)[8] who applied more recent WaveNet architecture to several short univariate and bivariate time-series (including financial ones). Heaton et al. (2016)[9] claimed to use autoencoders with a single hidden layer to compress multivariate financial data. Neil et al. (2016)[10] presented augmentation of LSTM architecture suitable for asynchronous series, which stimulates learning dependencies of different frequencies through time gate. <br />
<br />
In this paper, the authors examine the capabilities of several architectures (CNN, residual network, multi-layer LSTM, and phase LSTM) on AR-like artificial asynchronous and noisy time series, household electricity consumption dataset, and on real financial data from the credit default swap market with some inefficiencies.<br />
<br />
====AR Model====<br />
<br />
An autoregressive (AR) model describes the next value in a time-series as a combination of previous values, scaling factors, a bias, and noise [https://onlinecourses.science.psu.edu/stat501/node/358/ (source)]. For a p-th order (relating the current state to the p last states), the equation of the model is:<br />
<br />
<math> X_t = c + \sum_{i=1}^p \varphi_i X_{t-i}+ \varepsilon_t \,</math> [https://en.wikipedia.org/wiki/Autoregressive_model#Definition (equation source)]<br />
<br />
With parameters/coefficients <math>\varphi_i</math>, constant <math>c</math>, and noise <math>\varepsilon_t</math> This can be extended to vector form to create the VAR model mentioned in the paper.<br />
<br />
===Gating and weighting mechanisms===<br />
Gating mechanisms for neural networks has ability to overcome the problem of vanishing gradient, and can be expressed as <math display="inline">f(x)=c(x) \otimes \sigma(x)</math>, where <math>f</math> is the output function, <math>c</math> is a "candidate output" (a nonlinear function of <math>x</math>), <math>\otimes</math> is an element-wise matrix product, and <math>\sigma : \mathbb{R} \rightarrow [0,1] </math> is a sigmoid nonlinearity that controls the amount of output passed to the next layer. Different composition of functions of the same type as described above have proven to be an essential ingredient in popular recurrent architecture such as LSTM and GRU[11].<br />
<br />
The main purpose of the proposed gating system is to weight the outputs of the intermediate layers within neural networks, and is most closely related to softmax gating used in MuFuRu(Multi-Function Recurrent Unit)[12], i.e.<br />
<math display="inline"> f(x) = \sum_{l=1}^L p^l(x) \otimes f^l(x)\text{,}\ p(x)=\text{softmax}(\widehat{p}(x)), </math>, where <math>(f^l)_{l=1}^L </math>are candidate outputs (composition operators in MuFuRu), <math>(\widehat{p}^l)_{l=1}^L </math>are linear functions of inputs. <br />
<br />
This idea is also successfully used in attention networks[13] such as image captioning and machine translation. In this paper, the proposed method is similar as the separate inputs (time series steps in this case) are weighted in accordance with learned functions of these inputs. The difference is that the functions are being modeled using multi-layer CNNs. Another difference is that the proposed method is not using recurrent layers, which enables the network to remember parts of the sentence/image already translated/described.<br />
<br />
=Motivation=<br />
There are mainly five motivations that are stated in the paper by the authors:<br />
#The forecasting problem in this paper has been done almost independently by econometrics and machine learning communities. Unlike in machine learning, research in econometrics is more likely to explain variables rather than improving out-of-sample prediction power. These models tend to 'over-fit' on financial time series, their parameters are unstable and have poor performance on out-of-sample prediction.<br />
#It is difficult for the learning algorithms to deal with time series data where the observations have been made irregularly. Although Gaussian processes provide a useful theoretical framework that is able to handle asynchronous data, they are not suitable for financial datasets, which often follow heavy-tailed distribution .<br />
#Predictions of autoregressive time series may involve highly nonlinear functions if sampled irregularly. For AR time series with higher order and have more past observations, the expectation of it <math display="inline">\mathbb{E}[X(t)|{X(t-m), m=1,...,M}]</math> may involve more complicated functions that in general may not allow closed-form expression.<br />
#In practice, the dimensions of multivariate time series are often observed separately and asynchronously, such series at fixed frequency may lead to lose information or enlarge the dataset, which is shown in Figure 2(a). Therefore, the core of the proposed architecture SOCNN represents separate dimensions as a single one with dimension and duration indicators as additional features(Figure 2(b)).<br />
#Given a series of pairs of consecutive input values and corresponding durations, <math display="inline"> x_n = (X(t_n),t_n-t_{n-1}) </math>. One may expect that LSTM may memorize the input values in each step and weight them at the output according to the duration, but this approach may lead to an imbalance between the needs for memory and for linearity. The weights that are assigned to the memorized observations potentially require several layers of nonlinearity to be computed properly, while past observations might just need to be memorized as they are.<br />
<br />
[[File:Junyi2.png | 550px|thumb|center|Figure 2: (a) Fixed sampling frequency and its drawbacks; keep- ing all available information leads to much more datapoints. (b) Proposed data representation for the asynchronous series. Consecutive observations are stored together as a single value series, regardless of which series they belong to; this information, however, is stored in indicator features, alongside durations between observations.]]<br />
<br />
=Model Architecture=<br />
Suppose there exists a multivariate time series <math display="inline">(x_n)_{n=0}^{\infty} \subset \mathbb{R}^d </math>, we want to predict the conditional future values of a subset of elements of <math>x_n</math><br />
<div style="text-align: center;"><math>y_n = \mathbb{E} [x_n^I | \{x_{n-m}, m=1,2,...\}], </math></div><br />
where <math> I=\{i_1,i_2,...i_{d_I}\} \subset \{1,2,...,d\} </math> is a subset of features of <math>x_n</math>.<br />
<br />
Let <math> \textbf{x}_n^{-M} = (x_{n-m})_{m=1}^M </math>. <br />
<br />
The estimator of <math>y_n</math> can be expressed as:<br />
<div style="text-align: center;"><math>\hat{y}_n = \sum_{m=1}^M [F(\textbf{x}_n^{-M}) \otimes \sigma(S(\textbf{x}_n^{-M}))].,_m ,</math></div><br />
The estimate is the summation of the columns of the matrix in bracket. Here<br />
#<math>F,S : \mathbb{R}^{d \times M} \rightarrow \mathbb{R}^{d_I \times M}</math> are neural networks. <br />
#* <math>S</math> is a fully convolutional network which is composed of convolutional layers only. <br />
#* <math display="inline">F(\textbf{x}_n^{-M}) = W \otimes [\text{off}(x_{n-m}) + x_{n-m}^I)]_{m=1}^M </math> <br />
#** <math> W \in \mathbb{R}^{d_I \times M}</math> <br />
#** <math> \text{off}: \mathbb{R}^d \rightarrow \mathbb{R}^{d_I} </math> is a multilayer perceptron.<br />
<br />
#<math>\sigma</math> is a normalized activation function independent at each row, i.e. <math display="inline"> \sigma ((a_1^T, ..., a_{d_I}^T)^T)=(\sigma(a_1)^T,..., \sigma(a_{d_I})^T)^T </math><br />
#* for any <math>a_{i} \in \mathbb{R}^{M}</math><br />
#* and <math>\sigma </math> is defined such that <math>\sigma(a)^{T} \mathbf{1}_{M}=1</math> for any <math>a \in \mathbb{R}^M</math>.<br />
# <math>\otimes</math> is element-wise matrix multiplication (also known as Hadamard matrix multiplication).<br />
#<math>A.,_m</math> denotes the m-th column of a matrix A.<br />
<br />
Since <math>\sum_{m=1}^M W.,_m=W\cdot(1,1,...,1)^T</math> and <math>\sum_{m=1}^M S.,_m=S\cdot(1,1,...,1)^T</math>, we can express <math>\hat{y}_n</math> as:<br />
<div style="text-align: center;"><math>\hat{y}_n = \sum_{m=1}^M W.,_m \otimes (off(x_{n-m}) + x_{n-m}^I) \otimes \sigma(S.,_m(\textbf{x}_n^{-M}))</math></div><br />
This is the proposed network, Significance-Offset Convolutional Neural Network, <math>\text{off}</math> and <math>S</math> in the equation are corresponding to Offset and Significance in the name respectively.<br />
Figure 3 shows the scheme of network.<br />
<br />
[[File:Junyi3.png | 600px|thumb|center|Figure 3: A scheme of the proposed SOCNN architecture. The network preserves the time-dimension up to the top layer, while the number of features per timestep (filters) in the hidden layers is custom. The last convolutional layer, however, has the number of filters equal to dimension of the output. The Weighting frame shows how outputs from offset and significance networks are combined in accordance with Eq. of <math>\hat{y}_n</math>.]]<br />
<br />
The form of <math>\hat{y}_n</math> ensures the separation of the temporal dependence (obtained in weights <math>W_m</math>). <math>S</math>, which represents the local significance of observations, is determined by its filters which capture local dependencies and are independent of the relative position in time, and the predictors <math>\text{off}(x_{n-m})</math> are completely independent of position in time. An adjusted single regressor for the target variable is provided by each past observation through the offset network. Since in asynchronous sampling procedure, consecutive values of x come from different signals and might be heterogeneous, therefore adjustment of offset network is important. In addition, significance network provides data-dependent weight for each regressor and sums them up in an autoregressive manner.<br />
<br />
===Relation to asynchronous data===<br />
One common problem of time series is that durations are varying between consecutive observations, the paper states two ways to solve this problem<br />
#Data preprocessing: aligning the observations at some fixed frequency e.g. duplicating and interpolating observations as shown in Figure 2(a). However, as mentioned in the figure, this approach will tend to loss of information and enlarge the size of the dataset and model complexity.<br />
#Add additional features: Treating the duration or time of the observations as additional features, it is the core of SOCNN, which is shown in Figure 2(b).<br />
<br />
===Loss function===<br />
The output of the offset network is series of separate predictors of changes between corresponding observations <math>x_{n-m}^I</math> and the target value<math>y_n</math>, this is the reason why we use auxiliary loss function, which equals to mean squared error of such intermediate predictions:<br />
<div style="text-align: center;"><math>L^{aux}(\textbf{x}_n^{-M}, y_n)=\frac{1}{M} \sum_{m=1}^M ||off(x_{n-m}) + x_{n-m}^I -y_n||^2 </math></div><br />
The total loss for the sample <math> \textbf{x}_n^{-M},y_n) </math> is then given by:<br />
<div style="text-align: center;"><math>L^{tot}(\textbf{x}_n^{-M}, y_n)=L^2(\widehat{y}_n, y_n)+\alpha L^{aux}(\textbf{x}_n^{-M}, y_n)</math></div><br />
where <math>\widehat{y}_n</math> was mentioned before, <math>\alpha \geq 0</math> is a constant.<br />
<br />
=Experiments=<br />
The paper evaluated SOCNN architecture on three datasets: artificially generated datasets, [https://archive.ics.uci.edu/ml/datasets/Individual+household+electric+power+consumption household electric power consumption dataset], and the financial dataset of bid/ask quotes sent by several market participants active in the credit derivatives market. Comparing its performance with simple CNN, single and multiplayer LSTM and 25-layer ResNet. Apart from the evaluation of the SOCNN architecture the paper also discusses the impact of network components such as: such as auxiliary<br />
loss and the depth of the offset sub-network. The code and datasets are available [https://github.com/mbinkowski/nntimeseries here]<br />
<br />
==Datasets==<br />
Artificial data: They generated 4 artificial series, <math> X_{K \times N}</math>, where <math>K \in \{16,64\} </math>. Therefore there is a synchronous and an asynchronous series for each K value.<br />
<br />
Electricity data: This UCI dataset contains 7 different features excluding date and time. The features include global active power, global reactive power, voltage, global intensity, sub-metering 1, sub-metering 2 and sub-metering 3, recorded every minute for 47 months. The data has been altered so that one observation contains only one value of 7 features, while durations between consecutive observations are ranged from 1 to 7 minutes. The goal is to predict all 7 features for the next time step.<br />
<br />
Non-anonymous quotes: The dataset contains 2.1 million quotes from 28 different sources from different market participants such as analysts, banks etc. Each quote is characterized by 31 features: the offered price, 28 indicators of the quoting source, the direction indicator (the quote refers to either a buy or a sell offer) and duration from the previous quote. For each source and direction, we want to predict the next quoted price from this given source and direction considering the last 60 quotes.<br />
<br />
==Training details==<br />
They applied grid search on some hyperparameters in order to get the significance of its components. The hyperparameters include the offset sub-network's depth and the auxiliary weight <math>\alpha</math>. For offset sub-network's depth, they use 1, 10,1 for artificial, electricity and quotes dataset respectively; and they compared the values of <math>\alpha</math> in {0,0.1,0.01}.<br />
<br />
They chose LeakyReLU as activation function for all networks:<br />
<div style="text-align: center;"><math>\sigma^{LeakyReLU}(x) = x</math> if <math>x\geq 0</math>, and <math>0.1x</math> otherwise </div><br />
They use the same number of layers, same stride and similar kernel size structure in CNN. In each trained CNN, they applied max pooling with the pool size of 2 every 2 convolutional layers.<br />
<br />
Table 1 presents the configuration of network hyperparameters used in comparison<br />
<br />
[[File:Junyi4.png | 400px|center|]]<br />
<br />
===Network Training===<br />
The training and validation data were sampled randomly from the first 80% of timesteps in each series, with ratio of 3 to 1. The remaining 20% of data was used as a test set.<br />
<br />
All models were trained using Adam optimizer because the authors found that its rate of convergence was much faster than standard Stochastic Gradient Descent in early tests.<br />
<br />
They used a batch size of 128 for artificial and electricity data, and 256 for quotes dataset, and applied batch normalization between each convolution and the following activation. <br />
<br />
At the beginning of each epoch, the training samples were randomly sampled. To prevent overfitting, they applied dropout and early stopping.<br />
<br />
Weights were initialized using the normalized uniform procedure proposed by Glorot & Bengio (2010).[14]<br />
<br />
The authors carried out the experiments on Tensorflow and Keras and used different GPU to optimize the model for different datasets.<br />
<br />
==Results==<br />
Table 2 shows all results performed from all datasets.<br />
[[File:Junyi5.png | 600px|center|]]<br />
We can see that SOCNN outperforms in all asynchronous artificial, electricity and quotes datasets. For synchronous data, LSTM might be slightly better, but SOCNN almost has the same results with LSTM. Phased LSTM and ResNet have performed really bad on artificial asynchronous dataset and quotes dataset respectively. Notice that having more than one layer of offset network would have negative impact on results. Also, the higher weights of auxiliary loss(<math>\alpha</math>considerably improved the test error on asynchronous dataset, see Table 3. However, for other datasets, its impact was negligible.<br />
[[File:Junyi6.png | 400px|center|]]<br />
In general, SOCNN has significantly lower variance of the test and validation errors, especially in the early stage of the training process and for quotes dataset. This effect can be seen in the learning curves for Asynchronous 64 artificial dataset presented in Figure 5.<br />
[[File:Junyi7.png | 500px|thumb|center|Figure 5: Learning curves with different auxiliary weights for SOCNN model trained on Asynchronous 64 dataset. The solid lines indicate the test error while the dashed lines indicate the training error.]]<br />
<br />
Finally, we want to test the robustness of the proposed model SOCNN, adding noise terms to asynchronous 16 dataset and check how these networks perform. The result is shown in Figure 6.<br />
[[File:Junyi8.png | 600px|thumb|center|Figure 6: Experiment comparing robustness of the considered networks for Asynchronous 16 dataset. The plots show how the error would change if an additional noise term was added to the input series. The dotted curves show the total significance and average absolute offset (not to scale) outputs for the noisy observations. Interestingly, the significance of the noisy observations increases with the magnitude of noise; i.e. noisy observations are far from being discarded by SOCNN.]]<br />
From Figure 6, the purple line and green line seems staying at the same position in training and testing process. SOCNN and single-layer LSTM are most robust compared to other networks, and least prone to overfitting.<br />
<br />
=Conclusion and Discussion=<br />
In this paper, the authors have proposed a new architecture called Significance-Offset Convolutional Neural Network, which combines AR-like weighting mechanism and convolutional neural network. This new architecture is designed for high-noise asynchronous time series and achieves outperformance in forecasting several asynchronous time series compared to popular convolutional and recurrent networks. <br />
<br />
The SOCNN can be extended further by adding intermediate weighting layers of the same type in the network structure. Another possible extension but needs further empirical studies is that we consider not just <math>1 \times 1</math> convolutional kernels on the offset sub-network. Also, this new architecture might be tested on other real-life datasets with relevant characteristics in the future, especially on econometric datasets and more generally for time series (stochastic processes) regression.<br />
<br />
=Critiques=<br />
#The paper is most likely an application paper, and the proposed new architecture shows improved performance over baselines in the asynchronous time series.<br />
#The quote data cannot be reached, only two datasets available.<br />
#The 'Significance' network was described as critical to the model in paper, but they did not show how the performance of SOCNN with respect to the significance network.<br />
#The transform of the original data to asynchronous data is not clear.<br />
#The experiments on the main application are not reproducible because the data is proprietary.<br />
#The way that train and test data were split is unclear. This could be important in the case of the financial data set.<br />
#Although the auxiliary loss function was mentioned as an important part, the advantages of it was not too clear in the paper. Maybe it is better that the paper describes a little more about its effectiveness.<br />
#It was not mentioned clearly in the paper whether the model training was done on a rolling basis for time series forecasting.<br />
#The noise term used in section 5's model robustness analysis uses evenly distributed noise (see Appendix B). While the analysis is a good start, analysis with different noise distributions would make the findings more generalizable.<br />
#The paper uses financial/economic data as one of its testing data set. Instead of comparing neural network models such as CNN which is known to work badly on time series data, it would be much better if the author compared to well-known econometric time series models such as GARCH and VAR.<br />
<br />
=References=<br />
[1] Hamilton, J. D. Time series analysis, volume 2. Princeton university press Princeton, 1994. <br />
<br />
[2] Fama, E. F. Efficient capital markets: A review of theory and empirical work. The journal of Finance, 25(2):383–417, 1970.<br />
<br />
[3] Petelin, D., Sˇindela ́ˇr, J., Pˇrikryl, J., and Kocijan, J. Financial modeling using gaussian process models. In Intelligent Data Acquisition and Advanced Computing Systems (IDAACS), 2011 IEEE 6th International Conference on, volume 2, pp. 672–677. IEEE, 2011.<br />
<br />
[4] Tobar, F., Bui, T. D., and Turner, R. E. Learning stationary time series using gaussian processes with nonparametric kernels. In Advances in Neural Information Processing Systems, pp. 3501–3509, 2015.<br />
<br />
[5] Hwang, Y., Tong, A., and Choi, J. Automatic construction of nonparametric relational regression models for multiple time series. In Proceedings of the 33rd International Conference on Machine Learning, 2016.<br />
<br />
[6] Wilson, A. and Ghahramani, Z. Copula processes. In Advances in Neural Information Processing Systems, pp. 2460–2468, 2010.<br />
<br />
[7] Sirignano, J. Extended abstract: Neural networks for limit order books, February 2016.<br />
<br />
[8] Borovykh, A., Bohte, S., and Oosterlee, C. W. Condi- tional time series forecasting with convolutional neural networks, March 2017.<br />
<br />
[9] Heaton, J. B., Polson, N. G., and Witte, J. H. Deep learn- ing in finance, February 2016.<br />
<br />
[10] Neil, D., Pfeiffer, M., and Liu, S.-C. Phased lstm: Acceler- ating recurrent network training for long or event-based sequences. In Advances In Neural Information Process- ing Systems, pp. 3882–3890, 2016.<br />
<br />
[11] Chung, J., Gulcehre, C., Cho, K., and Bengio, Y. Em- pirical evaluation of gated recurrent neural networks on sequence modeling, December 2014.<br />
<br />
[12] Weissenborn, D. and Rockta ̈schel, T. MuFuRU: The Multi-Function recurrent unit, June 2016.<br />
<br />
[13] Cho, K., Courville, A., and Bengio, Y. Describing multi- media content using attention-based Encoder–Decoder networks. IEEE Transactions on Multimedia, 17(11): 1875–1886, July 2015. ISSN 1520-9210.<br />
<br />
[14] Glorot, X. and Bengio, Y. Understanding the dif- ficulty of training deep feedforward neural net- works. In In Proceedings of the International Con- ference on Artificial Intelligence and Statistics (AIS- TATSaˆ10). Society for Artificial Intelligence and Statistics, 2010.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=stat946F18/Autoregressive_Convolutional_Neural_Networks_for_Asynchronous_Time_Series&diff=41757stat946F18/Autoregressive Convolutional Neural Networks for Asynchronous Time Series2018-11-28T20:35:35Z<p>Ka2khan: /* Model Architecture */</p>
<hr />
<div>This page is a summary of the paper "[http://proceedings.mlr.press/v80/binkowski18a/binkowski18a.pdf Autoregressive Convolutional Neural Networks for Asynchronous Time Series]" by Mikołaj Binkowski, Gautier Marti, Philippe Donnat. It was published at ICML in 2018. The code for this paper is provided [https://github.com/mbinkowski/nntimeseries here].<br />
<br />
=Introduction=<br />
In this paper, the authors propose a deep convolutional network architecture called Significance-Offset Convolutional Neural Network for regression of multivariate asynchronous time series. The model is inspired by standard autoregressive(AR) models and gating systems used in recurrent neural networks. The model is evaluated on various time series data including:<br />
# Hedge fund proprietary dataset of over 2 million quotes for a credit derivative index, <br />
# An artificially generated noisy auto-regressive series, <br />
# A UCI household electricity consumption dataset. <br />
<br />
This paper focuses on time series with multi-variate and noisy signals, especially financial data. Financial time series is challenging to predict due to their low signal-to-noise ratio and heavy-tailed distributions. For example, the same signal (e.g. price of a stock) is obtained from different sources (e.g. financial news, an investment bank, financial analyst etc.) asynchronously. Each source may have a different bias or noise. (Figure 1) The investment bank with more clients can update their information more precisely than the investment bank with fewer clients, then the significance of each past observations may depend on other factors that change in time. Therefore, the traditional econometric models such as AR, VAR, VARMA[1] might not be sufficient. However, their relatively good performance could allow us to combine such linear econometric models with deep neural networks that can learn highly nonlinear relationships. This model is inspired by the gating mechanism which is successful in RNNs and Highway Networks.<br />
<br />
The time series forecasting problem can be expressed as a conditional probability distribution below,<br />
<div style="text-align: center;"><math>p(X_{t+d}|X_t,X_{t-1},...) = f(X_t,X_{t-1},...)</math></div><br />
Thus, we focus on modeling the predictors of future values of time series given their past values. <br />
The predictability of financial dataset still remains an open problem and is discussed in various publications [2].<br />
<br />
[[File:Junyi1.png | 500px|thumb|center|Figure 1: Quotes from four different market participants (sources) for the same credit default swaps (CDS) throughout one day. Each trader displays from time to time the prices for which he offers to buy (bid) and sell (ask) the underlying CDS. The filled area marks the difference between the best sell and buy offers (spread) at each time.]]<br />
<br />
The paper also provides empirical evidence that their model which combines linear models with deep learning models could perform better than just DL models like CNN, LSTMs and Phased LSTMs.<br />
<br />
=Related Work=<br />
===Time series forecasting===<br />
From recent proceedings in main machine learning venues i.e. ICML, NIPS, AISTATS, UAI, we can notice that time series are often forecast using Gaussian processes[3,4], especially for irregularly sampled time series[5]. Though still largely independent, combined models have started to appear, for example, the Gaussian Copula Process Volatility model[6]. For this paper, the authors use coupling AR models and neural networks to achieve such combined models.<br />
<br />
Although deep neural networks have been applied into many fields and produced satisfactory results, there still is little literature on deep learning for time series forecasting. More recently, the papers include Sirignano (2016)[7] that used 4-layer perceptrons in modeling price change distributions in Limit Order Books, and Borovykh et al. (2017)[8] who applied more recent WaveNet architecture to several short univariate and bivariate time-series (including financial ones). Heaton et al. (2016)[9] claimed to use autoencoders with a single hidden layer to compress multivariate financial data. Neil et al. (2016)[10] presented augmentation of LSTM architecture suitable for asynchronous series, which stimulates learning dependencies of different frequencies through time gate. <br />
<br />
In this paper, the authors examine the capabilities of several architectures (CNN, residual network, multi-layer LSTM, and phase LSTM) on AR-like artificial asynchronous and noisy time series, household electricity consumption dataset, and on real financial data from the credit default swap market with some inefficiencies.<br />
<br />
====AR Model====<br />
<br />
An autoregressive (AR) model describes the next value in a time-series as a combination of previous values, scaling factors, a bias, and noise [https://onlinecourses.science.psu.edu/stat501/node/358/ (source)]. For a p-th order (relating the current state to the p last states), the equation of the model is:<br />
<br />
<math> X_t = c + \sum_{i=1}^p \varphi_i X_{t-i}+ \varepsilon_t \,</math> [https://en.wikipedia.org/wiki/Autoregressive_model#Definition (equation source)]<br />
<br />
With parameters/coefficients <math>\varphi_i</math>, constant <math>c</math>, and noise <math>\varepsilon_t</math> This can be extended to vector form to create the VAR model mentioned in the paper.<br />
<br />
===Gating and weighting mechanisms===<br />
Gating mechanisms for neural networks has ability to overcome the problem of vanishing gradient, and can be expressed as <math display="inline">f(x)=c(x) \otimes \sigma(x)</math>, where <math>f</math> is the output function, <math>c</math> is a "candidate output" (a nonlinear function of <math>x</math>), <math>\otimes</math> is an element-wise matrix product, and <math>\sigma : \mathbb{R} \rightarrow [0,1] </math> is a sigmoid nonlinearity that controls the amount of output passed to the next layer. Different composition of functions of the same type as described above have proven to be an essential ingredient in popular recurrent architecture such as LSTM and GRU[11].<br />
<br />
The main purpose of the proposed gating system is to weight the outputs of the intermediate layers within neural networks, and is most closely related to softmax gating used in MuFuRu(Multi-Function Recurrent Unit)[12], i.e.<br />
<math display="inline"> f(x) = \sum_{l=1}^L p^l(x) \otimes f^l(x)\text{,}\ p(x)=\text{softmax}(\widehat{p}(x)), </math>, where <math>(f^l)_{l=1}^L </math>are candidate outputs (composition operators in MuFuRu), <math>(\widehat{p}^l)_{l=1}^L </math>are linear functions of inputs. <br />
<br />
This idea is also successfully used in attention networks[13] such as image captioning and machine translation. In this paper, the proposed method is similar as the separate inputs (time series steps in this case) are weighted in accordance with learned functions of these inputs. The difference is that the functions are being modeled using multi-layer CNNs. Another difference is that the proposed method is not using recurrent layers, which enables the network to remember parts of the sentence/image already translated/described.<br />
<br />
=Motivation=<br />
There are mainly five motivations that are stated in the paper by the authors:<br />
#The forecasting problem in this paper has been done almost independently by econometrics and machine learning communities. Unlike in machine learning, research in econometrics is more likely to explain variables rather than improving out-of-sample prediction power. These models tend to 'over-fit' on financial time series, their parameters are unstable and have poor performance on out-of-sample prediction.<br />
#It is difficult for the learning algorithms to deal with time series data where the observations have been made irregularly. Although Gaussian processes provide a useful theoretical framework that is able to handle asynchronous data, they are not suitable for financial datasets, which often follow heavy-tailed distribution .<br />
#Predictions of autoregressive time series may involve highly nonlinear functions if sampled irregularly. For AR time series with higher order and have more past observations, the expectation of it <math display="inline">\mathbb{E}[X(t)|{X(t-m), m=1,...,M}]</math> may involve more complicated functions that in general may not allow closed-form expression.<br />
#In practice, the dimensions of multivariate time series are often observed separately and asynchronously, such series at fixed frequency may lead to lose information or enlarge the dataset, which is shown in Figure 2(a). Therefore, the core of the proposed architecture SOCNN represents separate dimensions as a single one with dimension and duration indicators as additional features(Figure 2(b)).<br />
#Given a series of pairs of consecutive input values and corresponding durations, <math display="inline"> x_n = (X(t_n),t_n-t_{n-1}) </math>. One may expect that LSTM may memorize the input values in each step and weight them at the output according to the duration, but this approach may lead to an imbalance between the needs for memory and for linearity. The weights that are assigned to the memorized observations potentially require several layers of nonlinearity to be computed properly, while past observations might just need to be memorized as they are.<br />
<br />
[[File:Junyi2.png | 550px|thumb|center|Figure 2: (a) Fixed sampling frequency and its drawbacks; keep- ing all available information leads to much more datapoints. (b) Proposed data representation for the asynchronous series. Consecutive observations are stored together as a single value series, regardless of which series they belong to; this information, however, is stored in indicator features, alongside durations between observations.]]<br />
<br />
=Model Architecture=<br />
Suppose there exists a multivariate time series <math display="inline">(x_n)_{n=0}^{\infty} \subset \mathbb{R}^d </math>, we want to predict the conditional future values of a subset of elements of <math>x_n</math><br />
<div style="text-align: center;"><math>y_n = \mathbb{E} [x_n^I | \{x_{n-m}, m=1,2,...\}], </math></div><br />
where <math> I=\{i_1,i_2,...i_{d_I}\} \subset \{1,2,...,d\} </math> is a subset of features of <math>x_n</math>.<br />
<br />
Let <math> \textbf{x}_n^{-M} = (x_{n-m})_{m=1}^M </math>. <br />
<br />
The estimator of <math>y_n</math> can be expressed as:<br />
<div style="text-align: center;"><math>\hat{y}_n = \sum_{m=1}^M [F(\textbf{x}_n^{-M}) \otimes \sigma(S(\textbf{x}_n^{-M}))].,_m ,</math></div><br />
The estimate is the summation of the columns of the matrix in bracket. Here<br />
#<math>F,S : \mathbb{R}^{d \times M} \rightarrow \mathbb{R}^{d_I \times M}</math> are neural networks. <br />
#* <math>S</math> is a fully convolutional network which is composed of convolutional layers only. <br />
#* <math display="inline">F(\textbf{x}_n^{-M}) = W \otimes [\text{off}(x_{n-m}) + x_{n-m}^I)]_{m=1}^M </math> <br />
#** <math> W \in \mathbb{R}^{d_I \times M}</math> <br />
#** <math> \text{off}: \mathbb{R}^d \rightarrow \mathbb{R}^{d_I} </math> is a multilayer perceptron.<br />
<br />
#<math>\sigma</math> is a normalized activation function independent at each row, i.e. <math display="inline"> \sigma ((a_1^T, ..., a_{d_I}^T)^T)=(\sigma(a_1)^T,..., \sigma(a_{d_I})^T)^T </math><br />
#* for any <math>a_{i} \in \mathbb{R}^{M}</math><br />
#* and <math>\sigma </math> is defined such that <math>\sigma(a)^{T} \mathbf{1}_{M}=1</math> for any <math>a \in \mathbb{R}^M</math>.<br />
# <math>\otimes</math> is element-wise matrix multiplication (also known as Hadamard matrix multiplication).<br />
#<math>A.,_m</math> denotes the m-th column of a matrix A.<br />
<br />
Since <math>\sum_{m=1}^M W.,_m=W\cdot(1,1,...,1)^T</math> and <math>\sum_{m=1}^M S.,_m=S\cdot(1,1,...,1)^T</math>, we can express <math>\hat{y}_n</math> as:<br />
<div style="text-align: center;"><math>\hat{y}_n = \sum_{m=1}^M W.,_m \otimes (off(x_{n-m}) + x_{n-m}^I) \otimes \sigma(S.,_m(\textbf{x}_n^{-M}))</math></div><br />
This is the proposed network, Significance-Offset Convolutional Neural Network, <math>\text{off}</math> and <math>S</math> in the equation are corresponding to Offset and Significance in the name respectively.<br />
Figure 3 shows the scheme of network.<br />
<br />
[[File:Junyi3.png | 600px|thumb|center|Figure 3: A scheme of the proposed SOCNN architecture. The network preserves the time-dimension up to the top layer, while the number of features per timestep (filters) in the hidden layers is custom. The last convolutional layer, however, has the number of filters equal to dimension of the output. The Weighting frame shows how outputs from offset and significance networks are combined in accordance with Eq. of <math>\hat{y}_n</math>.]]<br />
<br />
The form of <math>\hat{y}_n</math> forced to separate the temporal dependence (obtained in weights <math>W_m</math>). S is determined by its filters which capture local dependencies and are independent of the relative position in time, the predictors <math>off(x_{n-m})</math> are completely independent of position in time. An adjusted single regressor for the target variable is provided by each past observation through the offset network. Since in asynchronous sampling procedure, consecutive values of x come from different signals and might be heterogeneous, therefore adjustment of offset network is important. In addition, significance network provides data-dependent weight for each regressor and sums them up in an autoregressive manner.<br />
<br />
===Relation to asynchronous data===<br />
One common problem of time series is that durations are varying between consecutive observations, the paper states two ways to solve this problem<br />
#Data preprocessing: aligning the observations at some fixed frequency e.g. duplicating and interpolating observations as shown in Figure 2(a). However, as mentioned in the figure, this approach will tend to loss of information and enlarge the size of the dataset and model complexity.<br />
#Add additional features: Treating the duration or time of the observations as additional features, it is the core of SOCNN, which is shown in Figure 2(b).<br />
<br />
===Loss function===<br />
The output of the offset network is series of separate predictors of changes between corresponding observations <math>x_{n-m}^I</math> and the target value<math>y_n</math>, this is the reason why we use auxiliary loss function, which equals to mean squared error of such intermediate predictions:<br />
<div style="text-align: center;"><math>L^{aux}(\textbf{x}_n^{-M}, y_n)=\frac{1}{M} \sum_{m=1}^M ||off(x_{n-m}) + x_{n-m}^I -y_n||^2 </math></div><br />
The total loss for the sample <math> \textbf{x}_n^{-M},y_n) </math> is then given by:<br />
<div style="text-align: center;"><math>L^{tot}(\textbf{x}_n^{-M}, y_n)=L^2(\widehat{y}_n, y_n)+\alpha L^{aux}(\textbf{x}_n^{-M}, y_n)</math></div><br />
where <math>\widehat{y}_n</math> was mentioned before, <math>\alpha \geq 0</math> is a constant.<br />
<br />
=Experiments=<br />
The paper evaluated SOCNN architecture on three datasets: artificially generated datasets, [https://archive.ics.uci.edu/ml/datasets/Individual+household+electric+power+consumption household electric power consumption dataset], and the financial dataset of bid/ask quotes sent by several market participants active in the credit derivatives market. Comparing its performance with simple CNN, single and multiplayer LSTM and 25-layer ResNet. Apart from the evaluation of the SOCNN architecture the paper also discusses the impact of network components such as: such as auxiliary<br />
loss and the depth of the offset sub-network. The code and datasets are available [https://github.com/mbinkowski/nntimeseries here]<br />
<br />
==Datasets==<br />
Artificial data: They generated 4 artificial series, <math> X_{K \times N}</math>, where <math>K \in \{16,64\} </math>. Therefore there is a synchronous and an asynchronous series for each K value.<br />
<br />
Electricity data: This UCI dataset contains 7 different features excluding date and time. The features include global active power, global reactive power, voltage, global intensity, sub-metering 1, sub-metering 2 and sub-metering 3, recorded every minute for 47 months. The data has been altered so that one observation contains only one value of 7 features, while durations between consecutive observations are ranged from 1 to 7 minutes. The goal is to predict all 7 features for the next time step.<br />
<br />
Non-anonymous quotes: The dataset contains 2.1 million quotes from 28 different sources from different market participants such as analysts, banks etc. Each quote is characterized by 31 features: the offered price, 28 indicators of the quoting source, the direction indicator (the quote refers to either a buy or a sell offer) and duration from the previous quote. For each source and direction, we want to predict the next quoted price from this given source and direction considering the last 60 quotes.<br />
<br />
==Training details==<br />
They applied grid search on some hyperparameters in order to get the significance of its components. The hyperparameters include the offset sub-network's depth and the auxiliary weight <math>\alpha</math>. For offset sub-network's depth, they use 1, 10,1 for artificial, electricity and quotes dataset respectively; and they compared the values of <math>\alpha</math> in {0,0.1,0.01}.<br />
<br />
They chose LeakyReLU as activation function for all networks:<br />
<div style="text-align: center;"><math>\sigma^{LeakyReLU}(x) = x</math> if <math>x\geq 0</math>, and <math>0.1x</math> otherwise </div><br />
They use the same number of layers, same stride and similar kernel size structure in CNN. In each trained CNN, they applied max pooling with the pool size of 2 every 2 convolutional layers.<br />
<br />
Table 1 presents the configuration of network hyperparameters used in comparison<br />
<br />
[[File:Junyi4.png | 400px|center|]]<br />
<br />
===Network Training===<br />
The training and validation data were sampled randomly from the first 80% of timesteps in each series, with ratio of 3 to 1. The remaining 20% of data was used as a test set.<br />
<br />
All models were trained using Adam optimizer because the authors found that its rate of convergence was much faster than standard Stochastic Gradient Descent in early tests.<br />
<br />
They used a batch size of 128 for artificial and electricity data, and 256 for quotes dataset, and applied batch normalization between each convolution and the following activation. <br />
<br />
At the beginning of each epoch, the training samples were randomly sampled. To prevent overfitting, they applied dropout and early stopping.<br />
<br />
Weights were initialized using the normalized uniform procedure proposed by Glorot & Bengio (2010).[14]<br />
<br />
The authors carried out the experiments on Tensorflow and Keras and used different GPU to optimize the model for different datasets.<br />
<br />
==Results==<br />
Table 2 shows all results performed from all datasets.<br />
[[File:Junyi5.png | 600px|center|]]<br />
We can see that SOCNN outperforms in all asynchronous artificial, electricity and quotes datasets. For synchronous data, LSTM might be slightly better, but SOCNN almost has the same results with LSTM. Phased LSTM and ResNet have performed really bad on artificial asynchronous dataset and quotes dataset respectively. Notice that having more than one layer of offset network would have negative impact on results. Also, the higher weights of auxiliary loss(<math>\alpha</math>considerably improved the test error on asynchronous dataset, see Table 3. However, for other datasets, its impact was negligible.<br />
[[File:Junyi6.png | 400px|center|]]<br />
In general, SOCNN has significantly lower variance of the test and validation errors, especially in the early stage of the training process and for quotes dataset. This effect can be seen in the learning curves for Asynchronous 64 artificial dataset presented in Figure 5.<br />
[[File:Junyi7.png | 500px|thumb|center|Figure 5: Learning curves with different auxiliary weights for SOCNN model trained on Asynchronous 64 dataset. The solid lines indicate the test error while the dashed lines indicate the training error.]]<br />
<br />
Finally, we want to test the robustness of the proposed model SOCNN, adding noise terms to asynchronous 16 dataset and check how these networks perform. The result is shown in Figure 6.<br />
[[File:Junyi8.png | 600px|thumb|center|Figure 6: Experiment comparing robustness of the considered networks for Asynchronous 16 dataset. The plots show how the error would change if an additional noise term was added to the input series. The dotted curves show the total significance and average absolute offset (not to scale) outputs for the noisy observations. Interestingly, the significance of the noisy observations increases with the magnitude of noise; i.e. noisy observations are far from being discarded by SOCNN.]]<br />
From Figure 6, the purple line and green line seems staying at the same position in training and testing process. SOCNN and single-layer LSTM are most robust compared to other networks, and least prone to overfitting.<br />
<br />
=Conclusion and Discussion=<br />
In this paper, the authors have proposed a new architecture called Significance-Offset Convolutional Neural Network, which combines AR-like weighting mechanism and convolutional neural network. This new architecture is designed for high-noise asynchronous time series and achieves outperformance in forecasting several asynchronous time series compared to popular convolutional and recurrent networks. <br />
<br />
The SOCNN can be extended further by adding intermediate weighting layers of the same type in the network structure. Another possible extension but needs further empirical studies is that we consider not just <math>1 \times 1</math> convolutional kernels on the offset sub-network. Also, this new architecture might be tested on other real-life datasets with relevant characteristics in the future, especially on econometric datasets and more generally for time series (stochastic processes) regression.<br />
<br />
=Critiques=<br />
#The paper is most likely an application paper, and the proposed new architecture shows improved performance over baselines in the asynchronous time series.<br />
#The quote data cannot be reached, only two datasets available.<br />
#The 'Significance' network was described as critical to the model in paper, but they did not show how the performance of SOCNN with respect to the significance network.<br />
#The transform of the original data to asynchronous data is not clear.<br />
#The experiments on the main application are not reproducible because the data is proprietary.<br />
#The way that train and test data were split is unclear. This could be important in the case of the financial data set.<br />
#Although the auxiliary loss function was mentioned as an important part, the advantages of it was not too clear in the paper. Maybe it is better that the paper describes a little more about its effectiveness.<br />
#It was not mentioned clearly in the paper whether the model training was done on a rolling basis for time series forecasting.<br />
#The noise term used in section 5's model robustness analysis uses evenly distributed noise (see Appendix B). While the analysis is a good start, analysis with different noise distributions would make the findings more generalizable.<br />
#The paper uses financial/economic data as one of its testing data set. Instead of comparing neural network models such as CNN which is known to work badly on time series data, it would be much better if the author compared to well-known econometric time series models such as GARCH and VAR.<br />
<br />
=References=<br />
[1] Hamilton, J. D. Time series analysis, volume 2. Princeton university press Princeton, 1994. <br />
<br />
[2] Fama, E. F. Efficient capital markets: A review of theory and empirical work. The journal of Finance, 25(2):383–417, 1970.<br />
<br />
[3] Petelin, D., Sˇindela ́ˇr, J., Pˇrikryl, J., and Kocijan, J. Financial modeling using gaussian process models. In Intelligent Data Acquisition and Advanced Computing Systems (IDAACS), 2011 IEEE 6th International Conference on, volume 2, pp. 672–677. IEEE, 2011.<br />
<br />
[4] Tobar, F., Bui, T. D., and Turner, R. E. Learning stationary time series using gaussian processes with nonparametric kernels. In Advances in Neural Information Processing Systems, pp. 3501–3509, 2015.<br />
<br />
[5] Hwang, Y., Tong, A., and Choi, J. Automatic construction of nonparametric relational regression models for multiple time series. In Proceedings of the 33rd International Conference on Machine Learning, 2016.<br />
<br />
[6] Wilson, A. and Ghahramani, Z. Copula processes. In Advances in Neural Information Processing Systems, pp. 2460–2468, 2010.<br />
<br />
[7] Sirignano, J. Extended abstract: Neural networks for limit order books, February 2016.<br />
<br />
[8] Borovykh, A., Bohte, S., and Oosterlee, C. W. Condi- tional time series forecasting with convolutional neural networks, March 2017.<br />
<br />
[9] Heaton, J. B., Polson, N. G., and Witte, J. H. Deep learn- ing in finance, February 2016.<br />
<br />
[10] Neil, D., Pfeiffer, M., and Liu, S.-C. Phased lstm: Acceler- ating recurrent network training for long or event-based sequences. In Advances In Neural Information Process- ing Systems, pp. 3882–3890, 2016.<br />
<br />
[11] Chung, J., Gulcehre, C., Cho, K., and Bengio, Y. Em- pirical evaluation of gated recurrent neural networks on sequence modeling, December 2014.<br />
<br />
[12] Weissenborn, D. and Rockta ̈schel, T. MuFuRU: The Multi-Function recurrent unit, June 2016.<br />
<br />
[13] Cho, K., Courville, A., and Bengio, Y. Describing multi- media content using attention-based Encoder–Decoder networks. IEEE Transactions on Multimedia, 17(11): 1875–1886, July 2015. ISSN 1520-9210.<br />
<br />
[14] Glorot, X. and Bengio, Y. Understanding the dif- ficulty of training deep feedforward neural net- works. In In Proceedings of the International Con- ference on Artificial Intelligence and Statistics (AIS- TATSaˆ10). Society for Artificial Intelligence and Statistics, 2010.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=stat946F18/Autoregressive_Convolutional_Neural_Networks_for_Asynchronous_Time_Series&diff=41756stat946F18/Autoregressive Convolutional Neural Networks for Asynchronous Time Series2018-11-28T20:18:54Z<p>Ka2khan: /* Motivation */</p>
<hr />
<div>This page is a summary of the paper "[http://proceedings.mlr.press/v80/binkowski18a/binkowski18a.pdf Autoregressive Convolutional Neural Networks for Asynchronous Time Series]" by Mikołaj Binkowski, Gautier Marti, Philippe Donnat. It was published at ICML in 2018. The code for this paper is provided [https://github.com/mbinkowski/nntimeseries here].<br />
<br />
=Introduction=<br />
In this paper, the authors propose a deep convolutional network architecture called Significance-Offset Convolutional Neural Network for regression of multivariate asynchronous time series. The model is inspired by standard autoregressive(AR) models and gating systems used in recurrent neural networks. The model is evaluated on various time series data including:<br />
# Hedge fund proprietary dataset of over 2 million quotes for a credit derivative index, <br />
# An artificially generated noisy auto-regressive series, <br />
# A UCI household electricity consumption dataset. <br />
<br />
This paper focuses on time series with multi-variate and noisy signals, especially financial data. Financial time series is challenging to predict due to their low signal-to-noise ratio and heavy-tailed distributions. For example, the same signal (e.g. price of a stock) is obtained from different sources (e.g. financial news, an investment bank, financial analyst etc.) asynchronously. Each source may have a different bias or noise. (Figure 1) The investment bank with more clients can update their information more precisely than the investment bank with fewer clients, then the significance of each past observations may depend on other factors that change in time. Therefore, the traditional econometric models such as AR, VAR, VARMA[1] might not be sufficient. However, their relatively good performance could allow us to combine such linear econometric models with deep neural networks that can learn highly nonlinear relationships. This model is inspired by the gating mechanism which is successful in RNNs and Highway Networks.<br />
<br />
The time series forecasting problem can be expressed as a conditional probability distribution below,<br />
<div style="text-align: center;"><math>p(X_{t+d}|X_t,X_{t-1},...) = f(X_t,X_{t-1},...)</math></div><br />
Thus, we focus on modeling the predictors of future values of time series given their past values. <br />
The predictability of financial dataset still remains an open problem and is discussed in various publications [2].<br />
<br />
[[File:Junyi1.png | 500px|thumb|center|Figure 1: Quotes from four different market participants (sources) for the same credit default swaps (CDS) throughout one day. Each trader displays from time to time the prices for which he offers to buy (bid) and sell (ask) the underlying CDS. The filled area marks the difference between the best sell and buy offers (spread) at each time.]]<br />
<br />
The paper also provides empirical evidence that their model which combines linear models with deep learning models could perform better than just DL models like CNN, LSTMs and Phased LSTMs.<br />
<br />
=Related Work=<br />
===Time series forecasting===<br />
From recent proceedings in main machine learning venues i.e. ICML, NIPS, AISTATS, UAI, we can notice that time series are often forecast using Gaussian processes[3,4], especially for irregularly sampled time series[5]. Though still largely independent, combined models have started to appear, for example, the Gaussian Copula Process Volatility model[6]. For this paper, the authors use coupling AR models and neural networks to achieve such combined models.<br />
<br />
Although deep neural networks have been applied into many fields and produced satisfactory results, there still is little literature on deep learning for time series forecasting. More recently, the papers include Sirignano (2016)[7] that used 4-layer perceptrons in modeling price change distributions in Limit Order Books, and Borovykh et al. (2017)[8] who applied more recent WaveNet architecture to several short univariate and bivariate time-series (including financial ones). Heaton et al. (2016)[9] claimed to use autoencoders with a single hidden layer to compress multivariate financial data. Neil et al. (2016)[10] presented augmentation of LSTM architecture suitable for asynchronous series, which stimulates learning dependencies of different frequencies through time gate. <br />
<br />
In this paper, the authors examine the capabilities of several architectures (CNN, residual network, multi-layer LSTM, and phase LSTM) on AR-like artificial asynchronous and noisy time series, household electricity consumption dataset, and on real financial data from the credit default swap market with some inefficiencies.<br />
<br />
====AR Model====<br />
<br />
An autoregressive (AR) model describes the next value in a time-series as a combination of previous values, scaling factors, a bias, and noise [https://onlinecourses.science.psu.edu/stat501/node/358/ (source)]. For a p-th order (relating the current state to the p last states), the equation of the model is:<br />
<br />
<math> X_t = c + \sum_{i=1}^p \varphi_i X_{t-i}+ \varepsilon_t \,</math> [https://en.wikipedia.org/wiki/Autoregressive_model#Definition (equation source)]<br />
<br />
With parameters/coefficients <math>\varphi_i</math>, constant <math>c</math>, and noise <math>\varepsilon_t</math> This can be extended to vector form to create the VAR model mentioned in the paper.<br />
<br />
===Gating and weighting mechanisms===<br />
Gating mechanisms for neural networks has ability to overcome the problem of vanishing gradient, and can be expressed as <math display="inline">f(x)=c(x) \otimes \sigma(x)</math>, where <math>f</math> is the output function, <math>c</math> is a "candidate output" (a nonlinear function of <math>x</math>), <math>\otimes</math> is an element-wise matrix product, and <math>\sigma : \mathbb{R} \rightarrow [0,1] </math> is a sigmoid nonlinearity that controls the amount of output passed to the next layer. Different composition of functions of the same type as described above have proven to be an essential ingredient in popular recurrent architecture such as LSTM and GRU[11].<br />
<br />
The main purpose of the proposed gating system is to weight the outputs of the intermediate layers within neural networks, and is most closely related to softmax gating used in MuFuRu(Multi-Function Recurrent Unit)[12], i.e.<br />
<math display="inline"> f(x) = \sum_{l=1}^L p^l(x) \otimes f^l(x)\text{,}\ p(x)=\text{softmax}(\widehat{p}(x)), </math>, where <math>(f^l)_{l=1}^L </math>are candidate outputs (composition operators in MuFuRu), <math>(\widehat{p}^l)_{l=1}^L </math>are linear functions of inputs. <br />
<br />
This idea is also successfully used in attention networks[13] such as image captioning and machine translation. In this paper, the proposed method is similar as the separate inputs (time series steps in this case) are weighted in accordance with learned functions of these inputs. The difference is that the functions are being modeled using multi-layer CNNs. Another difference is that the proposed method is not using recurrent layers, which enables the network to remember parts of the sentence/image already translated/described.<br />
<br />
=Motivation=<br />
There are mainly five motivations that are stated in the paper by the authors:<br />
#The forecasting problem in this paper has been done almost independently by econometrics and machine learning communities. Unlike in machine learning, research in econometrics is more likely to explain variables rather than improving out-of-sample prediction power. These models tend to 'over-fit' on financial time series, their parameters are unstable and have poor performance on out-of-sample prediction.<br />
#It is difficult for the learning algorithms to deal with time series data where the observations have been made irregularly. Although Gaussian processes provide a useful theoretical framework that is able to handle asynchronous data, they are not suitable for financial datasets, which often follow heavy-tailed distribution .<br />
#Predictions of autoregressive time series may involve highly nonlinear functions if sampled irregularly. For AR time series with higher order and have more past observations, the expectation of it <math display="inline">\mathbb{E}[X(t)|{X(t-m), m=1,...,M}]</math> may involve more complicated functions that in general may not allow closed-form expression.<br />
#In practice, the dimensions of multivariate time series are often observed separately and asynchronously, such series at fixed frequency may lead to lose information or enlarge the dataset, which is shown in Figure 2(a). Therefore, the core of the proposed architecture SOCNN represents separate dimensions as a single one with dimension and duration indicators as additional features(Figure 2(b)).<br />
#Given a series of pairs of consecutive input values and corresponding durations, <math display="inline"> x_n = (X(t_n),t_n-t_{n-1}) </math>. One may expect that LSTM may memorize the input values in each step and weight them at the output according to the duration, but this approach may lead to an imbalance between the needs for memory and for linearity. The weights that are assigned to the memorized observations potentially require several layers of nonlinearity to be computed properly, while past observations might just need to be memorized as they are.<br />
<br />
[[File:Junyi2.png | 550px|thumb|center|Figure 2: (a) Fixed sampling frequency and its drawbacks; keep- ing all available information leads to much more datapoints. (b) Proposed data representation for the asynchronous series. Consecutive observations are stored together as a single value series, regardless of which series they belong to; this information, however, is stored in indicator features, alongside durations between observations.]]<br />
<br />
=Model Architecture=<br />
Suppose there's a multivariate time series <math display="inline">(x_n)_{n=0}^{\infty} \subset \mathbb{R}^d </math>, we want to predict the conditional future values of a subset of elements of <math>x_n</math><br />
<div style="text-align: center;"><math>y_n = \mathbb{E} [x_n^I | {x_{n-m}, m=1,2,...}], </math></div><br />
where <math> I=\{i_1,i_2,...i_{d_I}\} \subset \{1,2,...,d\} </math> is a subset of features of <math>x_n</math>.<br />
<br />
Let <math> \textbf{x}_n^{-M} = (x_{n-m})_{m=1}^M </math>. <br />
<br />
The estimator of <math>y_n</math> can be expressed as:<br />
<div style="text-align: center;"><math>\hat{y}_n = \sum_{m=1}^M [F(\textbf{x}_n^{-M}) \otimes \sigma(S(\textbf{x}_n^{-M}))].,_m ,</math></div><br />
The estimate is the summation of the columns of the matrix in bracket. Here<br />
#<math>F,S : \mathbb{R}^{d \times M} \rightarrow \mathbb{R}^{d_I \times M}</math> are neural networks. <br />
#* <math>S</math> is a fully convolutional network which is composed of convolutional layers only. <br />
#* <math display="inline">F(\textbf{x}_n^{-M}) = W \otimes [off(x_{n-m}) + x_{n-m}^I)]_{m=1}^M </math> <br />
#** <math> W \in \mathbb{R}^{d_I \times M}</math> <br />
#** <math> off: \mathbb{R}^d \rightarrow \mathbb{R}^{d_I} </math> is a multilayer perceptron.<br />
<br />
#<math>\sigma</math> is a normalized activation function independent at each row, i.e. <math display="inline"> \sigma ((a_1^T, ..., a_{d_I}^T)^T)=(\sigma(a_1)^T,..., \sigma(a_{d_I})^T)^T </math><br />
#* <math>a_{i} \in \mathbb{R}^{M}</math><br />
#* and <math>\sigma </math> is defined such that <math>\sigma(a)^{T} \mathbf{1}_{M}=1</math><br />
# <math>\otimes</math> is element-wise matrix multiplication.<br />
#<math>A.,_m</math> denotes the m-th column of a matrix A.<br />
<br />
Since <math>\sum_{m=1}^M W.,_m=W(1,1,...,1)^T</math> and <math>\sum_{m=1}^M S.,_m=S(1,1,...,1)^T</math>, we can express <math>\hat{y}_n</math> as:<br />
<div style="text-align: center;"><math>\hat{y}_n = \sum_{m=1}^M W.,_m \otimes (off(x_{n-m}) + x_{n-m}^I) \otimes \sigma(S.,_m(\textbf{x}_n^{-M}))</math></div><br />
This is the proposed network, Significance-Offset Convolutional Neural Network, <math>off</math> and <math>S</math> in the equation are corresponding to Offset and Significance in the name respectively.<br />
Figure 3 shows the scheme of network.<br />
<br />
[[File:Junyi3.png | 600px|thumb|center|Figure 3: A scheme of the proposed SOCNN architecture. The network preserves the time-dimension up to the top layer, while the number of features per timestep (filters) in the hidden layers is custom. The last convolutional layer, however, has the number of filters equal to dimension of the output. The Weighting frame shows how outputs from offset and significance networks are combined in accordance with Eq. of <math>\hat{y}_n</math>.]]<br />
<br />
The form of <math>\hat{y}_n</math> forced to separate the temporal dependence (obtained in weights <math>W_m</math>). S is determined by its filters which capture local dependencies and are independent of the relative position in time, the predictors <math>off(x_{n-m})</math> are completely independent of position in time. An adjusted single regressor for the target variable is provided by each past observation through the offset network. Since in asynchronous sampling procedure, consecutive values of x come from different signals and might be heterogeneous, therefore adjustment of offset network is important. In addition, significance network provides data-dependent weight for each regressor and sums them up in an autoregressive manner.<br />
<br />
===Relation to asynchronous data===<br />
One common problem of time series is that durations are varying between consecutive observations, the paper states two ways to solve this problem<br />
#Data preprocessing: aligning the observations at some fixed frequency e.g. duplicating and interpolating observations as shown in Figure 2(a). However, as mentioned in the figure, this approach will tend to loss of information and enlarge the size of the dataset and model complexity.<br />
#Add additional features: Treating the duration or time of the observations as additional features, it is the core of SOCNN, which is shown in Figure 2(b).<br />
<br />
===Loss function===<br />
The output of the offset network is series of separate predictors of changes between corresponding observations <math>x_{n-m}^I</math> and the target value<math>y_n</math>, this is the reason why we use auxiliary loss function, which equals to mean squared error of such intermediate predictions:<br />
<div style="text-align: center;"><math>L^{aux}(\textbf{x}_n^{-M}, y_n)=\frac{1}{M} \sum_{m=1}^M ||off(x_{n-m}) + x_{n-m}^I -y_n||^2 </math></div><br />
The total loss for the sample <math> \textbf{x}_n^{-M},y_n) </math> is then given by:<br />
<div style="text-align: center;"><math>L^{tot}(\textbf{x}_n^{-M}, y_n)=L^2(\widehat{y}_n, y_n)+\alpha L^{aux}(\textbf{x}_n^{-M}, y_n)</math></div><br />
where <math>\widehat{y}_n</math> was mentioned before, <math>\alpha \geq 0</math> is a constant.<br />
<br />
=Experiments=<br />
The paper evaluated SOCNN architecture on three datasets: artificially generated datasets, [https://archive.ics.uci.edu/ml/datasets/Individual+household+electric+power+consumption household electric power consumption dataset], and the financial dataset of bid/ask quotes sent by several market participants active in the credit derivatives market. Comparing its performance with simple CNN, single and multiplayer LSTM and 25-layer ResNet. Apart from the evaluation of the SOCNN architecture the paper also discusses the impact of network components such as: such as auxiliary<br />
loss and the depth of the offset sub-network. The code and datasets are available [https://github.com/mbinkowski/nntimeseries here]<br />
<br />
==Datasets==<br />
Artificial data: They generated 4 artificial series, <math> X_{K \times N}</math>, where <math>K \in \{16,64\} </math>. Therefore there is a synchronous and an asynchronous series for each K value.<br />
<br />
Electricity data: This UCI dataset contains 7 different features excluding date and time. The features include global active power, global reactive power, voltage, global intensity, sub-metering 1, sub-metering 2 and sub-metering 3, recorded every minute for 47 months. The data has been altered so that one observation contains only one value of 7 features, while durations between consecutive observations are ranged from 1 to 7 minutes. The goal is to predict all 7 features for the next time step.<br />
<br />
Non-anonymous quotes: The dataset contains 2.1 million quotes from 28 different sources from different market participants such as analysts, banks etc. Each quote is characterized by 31 features: the offered price, 28 indicators of the quoting source, the direction indicator (the quote refers to either a buy or a sell offer) and duration from the previous quote. For each source and direction, we want to predict the next quoted price from this given source and direction considering the last 60 quotes.<br />
<br />
==Training details==<br />
They applied grid search on some hyperparameters in order to get the significance of its components. The hyperparameters include the offset sub-network's depth and the auxiliary weight <math>\alpha</math>. For offset sub-network's depth, they use 1, 10,1 for artificial, electricity and quotes dataset respectively; and they compared the values of <math>\alpha</math> in {0,0.1,0.01}.<br />
<br />
They chose LeakyReLU as activation function for all networks:<br />
<div style="text-align: center;"><math>\sigma^{LeakyReLU}(x) = x</math> if <math>x\geq 0</math>, and <math>0.1x</math> otherwise </div><br />
They use the same number of layers, same stride and similar kernel size structure in CNN. In each trained CNN, they applied max pooling with the pool size of 2 every 2 convolutional layers.<br />
<br />
Table 1 presents the configuration of network hyperparameters used in comparison<br />
<br />
[[File:Junyi4.png | 400px|center|]]<br />
<br />
===Network Training===<br />
The training and validation data were sampled randomly from the first 80% of timesteps in each series, with ratio of 3 to 1. The remaining 20% of data was used as a test set.<br />
<br />
All models were trained using Adam optimizer because the authors found that its rate of convergence was much faster than standard Stochastic Gradient Descent in early tests.<br />
<br />
They used a batch size of 128 for artificial and electricity data, and 256 for quotes dataset, and applied batch normalization between each convolution and the following activation. <br />
<br />
At the beginning of each epoch, the training samples were randomly sampled. To prevent overfitting, they applied dropout and early stopping.<br />
<br />
Weights were initialized using the normalized uniform procedure proposed by Glorot & Bengio (2010).[14]<br />
<br />
The authors carried out the experiments on Tensorflow and Keras and used different GPU to optimize the model for different datasets.<br />
<br />
==Results==<br />
Table 2 shows all results performed from all datasets.<br />
[[File:Junyi5.png | 600px|center|]]<br />
We can see that SOCNN outperforms in all asynchronous artificial, electricity and quotes datasets. For synchronous data, LSTM might be slightly better, but SOCNN almost has the same results with LSTM. Phased LSTM and ResNet have performed really bad on artificial asynchronous dataset and quotes dataset respectively. Notice that having more than one layer of offset network would have negative impact on results. Also, the higher weights of auxiliary loss(<math>\alpha</math>considerably improved the test error on asynchronous dataset, see Table 3. However, for other datasets, its impact was negligible.<br />
[[File:Junyi6.png | 400px|center|]]<br />
In general, SOCNN has significantly lower variance of the test and validation errors, especially in the early stage of the training process and for quotes dataset. This effect can be seen in the learning curves for Asynchronous 64 artificial dataset presented in Figure 5.<br />
[[File:Junyi7.png | 500px|thumb|center|Figure 5: Learning curves with different auxiliary weights for SOCNN model trained on Asynchronous 64 dataset. The solid lines indicate the test error while the dashed lines indicate the training error.]]<br />
<br />
Finally, we want to test the robustness of the proposed model SOCNN, adding noise terms to asynchronous 16 dataset and check how these networks perform. The result is shown in Figure 6.<br />
[[File:Junyi8.png | 600px|thumb|center|Figure 6: Experiment comparing robustness of the considered networks for Asynchronous 16 dataset. The plots show how the error would change if an additional noise term was added to the input series. The dotted curves show the total significance and average absolute offset (not to scale) outputs for the noisy observations. Interestingly, the significance of the noisy observations increases with the magnitude of noise; i.e. noisy observations are far from being discarded by SOCNN.]]<br />
From Figure 6, the purple line and green line seems staying at the same position in training and testing process. SOCNN and single-layer LSTM are most robust compared to other networks, and least prone to overfitting.<br />
<br />
=Conclusion and Discussion=<br />
In this paper, the authors have proposed a new architecture called Significance-Offset Convolutional Neural Network, which combines AR-like weighting mechanism and convolutional neural network. This new architecture is designed for high-noise asynchronous time series and achieves outperformance in forecasting several asynchronous time series compared to popular convolutional and recurrent networks. <br />
<br />
The SOCNN can be extended further by adding intermediate weighting layers of the same type in the network structure. Another possible extension but needs further empirical studies is that we consider not just <math>1 \times 1</math> convolutional kernels on the offset sub-network. Also, this new architecture might be tested on other real-life datasets with relevant characteristics in the future, especially on econometric datasets and more generally for time series (stochastic processes) regression.<br />
<br />
=Critiques=<br />
#The paper is most likely an application paper, and the proposed new architecture shows improved performance over baselines in the asynchronous time series.<br />
#The quote data cannot be reached, only two datasets available.<br />
#The 'Significance' network was described as critical to the model in paper, but they did not show how the performance of SOCNN with respect to the significance network.<br />
#The transform of the original data to asynchronous data is not clear.<br />
#The experiments on the main application are not reproducible because the data is proprietary.<br />
#The way that train and test data were split is unclear. This could be important in the case of the financial data set.<br />
#Although the auxiliary loss function was mentioned as an important part, the advantages of it was not too clear in the paper. Maybe it is better that the paper describes a little more about its effectiveness.<br />
#It was not mentioned clearly in the paper whether the model training was done on a rolling basis for time series forecasting.<br />
#The noise term used in section 5's model robustness analysis uses evenly distributed noise (see Appendix B). While the analysis is a good start, analysis with different noise distributions would make the findings more generalizable.<br />
#The paper uses financial/economic data as one of its testing data set. Instead of comparing neural network models such as CNN which is known to work badly on time series data, it would be much better if the author compared to well-known econometric time series models such as GARCH and VAR.<br />
<br />
=References=<br />
[1] Hamilton, J. D. Time series analysis, volume 2. Princeton university press Princeton, 1994. <br />
<br />
[2] Fama, E. F. Efficient capital markets: A review of theory and empirical work. The journal of Finance, 25(2):383–417, 1970.<br />
<br />
[3] Petelin, D., Sˇindela ́ˇr, J., Pˇrikryl, J., and Kocijan, J. Financial modeling using gaussian process models. In Intelligent Data Acquisition and Advanced Computing Systems (IDAACS), 2011 IEEE 6th International Conference on, volume 2, pp. 672–677. IEEE, 2011.<br />
<br />
[4] Tobar, F., Bui, T. D., and Turner, R. E. Learning stationary time series using gaussian processes with nonparametric kernels. In Advances in Neural Information Processing Systems, pp. 3501–3509, 2015.<br />
<br />
[5] Hwang, Y., Tong, A., and Choi, J. Automatic construction of nonparametric relational regression models for multiple time series. In Proceedings of the 33rd International Conference on Machine Learning, 2016.<br />
<br />
[6] Wilson, A. and Ghahramani, Z. Copula processes. In Advances in Neural Information Processing Systems, pp. 2460–2468, 2010.<br />
<br />
[7] Sirignano, J. Extended abstract: Neural networks for limit order books, February 2016.<br />
<br />
[8] Borovykh, A., Bohte, S., and Oosterlee, C. W. Condi- tional time series forecasting with convolutional neural networks, March 2017.<br />
<br />
[9] Heaton, J. B., Polson, N. G., and Witte, J. H. Deep learn- ing in finance, February 2016.<br />
<br />
[10] Neil, D., Pfeiffer, M., and Liu, S.-C. Phased lstm: Acceler- ating recurrent network training for long or event-based sequences. In Advances In Neural Information Process- ing Systems, pp. 3882–3890, 2016.<br />
<br />
[11] Chung, J., Gulcehre, C., Cho, K., and Bengio, Y. Em- pirical evaluation of gated recurrent neural networks on sequence modeling, December 2014.<br />
<br />
[12] Weissenborn, D. and Rockta ̈schel, T. MuFuRU: The Multi-Function recurrent unit, June 2016.<br />
<br />
[13] Cho, K., Courville, A., and Bengio, Y. Describing multi- media content using attention-based Encoder–Decoder networks. IEEE Transactions on Multimedia, 17(11): 1875–1886, July 2015. ISSN 1520-9210.<br />
<br />
[14] Glorot, X. and Bengio, Y. Understanding the dif- ficulty of training deep feedforward neural net- works. In In Proceedings of the International Con- ference on Artificial Intelligence and Statistics (AIS- TATSaˆ10). Society for Artificial Intelligence and Statistics, 2010.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=stat946F18/Autoregressive_Convolutional_Neural_Networks_for_Asynchronous_Time_Series&diff=41755stat946F18/Autoregressive Convolutional Neural Networks for Asynchronous Time Series2018-11-28T20:06:57Z<p>Ka2khan: /* Gating and weighting mechanisms */</p>
<hr />
<div>This page is a summary of the paper "[http://proceedings.mlr.press/v80/binkowski18a/binkowski18a.pdf Autoregressive Convolutional Neural Networks for Asynchronous Time Series]" by Mikołaj Binkowski, Gautier Marti, Philippe Donnat. It was published at ICML in 2018. The code for this paper is provided [https://github.com/mbinkowski/nntimeseries here].<br />
<br />
=Introduction=<br />
In this paper, the authors propose a deep convolutional network architecture called Significance-Offset Convolutional Neural Network for regression of multivariate asynchronous time series. The model is inspired by standard autoregressive(AR) models and gating systems used in recurrent neural networks. The model is evaluated on various time series data including:<br />
# Hedge fund proprietary dataset of over 2 million quotes for a credit derivative index, <br />
# An artificially generated noisy auto-regressive series, <br />
# A UCI household electricity consumption dataset. <br />
<br />
This paper focuses on time series with multi-variate and noisy signals, especially financial data. Financial time series is challenging to predict due to their low signal-to-noise ratio and heavy-tailed distributions. For example, the same signal (e.g. price of a stock) is obtained from different sources (e.g. financial news, an investment bank, financial analyst etc.) asynchronously. Each source may have a different bias or noise. (Figure 1) The investment bank with more clients can update their information more precisely than the investment bank with fewer clients, then the significance of each past observations may depend on other factors that change in time. Therefore, the traditional econometric models such as AR, VAR, VARMA[1] might not be sufficient. However, their relatively good performance could allow us to combine such linear econometric models with deep neural networks that can learn highly nonlinear relationships. This model is inspired by the gating mechanism which is successful in RNNs and Highway Networks.<br />
<br />
The time series forecasting problem can be expressed as a conditional probability distribution below,<br />
<div style="text-align: center;"><math>p(X_{t+d}|X_t,X_{t-1},...) = f(X_t,X_{t-1},...)</math></div><br />
Thus, we focus on modeling the predictors of future values of time series given their past values. <br />
The predictability of financial dataset still remains an open problem and is discussed in various publications [2].<br />
<br />
[[File:Junyi1.png | 500px|thumb|center|Figure 1: Quotes from four different market participants (sources) for the same credit default swaps (CDS) throughout one day. Each trader displays from time to time the prices for which he offers to buy (bid) and sell (ask) the underlying CDS. The filled area marks the difference between the best sell and buy offers (spread) at each time.]]<br />
<br />
The paper also provides empirical evidence that their model which combines linear models with deep learning models could perform better than just DL models like CNN, LSTMs and Phased LSTMs.<br />
<br />
=Related Work=<br />
===Time series forecasting===<br />
From recent proceedings in main machine learning venues i.e. ICML, NIPS, AISTATS, UAI, we can notice that time series are often forecast using Gaussian processes[3,4], especially for irregularly sampled time series[5]. Though still largely independent, combined models have started to appear, for example, the Gaussian Copula Process Volatility model[6]. For this paper, the authors use coupling AR models and neural networks to achieve such combined models.<br />
<br />
Although deep neural networks have been applied into many fields and produced satisfactory results, there still is little literature on deep learning for time series forecasting. More recently, the papers include Sirignano (2016)[7] that used 4-layer perceptrons in modeling price change distributions in Limit Order Books, and Borovykh et al. (2017)[8] who applied more recent WaveNet architecture to several short univariate and bivariate time-series (including financial ones). Heaton et al. (2016)[9] claimed to use autoencoders with a single hidden layer to compress multivariate financial data. Neil et al. (2016)[10] presented augmentation of LSTM architecture suitable for asynchronous series, which stimulates learning dependencies of different frequencies through time gate. <br />
<br />
In this paper, the authors examine the capabilities of several architectures (CNN, residual network, multi-layer LSTM, and phase LSTM) on AR-like artificial asynchronous and noisy time series, household electricity consumption dataset, and on real financial data from the credit default swap market with some inefficiencies.<br />
<br />
====AR Model====<br />
<br />
An autoregressive (AR) model describes the next value in a time-series as a combination of previous values, scaling factors, a bias, and noise [https://onlinecourses.science.psu.edu/stat501/node/358/ (source)]. For a p-th order (relating the current state to the p last states), the equation of the model is:<br />
<br />
<math> X_t = c + \sum_{i=1}^p \varphi_i X_{t-i}+ \varepsilon_t \,</math> [https://en.wikipedia.org/wiki/Autoregressive_model#Definition (equation source)]<br />
<br />
With parameters/coefficients <math>\varphi_i</math>, constant <math>c</math>, and noise <math>\varepsilon_t</math> This can be extended to vector form to create the VAR model mentioned in the paper.<br />
<br />
===Gating and weighting mechanisms===<br />
Gating mechanisms for neural networks has ability to overcome the problem of vanishing gradient, and can be expressed as <math display="inline">f(x)=c(x) \otimes \sigma(x)</math>, where <math>f</math> is the output function, <math>c</math> is a "candidate output" (a nonlinear function of <math>x</math>), <math>\otimes</math> is an element-wise matrix product, and <math>\sigma : \mathbb{R} \rightarrow [0,1] </math> is a sigmoid nonlinearity that controls the amount of output passed to the next layer. Different composition of functions of the same type as described above have proven to be an essential ingredient in popular recurrent architecture such as LSTM and GRU[11].<br />
<br />
The main purpose of the proposed gating system is to weight the outputs of the intermediate layers within neural networks, and is most closely related to softmax gating used in MuFuRu(Multi-Function Recurrent Unit)[12], i.e.<br />
<math display="inline"> f(x) = \sum_{l=1}^L p^l(x) \otimes f^l(x)\text{,}\ p(x)=\text{softmax}(\widehat{p}(x)), </math>, where <math>(f^l)_{l=1}^L </math>are candidate outputs (composition operators in MuFuRu), <math>(\widehat{p}^l)_{l=1}^L </math>are linear functions of inputs. <br />
<br />
This idea is also successfully used in attention networks[13] such as image captioning and machine translation. In this paper, the proposed method is similar as the separate inputs (time series steps in this case) are weighted in accordance with learned functions of these inputs. The difference is that the functions are being modeled using multi-layer CNNs. Another difference is that the proposed method is not using recurrent layers, which enables the network to remember parts of the sentence/image already translated/described.<br />
<br />
=Motivation=<br />
There are mainly five motivations they stated in the paper:<br />
#The forecasting problem in this paper has done almost independently by econometrics and machine learning communities. Unlike in machine learning, research in econometrics are more likely to explain variables rather than improving out-of-sample prediction power. These models tend to 'over-fit' on financial time series, their parameters are unstable and have poor performance on out-of-sample prediction.<br />
#Although Gaussian processes provide a useful theoretical framework that is able to handle asynchronous data, they often follow heavy-tailed distribution for financial datasets.<br />
#Predictions of autoregressive time series may involve highly nonlinear functions if sampled irregularly. For AR time series with higher order and have more past observations, the expectation of it <math display="inline">\mathbb{E}[X(t)|{X(t-m), m=1,...,M}]</math> may involve more complicated functions that in general may not allow closed-form expression.<br />
#In practice, the dimensions of multivariate time series are often observed separately and asynchronously, such series at fixed frequency may lead to lose information or enlarge the dataset, which is shown in Figure 2(a). Therefore, the core of the proposed architecture SOCNN represents separate dimensions as a single one with dimension and duration indicators as additional features(Figure 2(b)).<br />
#Given a series of pairs of consecutive input values and corresponding durations, <math display="inline"> x_n = (X(t_n),t_n-t_{n-1}) </math>. One may expect that LSTM may memorize the input values in each step and weight them at the output according to the duration, but this approach may lead to an imbalance between the needs for memory and for linearity. The weights that are assigned to the memorized observations potentially require several layers of nonlinearity to be computed properly, while past observations might just need to be memorized as they are.<br />
<br />
[[File:Junyi2.png | 550px|thumb|center|Figure 2: (a) Fixed sampling frequency and its drawbacks; keep- ing all available information leads to much more datapoints. (b) Proposed data representation for the asynchronous series. Consecutive observations are stored together as a single value series, regardless of which series they belong to; this information, however, is stored in indicator features, alongside durations between observations.]]<br />
<br />
<br />
=Model Architecture=<br />
Suppose there's a multivariate time series <math display="inline">(x_n)_{n=0}^{\infty} \subset \mathbb{R}^d </math>, we want to predict the conditional future values of a subset of elements of <math>x_n</math><br />
<div style="text-align: center;"><math>y_n = \mathbb{E} [x_n^I | {x_{n-m}, m=1,2,...}], </math></div><br />
where <math> I=\{i_1,i_2,...i_{d_I}\} \subset \{1,2,...,d\} </math> is a subset of features of <math>x_n</math>.<br />
<br />
Let <math> \textbf{x}_n^{-M} = (x_{n-m})_{m=1}^M </math>. <br />
<br />
The estimator of <math>y_n</math> can be expressed as:<br />
<div style="text-align: center;"><math>\hat{y}_n = \sum_{m=1}^M [F(\textbf{x}_n^{-M}) \otimes \sigma(S(\textbf{x}_n^{-M}))].,_m ,</math></div><br />
The estimate is the summation of the columns of the matrix in bracket. Here<br />
#<math>F,S : \mathbb{R}^{d \times M} \rightarrow \mathbb{R}^{d_I \times M}</math> are neural networks. <br />
#* <math>S</math> is a fully convolutional network which is composed of convolutional layers only. <br />
#* <math display="inline">F(\textbf{x}_n^{-M}) = W \otimes [off(x_{n-m}) + x_{n-m}^I)]_{m=1}^M </math> <br />
#** <math> W \in \mathbb{R}^{d_I \times M}</math> <br />
#** <math> off: \mathbb{R}^d \rightarrow \mathbb{R}^{d_I} </math> is a multilayer perceptron.<br />
<br />
#<math>\sigma</math> is a normalized activation function independent at each row, i.e. <math display="inline"> \sigma ((a_1^T, ..., a_{d_I}^T)^T)=(\sigma(a_1)^T,..., \sigma(a_{d_I})^T)^T </math><br />
#* <math>a_{i} \in \mathbb{R}^{M}</math><br />
#* and <math>\sigma </math> is defined such that <math>\sigma(a)^{T} \mathbf{1}_{M}=1</math><br />
# <math>\otimes</math> is element-wise matrix multiplication.<br />
#<math>A.,_m</math> denotes the m-th column of a matrix A.<br />
<br />
Since <math>\sum_{m=1}^M W.,_m=W(1,1,...,1)^T</math> and <math>\sum_{m=1}^M S.,_m=S(1,1,...,1)^T</math>, we can express <math>\hat{y}_n</math> as:<br />
<div style="text-align: center;"><math>\hat{y}_n = \sum_{m=1}^M W.,_m \otimes (off(x_{n-m}) + x_{n-m}^I) \otimes \sigma(S.,_m(\textbf{x}_n^{-M}))</math></div><br />
This is the proposed network, Significance-Offset Convolutional Neural Network, <math>off</math> and <math>S</math> in the equation are corresponding to Offset and Significance in the name respectively.<br />
Figure 3 shows the scheme of network.<br />
<br />
[[File:Junyi3.png | 600px|thumb|center|Figure 3: A scheme of the proposed SOCNN architecture. The network preserves the time-dimension up to the top layer, while the number of features per timestep (filters) in the hidden layers is custom. The last convolutional layer, however, has the number of filters equal to dimension of the output. The Weighting frame shows how outputs from offset and significance networks are combined in accordance with Eq. of <math>\hat{y}_n</math>.]]<br />
<br />
The form of <math>\hat{y}_n</math> forced to separate the temporal dependence (obtained in weights <math>W_m</math>). S is determined by its filters which capture local dependencies and are independent of the relative position in time, the predictors <math>off(x_{n-m})</math> are completely independent of position in time. An adjusted single regressor for the target variable is provided by each past observation through the offset network. Since in asynchronous sampling procedure, consecutive values of x come from different signals and might be heterogeneous, therefore adjustment of offset network is important. In addition, significance network provides data-dependent weight for each regressor and sums them up in an autoregressive manner.<br />
<br />
===Relation to asynchronous data===<br />
One common problem of time series is that durations are varying between consecutive observations, the paper states two ways to solve this problem<br />
#Data preprocessing: aligning the observations at some fixed frequency e.g. duplicating and interpolating observations as shown in Figure 2(a). However, as mentioned in the figure, this approach will tend to loss of information and enlarge the size of the dataset and model complexity.<br />
#Add additional features: Treating the duration or time of the observations as additional features, it is the core of SOCNN, which is shown in Figure 2(b).<br />
<br />
===Loss function===<br />
The output of the offset network is series of separate predictors of changes between corresponding observations <math>x_{n-m}^I</math> and the target value<math>y_n</math>, this is the reason why we use auxiliary loss function, which equals to mean squared error of such intermediate predictions:<br />
<div style="text-align: center;"><math>L^{aux}(\textbf{x}_n^{-M}, y_n)=\frac{1}{M} \sum_{m=1}^M ||off(x_{n-m}) + x_{n-m}^I -y_n||^2 </math></div><br />
The total loss for the sample <math> \textbf{x}_n^{-M},y_n) </math> is then given by:<br />
<div style="text-align: center;"><math>L^{tot}(\textbf{x}_n^{-M}, y_n)=L^2(\widehat{y}_n, y_n)+\alpha L^{aux}(\textbf{x}_n^{-M}, y_n)</math></div><br />
where <math>\widehat{y}_n</math> was mentioned before, <math>\alpha \geq 0</math> is a constant.<br />
<br />
=Experiments=<br />
The paper evaluated SOCNN architecture on three datasets: artificially generated datasets, [https://archive.ics.uci.edu/ml/datasets/Individual+household+electric+power+consumption household electric power consumption dataset], and the financial dataset of bid/ask quotes sent by several market participants active in the credit derivatives market. Comparing its performance with simple CNN, single and multiplayer LSTM and 25-layer ResNet. Apart from the evaluation of the SOCNN architecture the paper also discusses the impact of network components such as: such as auxiliary<br />
loss and the depth of the offset sub-network. The code and datasets are available [https://github.com/mbinkowski/nntimeseries here]<br />
<br />
==Datasets==<br />
Artificial data: They generated 4 artificial series, <math> X_{K \times N}</math>, where <math>K \in \{16,64\} </math>. Therefore there is a synchronous and an asynchronous series for each K value.<br />
<br />
Electricity data: This UCI dataset contains 7 different features excluding date and time. The features include global active power, global reactive power, voltage, global intensity, sub-metering 1, sub-metering 2 and sub-metering 3, recorded every minute for 47 months. The data has been altered so that one observation contains only one value of 7 features, while durations between consecutive observations are ranged from 1 to 7 minutes. The goal is to predict all 7 features for the next time step.<br />
<br />
Non-anonymous quotes: The dataset contains 2.1 million quotes from 28 different sources from different market participants such as analysts, banks etc. Each quote is characterized by 31 features: the offered price, 28 indicators of the quoting source, the direction indicator (the quote refers to either a buy or a sell offer) and duration from the previous quote. For each source and direction, we want to predict the next quoted price from this given source and direction considering the last 60 quotes.<br />
<br />
==Training details==<br />
They applied grid search on some hyperparameters in order to get the significance of its components. The hyperparameters include the offset sub-network's depth and the auxiliary weight <math>\alpha</math>. For offset sub-network's depth, they use 1, 10,1 for artificial, electricity and quotes dataset respectively; and they compared the values of <math>\alpha</math> in {0,0.1,0.01}.<br />
<br />
They chose LeakyReLU as activation function for all networks:<br />
<div style="text-align: center;"><math>\sigma^{LeakyReLU}(x) = x</math> if <math>x\geq 0</math>, and <math>0.1x</math> otherwise </div><br />
They use the same number of layers, same stride and similar kernel size structure in CNN. In each trained CNN, they applied max pooling with the pool size of 2 every 2 convolutional layers.<br />
<br />
Table 1 presents the configuration of network hyperparameters used in comparison<br />
<br />
[[File:Junyi4.png | 400px|center|]]<br />
<br />
===Network Training===<br />
The training and validation data were sampled randomly from the first 80% of timesteps in each series, with ratio of 3 to 1. The remaining 20% of data was used as a test set.<br />
<br />
All models were trained using Adam optimizer because the authors found that its rate of convergence was much faster than standard Stochastic Gradient Descent in early tests.<br />
<br />
They used a batch size of 128 for artificial and electricity data, and 256 for quotes dataset, and applied batch normalization between each convolution and the following activation. <br />
<br />
At the beginning of each epoch, the training samples were randomly sampled. To prevent overfitting, they applied dropout and early stopping.<br />
<br />
Weights were initialized using the normalized uniform procedure proposed by Glorot & Bengio (2010).[14]<br />
<br />
The authors carried out the experiments on Tensorflow and Keras and used different GPU to optimize the model for different datasets.<br />
<br />
==Results==<br />
Table 2 shows all results performed from all datasets.<br />
[[File:Junyi5.png | 600px|center|]]<br />
We can see that SOCNN outperforms in all asynchronous artificial, electricity and quotes datasets. For synchronous data, LSTM might be slightly better, but SOCNN almost has the same results with LSTM. Phased LSTM and ResNet have performed really bad on artificial asynchronous dataset and quotes dataset respectively. Notice that having more than one layer of offset network would have negative impact on results. Also, the higher weights of auxiliary loss(<math>\alpha</math>considerably improved the test error on asynchronous dataset, see Table 3. However, for other datasets, its impact was negligible.<br />
[[File:Junyi6.png | 400px|center|]]<br />
In general, SOCNN has significantly lower variance of the test and validation errors, especially in the early stage of the training process and for quotes dataset. This effect can be seen in the learning curves for Asynchronous 64 artificial dataset presented in Figure 5.<br />
[[File:Junyi7.png | 500px|thumb|center|Figure 5: Learning curves with different auxiliary weights for SOCNN model trained on Asynchronous 64 dataset. The solid lines indicate the test error while the dashed lines indicate the training error.]]<br />
<br />
Finally, we want to test the robustness of the proposed model SOCNN, adding noise terms to asynchronous 16 dataset and check how these networks perform. The result is shown in Figure 6.<br />
[[File:Junyi8.png | 600px|thumb|center|Figure 6: Experiment comparing robustness of the considered networks for Asynchronous 16 dataset. The plots show how the error would change if an additional noise term was added to the input series. The dotted curves show the total significance and average absolute offset (not to scale) outputs for the noisy observations. Interestingly, the significance of the noisy observations increases with the magnitude of noise; i.e. noisy observations are far from being discarded by SOCNN.]]<br />
From Figure 6, the purple line and green line seems staying at the same position in training and testing process. SOCNN and single-layer LSTM are most robust compared to other networks, and least prone to overfitting.<br />
<br />
=Conclusion and Discussion=<br />
In this paper, the authors have proposed a new architecture called Significance-Offset Convolutional Neural Network, which combines AR-like weighting mechanism and convolutional neural network. This new architecture is designed for high-noise asynchronous time series and achieves outperformance in forecasting several asynchronous time series compared to popular convolutional and recurrent networks. <br />
<br />
The SOCNN can be extended further by adding intermediate weighting layers of the same type in the network structure. Another possible extension but needs further empirical studies is that we consider not just <math>1 \times 1</math> convolutional kernels on the offset sub-network. Also, this new architecture might be tested on other real-life datasets with relevant characteristics in the future, especially on econometric datasets and more generally for time series (stochastic processes) regression.<br />
<br />
=Critiques=<br />
#The paper is most likely an application paper, and the proposed new architecture shows improved performance over baselines in the asynchronous time series.<br />
#The quote data cannot be reached, only two datasets available.<br />
#The 'Significance' network was described as critical to the model in paper, but they did not show how the performance of SOCNN with respect to the significance network.<br />
#The transform of the original data to asynchronous data is not clear.<br />
#The experiments on the main application are not reproducible because the data is proprietary.<br />
#The way that train and test data were split is unclear. This could be important in the case of the financial data set.<br />
#Although the auxiliary loss function was mentioned as an important part, the advantages of it was not too clear in the paper. Maybe it is better that the paper describes a little more about its effectiveness.<br />
#It was not mentioned clearly in the paper whether the model training was done on a rolling basis for time series forecasting.<br />
#The noise term used in section 5's model robustness analysis uses evenly distributed noise (see Appendix B). While the analysis is a good start, analysis with different noise distributions would make the findings more generalizable.<br />
#The paper uses financial/economic data as one of its testing data set. Instead of comparing neural network models such as CNN which is known to work badly on time series data, it would be much better if the author compared to well-known econometric time series models such as GARCH and VAR.<br />
<br />
=References=<br />
[1] Hamilton, J. D. Time series analysis, volume 2. Princeton university press Princeton, 1994. <br />
<br />
[2] Fama, E. F. Efficient capital markets: A review of theory and empirical work. The journal of Finance, 25(2):383–417, 1970.<br />
<br />
[3] Petelin, D., Sˇindela ́ˇr, J., Pˇrikryl, J., and Kocijan, J. Financial modeling using gaussian process models. In Intelligent Data Acquisition and Advanced Computing Systems (IDAACS), 2011 IEEE 6th International Conference on, volume 2, pp. 672–677. IEEE, 2011.<br />
<br />
[4] Tobar, F., Bui, T. D., and Turner, R. E. Learning stationary time series using gaussian processes with nonparametric kernels. In Advances in Neural Information Processing Systems, pp. 3501–3509, 2015.<br />
<br />
[5] Hwang, Y., Tong, A., and Choi, J. Automatic construction of nonparametric relational regression models for multiple time series. In Proceedings of the 33rd International Conference on Machine Learning, 2016.<br />
<br />
[6] Wilson, A. and Ghahramani, Z. Copula processes. In Advances in Neural Information Processing Systems, pp. 2460–2468, 2010.<br />
<br />
[7] Sirignano, J. Extended abstract: Neural networks for limit order books, February 2016.<br />
<br />
[8] Borovykh, A., Bohte, S., and Oosterlee, C. W. Condi- tional time series forecasting with convolutional neural networks, March 2017.<br />
<br />
[9] Heaton, J. B., Polson, N. G., and Witte, J. H. Deep learn- ing in finance, February 2016.<br />
<br />
[10] Neil, D., Pfeiffer, M., and Liu, S.-C. Phased lstm: Acceler- ating recurrent network training for long or event-based sequences. In Advances In Neural Information Process- ing Systems, pp. 3882–3890, 2016.<br />
<br />
[11] Chung, J., Gulcehre, C., Cho, K., and Bengio, Y. Em- pirical evaluation of gated recurrent neural networks on sequence modeling, December 2014.<br />
<br />
[12] Weissenborn, D. and Rockta ̈schel, T. MuFuRU: The Multi-Function recurrent unit, June 2016.<br />
<br />
[13] Cho, K., Courville, A., and Bengio, Y. Describing multi- media content using attention-based Encoder–Decoder networks. IEEE Transactions on Multimedia, 17(11): 1875–1886, July 2015. ISSN 1520-9210.<br />
<br />
[14] Glorot, X. and Bengio, Y. Understanding the dif- ficulty of training deep feedforward neural net- works. In In Proceedings of the International Con- ference on Artificial Intelligence and Statistics (AIS- TATSaˆ10). Society for Artificial Intelligence and Statistics, 2010.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=stat946F18/Autoregressive_Convolutional_Neural_Networks_for_Asynchronous_Time_Series&diff=41754stat946F18/Autoregressive Convolutional Neural Networks for Asynchronous Time Series2018-11-28T19:56:47Z<p>Ka2khan: /* Gating and weighting mechanisms */</p>
<hr />
<div>This page is a summary of the paper "[http://proceedings.mlr.press/v80/binkowski18a/binkowski18a.pdf Autoregressive Convolutional Neural Networks for Asynchronous Time Series]" by Mikołaj Binkowski, Gautier Marti, Philippe Donnat. It was published at ICML in 2018. The code for this paper is provided [https://github.com/mbinkowski/nntimeseries here].<br />
<br />
=Introduction=<br />
In this paper, the authors propose a deep convolutional network architecture called Significance-Offset Convolutional Neural Network for regression of multivariate asynchronous time series. The model is inspired by standard autoregressive(AR) models and gating systems used in recurrent neural networks. The model is evaluated on various time series data including:<br />
# Hedge fund proprietary dataset of over 2 million quotes for a credit derivative index, <br />
# An artificially generated noisy auto-regressive series, <br />
# A UCI household electricity consumption dataset. <br />
<br />
This paper focuses on time series with multi-variate and noisy signals, especially financial data. Financial time series is challenging to predict due to their low signal-to-noise ratio and heavy-tailed distributions. For example, the same signal (e.g. price of a stock) is obtained from different sources (e.g. financial news, an investment bank, financial analyst etc.) asynchronously. Each source may have a different bias or noise. (Figure 1) The investment bank with more clients can update their information more precisely than the investment bank with fewer clients, then the significance of each past observations may depend on other factors that change in time. Therefore, the traditional econometric models such as AR, VAR, VARMA[1] might not be sufficient. However, their relatively good performance could allow us to combine such linear econometric models with deep neural networks that can learn highly nonlinear relationships. This model is inspired by the gating mechanism which is successful in RNNs and Highway Networks.<br />
<br />
The time series forecasting problem can be expressed as a conditional probability distribution below,<br />
<div style="text-align: center;"><math>p(X_{t+d}|X_t,X_{t-1},...) = f(X_t,X_{t-1},...)</math></div><br />
Thus, we focus on modeling the predictors of future values of time series given their past values. <br />
The predictability of financial dataset still remains an open problem and is discussed in various publications [2].<br />
<br />
[[File:Junyi1.png | 500px|thumb|center|Figure 1: Quotes from four different market participants (sources) for the same credit default swaps (CDS) throughout one day. Each trader displays from time to time the prices for which he offers to buy (bid) and sell (ask) the underlying CDS. The filled area marks the difference between the best sell and buy offers (spread) at each time.]]<br />
<br />
The paper also provides empirical evidence that their model which combines linear models with deep learning models could perform better than just DL models like CNN, LSTMs and Phased LSTMs.<br />
<br />
=Related Work=<br />
===Time series forecasting===<br />
From recent proceedings in main machine learning venues i.e. ICML, NIPS, AISTATS, UAI, we can notice that time series are often forecast using Gaussian processes[3,4], especially for irregularly sampled time series[5]. Though still largely independent, combined models have started to appear, for example, the Gaussian Copula Process Volatility model[6]. For this paper, the authors use coupling AR models and neural networks to achieve such combined models.<br />
<br />
Although deep neural networks have been applied into many fields and produced satisfactory results, there still is little literature on deep learning for time series forecasting. More recently, the papers include Sirignano (2016)[7] that used 4-layer perceptrons in modeling price change distributions in Limit Order Books, and Borovykh et al. (2017)[8] who applied more recent WaveNet architecture to several short univariate and bivariate time-series (including financial ones). Heaton et al. (2016)[9] claimed to use autoencoders with a single hidden layer to compress multivariate financial data. Neil et al. (2016)[10] presented augmentation of LSTM architecture suitable for asynchronous series, which stimulates learning dependencies of different frequencies through time gate. <br />
<br />
In this paper, the authors examine the capabilities of several architectures (CNN, residual network, multi-layer LSTM, and phase LSTM) on AR-like artificial asynchronous and noisy time series, household electricity consumption dataset, and on real financial data from the credit default swap market with some inefficiencies.<br />
<br />
====AR Model====<br />
<br />
An autoregressive (AR) model describes the next value in a time-series as a combination of previous values, scaling factors, a bias, and noise [https://onlinecourses.science.psu.edu/stat501/node/358/ (source)]. For a p-th order (relating the current state to the p last states), the equation of the model is:<br />
<br />
<math> X_t = c + \sum_{i=1}^p \varphi_i X_{t-i}+ \varepsilon_t \,</math> [https://en.wikipedia.org/wiki/Autoregressive_model#Definition (equation source)]<br />
<br />
With parameters/coefficients <math>\varphi_i</math>, constant <math>c</math>, and noise <math>\varepsilon_t</math> This can be extended to vector form to create the VAR model mentioned in the paper.<br />
<br />
===Gating and weighting mechanisms===<br />
Gating mechanisms for neural networks has ability to overcome the problem of vanishing gradient, and can be expressed as <math display="inline">f(x)=c(x) \otimes \sigma(x)</math>, where <math>f</math> is the output function, <math>c</math> is a "candidate output" (a nonlinear function of <math>x</math>), <math>\otimes</math> is an element-wise matrix product, and <math>\sigma : \mathbb{R} \rightarrow [0,1] </math> is a sigmoid nonlinearity that controls the amount of output passed to the next layer. Different composition of functions of the same type as described above have proven to be an essential ingredient in popular recurrent architecture such as LSTM and GRU[11].<br />
<br />
The main purpose of the proposed gating system is to weight the outputs of the intermediate layers within neural networks, and is most closely related to softmax gating used in MuFuRu(Multi-Function Recurrent Unit)[12], i.e.<br />
<math display="inline"> f(x) = \sum_{l=1}^L p^l(x) \otimes f^l(x)\text{,}\ p(x)=\text{softmax}(\widehat{p}(x)), </math>, where <math>(f^l)_{l=1}^L </math>are candidate outputs (composition operators in MuFuRu), <math>(\widehat{p}^l)_{l=1}^L </math>are linear functions of inputs. <br />
<br />
This idea is also successfully used in attention networks[13] such as image captioning and machine translation. In this paper, the method is similar as this. The difference is that modeling the functions as multi-layer CNNs. Another difference is that not using recurrent layers, which can enable the network to remember the parts of the sentence/image already translated/described.<br />
<br />
=Motivation=<br />
There are mainly five motivations they stated in the paper:<br />
#The forecasting problem in this paper has done almost independently by econometrics and machine learning communities. Unlike in machine learning, research in econometrics are more likely to explain variables rather than improving out-of-sample prediction power. These models tend to 'over-fit' on financial time series, their parameters are unstable and have poor performance on out-of-sample prediction.<br />
#Although Gaussian processes provide a useful theoretical framework that is able to handle asynchronous data, they often follow heavy-tailed distribution for financial datasets.<br />
#Predictions of autoregressive time series may involve highly nonlinear functions if sampled irregularly. For AR time series with higher order and have more past observations, the expectation of it <math display="inline">\mathbb{E}[X(t)|{X(t-m), m=1,...,M}]</math> may involve more complicated functions that in general may not allow closed-form expression.<br />
#In practice, the dimensions of multivariate time series are often observed separately and asynchronously, such series at fixed frequency may lead to lose information or enlarge the dataset, which is shown in Figure 2(a). Therefore, the core of the proposed architecture SOCNN represents separate dimensions as a single one with dimension and duration indicators as additional features(Figure 2(b)).<br />
#Given a series of pairs of consecutive input values and corresponding durations, <math display="inline"> x_n = (X(t_n),t_n-t_{n-1}) </math>. One may expect that LSTM may memorize the input values in each step and weight them at the output according to the duration, but this approach may lead to an imbalance between the needs for memory and for linearity. The weights that are assigned to the memorized observations potentially require several layers of nonlinearity to be computed properly, while past observations might just need to be memorized as they are.<br />
<br />
[[File:Junyi2.png | 550px|thumb|center|Figure 2: (a) Fixed sampling frequency and its drawbacks; keep- ing all available information leads to much more datapoints. (b) Proposed data representation for the asynchronous series. Consecutive observations are stored together as a single value series, regardless of which series they belong to; this information, however, is stored in indicator features, alongside durations between observations.]]<br />
<br />
<br />
=Model Architecture=<br />
Suppose there's a multivariate time series <math display="inline">(x_n)_{n=0}^{\infty} \subset \mathbb{R}^d </math>, we want to predict the conditional future values of a subset of elements of <math>x_n</math><br />
<div style="text-align: center;"><math>y_n = \mathbb{E} [x_n^I | {x_{n-m}, m=1,2,...}], </math></div><br />
where <math> I=\{i_1,i_2,...i_{d_I}\} \subset \{1,2,...,d\} </math> is a subset of features of <math>x_n</math>.<br />
<br />
Let <math> \textbf{x}_n^{-M} = (x_{n-m})_{m=1}^M </math>. <br />
<br />
The estimator of <math>y_n</math> can be expressed as:<br />
<div style="text-align: center;"><math>\hat{y}_n = \sum_{m=1}^M [F(\textbf{x}_n^{-M}) \otimes \sigma(S(\textbf{x}_n^{-M}))].,_m ,</math></div><br />
The estimate is the summation of the columns of the matrix in bracket. Here<br />
#<math>F,S : \mathbb{R}^{d \times M} \rightarrow \mathbb{R}^{d_I \times M}</math> are neural networks. <br />
#* <math>S</math> is a fully convolutional network which is composed of convolutional layers only. <br />
#* <math display="inline">F(\textbf{x}_n^{-M}) = W \otimes [off(x_{n-m}) + x_{n-m}^I)]_{m=1}^M </math> <br />
#** <math> W \in \mathbb{R}^{d_I \times M}</math> <br />
#** <math> off: \mathbb{R}^d \rightarrow \mathbb{R}^{d_I} </math> is a multilayer perceptron.<br />
<br />
#<math>\sigma</math> is a normalized activation function independent at each row, i.e. <math display="inline"> \sigma ((a_1^T, ..., a_{d_I}^T)^T)=(\sigma(a_1)^T,..., \sigma(a_{d_I})^T)^T </math><br />
#* <math>a_{i} \in \mathbb{R}^{M}</math><br />
#* and <math>\sigma </math> is defined such that <math>\sigma(a)^{T} \mathbf{1}_{M}=1</math><br />
# <math>\otimes</math> is element-wise matrix multiplication.<br />
#<math>A.,_m</math> denotes the m-th column of a matrix A.<br />
<br />
Since <math>\sum_{m=1}^M W.,_m=W(1,1,...,1)^T</math> and <math>\sum_{m=1}^M S.,_m=S(1,1,...,1)^T</math>, we can express <math>\hat{y}_n</math> as:<br />
<div style="text-align: center;"><math>\hat{y}_n = \sum_{m=1}^M W.,_m \otimes (off(x_{n-m}) + x_{n-m}^I) \otimes \sigma(S.,_m(\textbf{x}_n^{-M}))</math></div><br />
This is the proposed network, Significance-Offset Convolutional Neural Network, <math>off</math> and <math>S</math> in the equation are corresponding to Offset and Significance in the name respectively.<br />
Figure 3 shows the scheme of network.<br />
<br />
[[File:Junyi3.png | 600px|thumb|center|Figure 3: A scheme of the proposed SOCNN architecture. The network preserves the time-dimension up to the top layer, while the number of features per timestep (filters) in the hidden layers is custom. The last convolutional layer, however, has the number of filters equal to dimension of the output. The Weighting frame shows how outputs from offset and significance networks are combined in accordance with Eq. of <math>\hat{y}_n</math>.]]<br />
<br />
The form of <math>\hat{y}_n</math> forced to separate the temporal dependence (obtained in weights <math>W_m</math>). S is determined by its filters which capture local dependencies and are independent of the relative position in time, the predictors <math>off(x_{n-m})</math> are completely independent of position in time. An adjusted single regressor for the target variable is provided by each past observation through the offset network. Since in asynchronous sampling procedure, consecutive values of x come from different signals and might be heterogeneous, therefore adjustment of offset network is important. In addition, significance network provides data-dependent weight for each regressor and sums them up in an autoregressive manner.<br />
<br />
===Relation to asynchronous data===<br />
One common problem of time series is that durations are varying between consecutive observations, the paper states two ways to solve this problem<br />
#Data preprocessing: aligning the observations at some fixed frequency e.g. duplicating and interpolating observations as shown in Figure 2(a). However, as mentioned in the figure, this approach will tend to loss of information and enlarge the size of the dataset and model complexity.<br />
#Add additional features: Treating the duration or time of the observations as additional features, it is the core of SOCNN, which is shown in Figure 2(b).<br />
<br />
===Loss function===<br />
The output of the offset network is series of separate predictors of changes between corresponding observations <math>x_{n-m}^I</math> and the target value<math>y_n</math>, this is the reason why we use auxiliary loss function, which equals to mean squared error of such intermediate predictions:<br />
<div style="text-align: center;"><math>L^{aux}(\textbf{x}_n^{-M}, y_n)=\frac{1}{M} \sum_{m=1}^M ||off(x_{n-m}) + x_{n-m}^I -y_n||^2 </math></div><br />
The total loss for the sample <math> \textbf{x}_n^{-M},y_n) </math> is then given by:<br />
<div style="text-align: center;"><math>L^{tot}(\textbf{x}_n^{-M}, y_n)=L^2(\widehat{y}_n, y_n)+\alpha L^{aux}(\textbf{x}_n^{-M}, y_n)</math></div><br />
where <math>\widehat{y}_n</math> was mentioned before, <math>\alpha \geq 0</math> is a constant.<br />
<br />
=Experiments=<br />
The paper evaluated SOCNN architecture on three datasets: artificially generated datasets, [https://archive.ics.uci.edu/ml/datasets/Individual+household+electric+power+consumption household electric power consumption dataset], and the financial dataset of bid/ask quotes sent by several market participants active in the credit derivatives market. Comparing its performance with simple CNN, single and multiplayer LSTM and 25-layer ResNet. Apart from the evaluation of the SOCNN architecture the paper also discusses the impact of network components such as: such as auxiliary<br />
loss and the depth of the offset sub-network. The code and datasets are available [https://github.com/mbinkowski/nntimeseries here]<br />
<br />
==Datasets==<br />
Artificial data: They generated 4 artificial series, <math> X_{K \times N}</math>, where <math>K \in \{16,64\} </math>. Therefore there is a synchronous and an asynchronous series for each K value.<br />
<br />
Electricity data: This UCI dataset contains 7 different features excluding date and time. The features include global active power, global reactive power, voltage, global intensity, sub-metering 1, sub-metering 2 and sub-metering 3, recorded every minute for 47 months. The data has been altered so that one observation contains only one value of 7 features, while durations between consecutive observations are ranged from 1 to 7 minutes. The goal is to predict all 7 features for the next time step.<br />
<br />
Non-anonymous quotes: The dataset contains 2.1 million quotes from 28 different sources from different market participants such as analysts, banks etc. Each quote is characterized by 31 features: the offered price, 28 indicators of the quoting source, the direction indicator (the quote refers to either a buy or a sell offer) and duration from the previous quote. For each source and direction, we want to predict the next quoted price from this given source and direction considering the last 60 quotes.<br />
<br />
==Training details==<br />
They applied grid search on some hyperparameters in order to get the significance of its components. The hyperparameters include the offset sub-network's depth and the auxiliary weight <math>\alpha</math>. For offset sub-network's depth, they use 1, 10,1 for artificial, electricity and quotes dataset respectively; and they compared the values of <math>\alpha</math> in {0,0.1,0.01}.<br />
<br />
They chose LeakyReLU as activation function for all networks:<br />
<div style="text-align: center;"><math>\sigma^{LeakyReLU}(x) = x</math> if <math>x\geq 0</math>, and <math>0.1x</math> otherwise </div><br />
They use the same number of layers, same stride and similar kernel size structure in CNN. In each trained CNN, they applied max pooling with the pool size of 2 every 2 convolutional layers.<br />
<br />
Table 1 presents the configuration of network hyperparameters used in comparison<br />
<br />
[[File:Junyi4.png | 400px|center|]]<br />
<br />
===Network Training===<br />
The training and validation data were sampled randomly from the first 80% of timesteps in each series, with ratio of 3 to 1. The remaining 20% of data was used as a test set.<br />
<br />
All models were trained using Adam optimizer because the authors found that its rate of convergence was much faster than standard Stochastic Gradient Descent in early tests.<br />
<br />
They used a batch size of 128 for artificial and electricity data, and 256 for quotes dataset, and applied batch normalization between each convolution and the following activation. <br />
<br />
At the beginning of each epoch, the training samples were randomly sampled. To prevent overfitting, they applied dropout and early stopping.<br />
<br />
Weights were initialized using the normalized uniform procedure proposed by Glorot & Bengio (2010).[14]<br />
<br />
The authors carried out the experiments on Tensorflow and Keras and used different GPU to optimize the model for different datasets.<br />
<br />
==Results==<br />
Table 2 shows all results performed from all datasets.<br />
[[File:Junyi5.png | 600px|center|]]<br />
We can see that SOCNN outperforms in all asynchronous artificial, electricity and quotes datasets. For synchronous data, LSTM might be slightly better, but SOCNN almost has the same results with LSTM. Phased LSTM and ResNet have performed really bad on artificial asynchronous dataset and quotes dataset respectively. Notice that having more than one layer of offset network would have negative impact on results. Also, the higher weights of auxiliary loss(<math>\alpha</math>considerably improved the test error on asynchronous dataset, see Table 3. However, for other datasets, its impact was negligible.<br />
[[File:Junyi6.png | 400px|center|]]<br />
In general, SOCNN has significantly lower variance of the test and validation errors, especially in the early stage of the training process and for quotes dataset. This effect can be seen in the learning curves for Asynchronous 64 artificial dataset presented in Figure 5.<br />
[[File:Junyi7.png | 500px|thumb|center|Figure 5: Learning curves with different auxiliary weights for SOCNN model trained on Asynchronous 64 dataset. The solid lines indicate the test error while the dashed lines indicate the training error.]]<br />
<br />
Finally, we want to test the robustness of the proposed model SOCNN, adding noise terms to asynchronous 16 dataset and check how these networks perform. The result is shown in Figure 6.<br />
[[File:Junyi8.png | 600px|thumb|center|Figure 6: Experiment comparing robustness of the considered networks for Asynchronous 16 dataset. The plots show how the error would change if an additional noise term was added to the input series. The dotted curves show the total significance and average absolute offset (not to scale) outputs for the noisy observations. Interestingly, the significance of the noisy observations increases with the magnitude of noise; i.e. noisy observations are far from being discarded by SOCNN.]]<br />
From Figure 6, the purple line and green line seems staying at the same position in training and testing process. SOCNN and single-layer LSTM are most robust compared to other networks, and least prone to overfitting.<br />
<br />
=Conclusion and Discussion=<br />
In this paper, the authors have proposed a new architecture called Significance-Offset Convolutional Neural Network, which combines AR-like weighting mechanism and convolutional neural network. This new architecture is designed for high-noise asynchronous time series and achieves outperformance in forecasting several asynchronous time series compared to popular convolutional and recurrent networks. <br />
<br />
The SOCNN can be extended further by adding intermediate weighting layers of the same type in the network structure. Another possible extension but needs further empirical studies is that we consider not just <math>1 \times 1</math> convolutional kernels on the offset sub-network. Also, this new architecture might be tested on other real-life datasets with relevant characteristics in the future, especially on econometric datasets and more generally for time series (stochastic processes) regression.<br />
<br />
=Critiques=<br />
#The paper is most likely an application paper, and the proposed new architecture shows improved performance over baselines in the asynchronous time series.<br />
#The quote data cannot be reached, only two datasets available.<br />
#The 'Significance' network was described as critical to the model in paper, but they did not show how the performance of SOCNN with respect to the significance network.<br />
#The transform of the original data to asynchronous data is not clear.<br />
#The experiments on the main application are not reproducible because the data is proprietary.<br />
#The way that train and test data were split is unclear. This could be important in the case of the financial data set.<br />
#Although the auxiliary loss function was mentioned as an important part, the advantages of it was not too clear in the paper. Maybe it is better that the paper describes a little more about its effectiveness.<br />
#It was not mentioned clearly in the paper whether the model training was done on a rolling basis for time series forecasting.<br />
#The noise term used in section 5's model robustness analysis uses evenly distributed noise (see Appendix B). While the analysis is a good start, analysis with different noise distributions would make the findings more generalizable.<br />
#The paper uses financial/economic data as one of its testing data set. Instead of comparing neural network models such as CNN which is known to work badly on time series data, it would be much better if the author compared to well-known econometric time series models such as GARCH and VAR.<br />
<br />
=References=<br />
[1] Hamilton, J. D. Time series analysis, volume 2. Princeton university press Princeton, 1994. <br />
<br />
[2] Fama, E. F. Efficient capital markets: A review of theory and empirical work. The journal of Finance, 25(2):383–417, 1970.<br />
<br />
[3] Petelin, D., Sˇindela ́ˇr, J., Pˇrikryl, J., and Kocijan, J. Financial modeling using gaussian process models. In Intelligent Data Acquisition and Advanced Computing Systems (IDAACS), 2011 IEEE 6th International Conference on, volume 2, pp. 672–677. IEEE, 2011.<br />
<br />
[4] Tobar, F., Bui, T. D., and Turner, R. E. Learning stationary time series using gaussian processes with nonparametric kernels. In Advances in Neural Information Processing Systems, pp. 3501–3509, 2015.<br />
<br />
[5] Hwang, Y., Tong, A., and Choi, J. Automatic construction of nonparametric relational regression models for multiple time series. In Proceedings of the 33rd International Conference on Machine Learning, 2016.<br />
<br />
[6] Wilson, A. and Ghahramani, Z. Copula processes. In Advances in Neural Information Processing Systems, pp. 2460–2468, 2010.<br />
<br />
[7] Sirignano, J. Extended abstract: Neural networks for limit order books, February 2016.<br />
<br />
[8] Borovykh, A., Bohte, S., and Oosterlee, C. W. Condi- tional time series forecasting with convolutional neural networks, March 2017.<br />
<br />
[9] Heaton, J. B., Polson, N. G., and Witte, J. H. Deep learn- ing in finance, February 2016.<br />
<br />
[10] Neil, D., Pfeiffer, M., and Liu, S.-C. Phased lstm: Acceler- ating recurrent network training for long or event-based sequences. In Advances In Neural Information Process- ing Systems, pp. 3882–3890, 2016.<br />
<br />
[11] Chung, J., Gulcehre, C., Cho, K., and Bengio, Y. Em- pirical evaluation of gated recurrent neural networks on sequence modeling, December 2014.<br />
<br />
[12] Weissenborn, D. and Rockta ̈schel, T. MuFuRU: The Multi-Function recurrent unit, June 2016.<br />
<br />
[13] Cho, K., Courville, A., and Bengio, Y. Describing multi- media content using attention-based Encoder–Decoder networks. IEEE Transactions on Multimedia, 17(11): 1875–1886, July 2015. ISSN 1520-9210.<br />
<br />
[14] Glorot, X. and Bengio, Y. Understanding the dif- ficulty of training deep feedforward neural net- works. In In Proceedings of the International Con- ference on Artificial Intelligence and Statistics (AIS- TATSaˆ10). Society for Artificial Intelligence and Statistics, 2010.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=stat946F18/Autoregressive_Convolutional_Neural_Networks_for_Asynchronous_Time_Series&diff=41753stat946F18/Autoregressive Convolutional Neural Networks for Asynchronous Time Series2018-11-28T19:51:50Z<p>Ka2khan: /* Gating and weighting mechanisms */</p>
<hr />
<div>This page is a summary of the paper "[http://proceedings.mlr.press/v80/binkowski18a/binkowski18a.pdf Autoregressive Convolutional Neural Networks for Asynchronous Time Series]" by Mikołaj Binkowski, Gautier Marti, Philippe Donnat. It was published at ICML in 2018. The code for this paper is provided [https://github.com/mbinkowski/nntimeseries here].<br />
<br />
=Introduction=<br />
In this paper, the authors propose a deep convolutional network architecture called Significance-Offset Convolutional Neural Network for regression of multivariate asynchronous time series. The model is inspired by standard autoregressive(AR) models and gating systems used in recurrent neural networks. The model is evaluated on various time series data including:<br />
# Hedge fund proprietary dataset of over 2 million quotes for a credit derivative index, <br />
# An artificially generated noisy auto-regressive series, <br />
# A UCI household electricity consumption dataset. <br />
<br />
This paper focuses on time series with multi-variate and noisy signals, especially financial data. Financial time series is challenging to predict due to their low signal-to-noise ratio and heavy-tailed distributions. For example, the same signal (e.g. price of a stock) is obtained from different sources (e.g. financial news, an investment bank, financial analyst etc.) asynchronously. Each source may have a different bias or noise. (Figure 1) The investment bank with more clients can update their information more precisely than the investment bank with fewer clients, then the significance of each past observations may depend on other factors that change in time. Therefore, the traditional econometric models such as AR, VAR, VARMA[1] might not be sufficient. However, their relatively good performance could allow us to combine such linear econometric models with deep neural networks that can learn highly nonlinear relationships. This model is inspired by the gating mechanism which is successful in RNNs and Highway Networks.<br />
<br />
The time series forecasting problem can be expressed as a conditional probability distribution below,<br />
<div style="text-align: center;"><math>p(X_{t+d}|X_t,X_{t-1},...) = f(X_t,X_{t-1},...)</math></div><br />
Thus, we focus on modeling the predictors of future values of time series given their past values. <br />
The predictability of financial dataset still remains an open problem and is discussed in various publications [2].<br />
<br />
[[File:Junyi1.png | 500px|thumb|center|Figure 1: Quotes from four different market participants (sources) for the same credit default swaps (CDS) throughout one day. Each trader displays from time to time the prices for which he offers to buy (bid) and sell (ask) the underlying CDS. The filled area marks the difference between the best sell and buy offers (spread) at each time.]]<br />
<br />
The paper also provides empirical evidence that their model which combines linear models with deep learning models could perform better than just DL models like CNN, LSTMs and Phased LSTMs.<br />
<br />
=Related Work=<br />
===Time series forecasting===<br />
From recent proceedings in main machine learning venues i.e. ICML, NIPS, AISTATS, UAI, we can notice that time series are often forecast using Gaussian processes[3,4], especially for irregularly sampled time series[5]. Though still largely independent, combined models have started to appear, for example, the Gaussian Copula Process Volatility model[6]. For this paper, the authors use coupling AR models and neural networks to achieve such combined models.<br />
<br />
Although deep neural networks have been applied into many fields and produced satisfactory results, there still is little literature on deep learning for time series forecasting. More recently, the papers include Sirignano (2016)[7] that used 4-layer perceptrons in modeling price change distributions in Limit Order Books, and Borovykh et al. (2017)[8] who applied more recent WaveNet architecture to several short univariate and bivariate time-series (including financial ones). Heaton et al. (2016)[9] claimed to use autoencoders with a single hidden layer to compress multivariate financial data. Neil et al. (2016)[10] presented augmentation of LSTM architecture suitable for asynchronous series, which stimulates learning dependencies of different frequencies through time gate. <br />
<br />
In this paper, the authors examine the capabilities of several architectures (CNN, residual network, multi-layer LSTM, and phase LSTM) on AR-like artificial asynchronous and noisy time series, household electricity consumption dataset, and on real financial data from the credit default swap market with some inefficiencies.<br />
<br />
====AR Model====<br />
<br />
An autoregressive (AR) model describes the next value in a time-series as a combination of previous values, scaling factors, a bias, and noise [https://onlinecourses.science.psu.edu/stat501/node/358/ (source)]. For a p-th order (relating the current state to the p last states), the equation of the model is:<br />
<br />
<math> X_t = c + \sum_{i=1}^p \varphi_i X_{t-i}+ \varepsilon_t \,</math> [https://en.wikipedia.org/wiki/Autoregressive_model#Definition (equation source)]<br />
<br />
With parameters/coefficients <math>\varphi_i</math>, constant <math>c</math>, and noise <math>\varepsilon_t</math> This can be extended to vector form to create the VAR model mentioned in the paper.<br />
<br />
===Gating and weighting mechanisms===<br />
Gating mechanisms for neural networks has ability to overcome the problem of vanishing gradient, and can be expressed as <math display="inline">f(x)=c(x) \otimes \sigma(x)</math>, where <math>f</math> is the output function, <math>c</math> is a "candidate output" (a nonlinear function of <math>x</math>), <math>\otimes</math> is an element-wise matrix product, and <math>\sigma : \mathbb{R} \rightarrow [0,1] </math> is a sigmoid nonlinearity that controls the amount of output passed to the next layer. Different composition of functions of the same type as described above have proven to be an essential ingredient in popular recurrent architecture such as LSTM and GRU[11].<br />
<br />
The idea of the gating system is aimed to weight outputs of the intermediate layers within neural networks, and is most closely related to softmax gating used in MuFuRu(Multi-Function Recurrent Unit)[12], i.e.<br />
<math display="inline"> f(x) = \sum_{l=1}^L p^l(x) \otimes f^l(x), p(x)=softmax(\widehat{p}(x)), </math>, where <math>(f^l)_{l=1}^L </math>are candidate outputs(composition operators in MuFuRu), <math>(\widehat{p}^l)_{l=1}^L </math>are linear functions of inputs. <br />
<br />
This idea is also successfully used in attention networks[13] such as image captioning and machine translation. In this paper, the method is similar as this. The difference is that modeling the functions as multi-layer CNNs. Another difference is that not using recurrent layers, which can enable the network to remember the parts of the sentence/image already translated/described.<br />
<br />
=Motivation=<br />
There are mainly five motivations they stated in the paper:<br />
#The forecasting problem in this paper has done almost independently by econometrics and machine learning communities. Unlike in machine learning, research in econometrics are more likely to explain variables rather than improving out-of-sample prediction power. These models tend to 'over-fit' on financial time series, their parameters are unstable and have poor performance on out-of-sample prediction.<br />
#Although Gaussian processes provide a useful theoretical framework that is able to handle asynchronous data, they often follow heavy-tailed distribution for financial datasets.<br />
#Predictions of autoregressive time series may involve highly nonlinear functions if sampled irregularly. For AR time series with higher order and have more past observations, the expectation of it <math display="inline">\mathbb{E}[X(t)|{X(t-m), m=1,...,M}]</math> may involve more complicated functions that in general may not allow closed-form expression.<br />
#In practice, the dimensions of multivariate time series are often observed separately and asynchronously, such series at fixed frequency may lead to lose information or enlarge the dataset, which is shown in Figure 2(a). Therefore, the core of the proposed architecture SOCNN represents separate dimensions as a single one with dimension and duration indicators as additional features(Figure 2(b)).<br />
#Given a series of pairs of consecutive input values and corresponding durations, <math display="inline"> x_n = (X(t_n),t_n-t_{n-1}) </math>. One may expect that LSTM may memorize the input values in each step and weight them at the output according to the duration, but this approach may lead to an imbalance between the needs for memory and for linearity. The weights that are assigned to the memorized observations potentially require several layers of nonlinearity to be computed properly, while past observations might just need to be memorized as they are.<br />
<br />
[[File:Junyi2.png | 550px|thumb|center|Figure 2: (a) Fixed sampling frequency and its drawbacks; keep- ing all available information leads to much more datapoints. (b) Proposed data representation for the asynchronous series. Consecutive observations are stored together as a single value series, regardless of which series they belong to; this information, however, is stored in indicator features, alongside durations between observations.]]<br />
<br />
<br />
=Model Architecture=<br />
Suppose there's a multivariate time series <math display="inline">(x_n)_{n=0}^{\infty} \subset \mathbb{R}^d </math>, we want to predict the conditional future values of a subset of elements of <math>x_n</math><br />
<div style="text-align: center;"><math>y_n = \mathbb{E} [x_n^I | {x_{n-m}, m=1,2,...}], </math></div><br />
where <math> I=\{i_1,i_2,...i_{d_I}\} \subset \{1,2,...,d\} </math> is a subset of features of <math>x_n</math>.<br />
<br />
Let <math> \textbf{x}_n^{-M} = (x_{n-m})_{m=1}^M </math>. <br />
<br />
The estimator of <math>y_n</math> can be expressed as:<br />
<div style="text-align: center;"><math>\hat{y}_n = \sum_{m=1}^M [F(\textbf{x}_n^{-M}) \otimes \sigma(S(\textbf{x}_n^{-M}))].,_m ,</math></div><br />
The estimate is the summation of the columns of the matrix in bracket. Here<br />
#<math>F,S : \mathbb{R}^{d \times M} \rightarrow \mathbb{R}^{d_I \times M}</math> are neural networks. <br />
#* <math>S</math> is a fully convolutional network which is composed of convolutional layers only. <br />
#* <math display="inline">F(\textbf{x}_n^{-M}) = W \otimes [off(x_{n-m}) + x_{n-m}^I)]_{m=1}^M </math> <br />
#** <math> W \in \mathbb{R}^{d_I \times M}</math> <br />
#** <math> off: \mathbb{R}^d \rightarrow \mathbb{R}^{d_I} </math> is a multilayer perceptron.<br />
<br />
#<math>\sigma</math> is a normalized activation function independent at each row, i.e. <math display="inline"> \sigma ((a_1^T, ..., a_{d_I}^T)^T)=(\sigma(a_1)^T,..., \sigma(a_{d_I})^T)^T </math><br />
#* <math>a_{i} \in \mathbb{R}^{M}</math><br />
#* and <math>\sigma </math> is defined such that <math>\sigma(a)^{T} \mathbf{1}_{M}=1</math><br />
# <math>\otimes</math> is element-wise matrix multiplication.<br />
#<math>A.,_m</math> denotes the m-th column of a matrix A.<br />
<br />
Since <math>\sum_{m=1}^M W.,_m=W(1,1,...,1)^T</math> and <math>\sum_{m=1}^M S.,_m=S(1,1,...,1)^T</math>, we can express <math>\hat{y}_n</math> as:<br />
<div style="text-align: center;"><math>\hat{y}_n = \sum_{m=1}^M W.,_m \otimes (off(x_{n-m}) + x_{n-m}^I) \otimes \sigma(S.,_m(\textbf{x}_n^{-M}))</math></div><br />
This is the proposed network, Significance-Offset Convolutional Neural Network, <math>off</math> and <math>S</math> in the equation are corresponding to Offset and Significance in the name respectively.<br />
Figure 3 shows the scheme of network.<br />
<br />
[[File:Junyi3.png | 600px|thumb|center|Figure 3: A scheme of the proposed SOCNN architecture. The network preserves the time-dimension up to the top layer, while the number of features per timestep (filters) in the hidden layers is custom. The last convolutional layer, however, has the number of filters equal to dimension of the output. The Weighting frame shows how outputs from offset and significance networks are combined in accordance with Eq. of <math>\hat{y}_n</math>.]]<br />
<br />
The form of <math>\hat{y}_n</math> forced to separate the temporal dependence (obtained in weights <math>W_m</math>). S is determined by its filters which capture local dependencies and are independent of the relative position in time, the predictors <math>off(x_{n-m})</math> are completely independent of position in time. An adjusted single regressor for the target variable is provided by each past observation through the offset network. Since in asynchronous sampling procedure, consecutive values of x come from different signals and might be heterogeneous, therefore adjustment of offset network is important. In addition, significance network provides data-dependent weight for each regressor and sums them up in an autoregressive manner.<br />
<br />
===Relation to asynchronous data===<br />
One common problem of time series is that durations are varying between consecutive observations, the paper states two ways to solve this problem<br />
#Data preprocessing: aligning the observations at some fixed frequency e.g. duplicating and interpolating observations as shown in Figure 2(a). However, as mentioned in the figure, this approach will tend to loss of information and enlarge the size of the dataset and model complexity.<br />
#Add additional features: Treating the duration or time of the observations as additional features, it is the core of SOCNN, which is shown in Figure 2(b).<br />
<br />
===Loss function===<br />
The output of the offset network is series of separate predictors of changes between corresponding observations <math>x_{n-m}^I</math> and the target value<math>y_n</math>, this is the reason why we use auxiliary loss function, which equals to mean squared error of such intermediate predictions:<br />
<div style="text-align: center;"><math>L^{aux}(\textbf{x}_n^{-M}, y_n)=\frac{1}{M} \sum_{m=1}^M ||off(x_{n-m}) + x_{n-m}^I -y_n||^2 </math></div><br />
The total loss for the sample <math> \textbf{x}_n^{-M},y_n) </math> is then given by:<br />
<div style="text-align: center;"><math>L^{tot}(\textbf{x}_n^{-M}, y_n)=L^2(\widehat{y}_n, y_n)+\alpha L^{aux}(\textbf{x}_n^{-M}, y_n)</math></div><br />
where <math>\widehat{y}_n</math> was mentioned before, <math>\alpha \geq 0</math> is a constant.<br />
<br />
=Experiments=<br />
The paper evaluated SOCNN architecture on three datasets: artificially generated datasets, [https://archive.ics.uci.edu/ml/datasets/Individual+household+electric+power+consumption household electric power consumption dataset], and the financial dataset of bid/ask quotes sent by several market participants active in the credit derivatives market. Comparing its performance with simple CNN, single and multiplayer LSTM and 25-layer ResNet. Apart from the evaluation of the SOCNN architecture the paper also discusses the impact of network components such as: such as auxiliary<br />
loss and the depth of the offset sub-network. The code and datasets are available [https://github.com/mbinkowski/nntimeseries here]<br />
<br />
==Datasets==<br />
Artificial data: They generated 4 artificial series, <math> X_{K \times N}</math>, where <math>K \in \{16,64\} </math>. Therefore there is a synchronous and an asynchronous series for each K value.<br />
<br />
Electricity data: This UCI dataset contains 7 different features excluding date and time. The features include global active power, global reactive power, voltage, global intensity, sub-metering 1, sub-metering 2 and sub-metering 3, recorded every minute for 47 months. The data has been altered so that one observation contains only one value of 7 features, while durations between consecutive observations are ranged from 1 to 7 minutes. The goal is to predict all 7 features for the next time step.<br />
<br />
Non-anonymous quotes: The dataset contains 2.1 million quotes from 28 different sources from different market participants such as analysts, banks etc. Each quote is characterized by 31 features: the offered price, 28 indicators of the quoting source, the direction indicator (the quote refers to either a buy or a sell offer) and duration from the previous quote. For each source and direction, we want to predict the next quoted price from this given source and direction considering the last 60 quotes.<br />
<br />
==Training details==<br />
They applied grid search on some hyperparameters in order to get the significance of its components. The hyperparameters include the offset sub-network's depth and the auxiliary weight <math>\alpha</math>. For offset sub-network's depth, they use 1, 10,1 for artificial, electricity and quotes dataset respectively; and they compared the values of <math>\alpha</math> in {0,0.1,0.01}.<br />
<br />
They chose LeakyReLU as activation function for all networks:<br />
<div style="text-align: center;"><math>\sigma^{LeakyReLU}(x) = x</math> if <math>x\geq 0</math>, and <math>0.1x</math> otherwise </div><br />
They use the same number of layers, same stride and similar kernel size structure in CNN. In each trained CNN, they applied max pooling with the pool size of 2 every 2 convolutional layers.<br />
<br />
Table 1 presents the configuration of network hyperparameters used in comparison<br />
<br />
[[File:Junyi4.png | 400px|center|]]<br />
<br />
===Network Training===<br />
The training and validation data were sampled randomly from the first 80% of timesteps in each series, with ratio of 3 to 1. The remaining 20% of data was used as a test set.<br />
<br />
All models were trained using Adam optimizer because the authors found that its rate of convergence was much faster than standard Stochastic Gradient Descent in early tests.<br />
<br />
They used a batch size of 128 for artificial and electricity data, and 256 for quotes dataset, and applied batch normalization between each convolution and the following activation. <br />
<br />
At the beginning of each epoch, the training samples were randomly sampled. To prevent overfitting, they applied dropout and early stopping.<br />
<br />
Weights were initialized using the normalized uniform procedure proposed by Glorot & Bengio (2010).[14]<br />
<br />
The authors carried out the experiments on Tensorflow and Keras and used different GPU to optimize the model for different datasets.<br />
<br />
==Results==<br />
Table 2 shows all results performed from all datasets.<br />
[[File:Junyi5.png | 600px|center|]]<br />
We can see that SOCNN outperforms in all asynchronous artificial, electricity and quotes datasets. For synchronous data, LSTM might be slightly better, but SOCNN almost has the same results with LSTM. Phased LSTM and ResNet have performed really bad on artificial asynchronous dataset and quotes dataset respectively. Notice that having more than one layer of offset network would have negative impact on results. Also, the higher weights of auxiliary loss(<math>\alpha</math>considerably improved the test error on asynchronous dataset, see Table 3. However, for other datasets, its impact was negligible.<br />
[[File:Junyi6.png | 400px|center|]]<br />
In general, SOCNN has significantly lower variance of the test and validation errors, especially in the early stage of the training process and for quotes dataset. This effect can be seen in the learning curves for Asynchronous 64 artificial dataset presented in Figure 5.<br />
[[File:Junyi7.png | 500px|thumb|center|Figure 5: Learning curves with different auxiliary weights for SOCNN model trained on Asynchronous 64 dataset. The solid lines indicate the test error while the dashed lines indicate the training error.]]<br />
<br />
Finally, we want to test the robustness of the proposed model SOCNN, adding noise terms to asynchronous 16 dataset and check how these networks perform. The result is shown in Figure 6.<br />
[[File:Junyi8.png | 600px|thumb|center|Figure 6: Experiment comparing robustness of the considered networks for Asynchronous 16 dataset. The plots show how the error would change if an additional noise term was added to the input series. The dotted curves show the total significance and average absolute offset (not to scale) outputs for the noisy observations. Interestingly, the significance of the noisy observations increases with the magnitude of noise; i.e. noisy observations are far from being discarded by SOCNN.]]<br />
From Figure 6, the purple line and green line seems staying at the same position in training and testing process. SOCNN and single-layer LSTM are most robust compared to other networks, and least prone to overfitting.<br />
<br />
=Conclusion and Discussion=<br />
In this paper, the authors have proposed a new architecture called Significance-Offset Convolutional Neural Network, which combines AR-like weighting mechanism and convolutional neural network. This new architecture is designed for high-noise asynchronous time series and achieves outperformance in forecasting several asynchronous time series compared to popular convolutional and recurrent networks. <br />
<br />
The SOCNN can be extended further by adding intermediate weighting layers of the same type in the network structure. Another possible extension but needs further empirical studies is that we consider not just <math>1 \times 1</math> convolutional kernels on the offset sub-network. Also, this new architecture might be tested on other real-life datasets with relevant characteristics in the future, especially on econometric datasets and more generally for time series (stochastic processes) regression.<br />
<br />
=Critiques=<br />
#The paper is most likely an application paper, and the proposed new architecture shows improved performance over baselines in the asynchronous time series.<br />
#The quote data cannot be reached, only two datasets available.<br />
#The 'Significance' network was described as critical to the model in paper, but they did not show how the performance of SOCNN with respect to the significance network.<br />
#The transform of the original data to asynchronous data is not clear.<br />
#The experiments on the main application are not reproducible because the data is proprietary.<br />
#The way that train and test data were split is unclear. This could be important in the case of the financial data set.<br />
#Although the auxiliary loss function was mentioned as an important part, the advantages of it was not too clear in the paper. Maybe it is better that the paper describes a little more about its effectiveness.<br />
#It was not mentioned clearly in the paper whether the model training was done on a rolling basis for time series forecasting.<br />
#The noise term used in section 5's model robustness analysis uses evenly distributed noise (see Appendix B). While the analysis is a good start, analysis with different noise distributions would make the findings more generalizable.<br />
#The paper uses financial/economic data as one of its testing data set. Instead of comparing neural network models such as CNN which is known to work badly on time series data, it would be much better if the author compared to well-known econometric time series models such as GARCH and VAR.<br />
<br />
=References=<br />
[1] Hamilton, J. D. Time series analysis, volume 2. Princeton university press Princeton, 1994. <br />
<br />
[2] Fama, E. F. Efficient capital markets: A review of theory and empirical work. The journal of Finance, 25(2):383–417, 1970.<br />
<br />
[3] Petelin, D., Sˇindela ́ˇr, J., Pˇrikryl, J., and Kocijan, J. Financial modeling using gaussian process models. In Intelligent Data Acquisition and Advanced Computing Systems (IDAACS), 2011 IEEE 6th International Conference on, volume 2, pp. 672–677. IEEE, 2011.<br />
<br />
[4] Tobar, F., Bui, T. D., and Turner, R. E. Learning stationary time series using gaussian processes with nonparametric kernels. In Advances in Neural Information Processing Systems, pp. 3501–3509, 2015.<br />
<br />
[5] Hwang, Y., Tong, A., and Choi, J. Automatic construction of nonparametric relational regression models for multiple time series. In Proceedings of the 33rd International Conference on Machine Learning, 2016.<br />
<br />
[6] Wilson, A. and Ghahramani, Z. Copula processes. In Advances in Neural Information Processing Systems, pp. 2460–2468, 2010.<br />
<br />
[7] Sirignano, J. Extended abstract: Neural networks for limit order books, February 2016.<br />
<br />
[8] Borovykh, A., Bohte, S., and Oosterlee, C. W. Condi- tional time series forecasting with convolutional neural networks, March 2017.<br />
<br />
[9] Heaton, J. B., Polson, N. G., and Witte, J. H. Deep learn- ing in finance, February 2016.<br />
<br />
[10] Neil, D., Pfeiffer, M., and Liu, S.-C. Phased lstm: Acceler- ating recurrent network training for long or event-based sequences. In Advances In Neural Information Process- ing Systems, pp. 3882–3890, 2016.<br />
<br />
[11] Chung, J., Gulcehre, C., Cho, K., and Bengio, Y. Em- pirical evaluation of gated recurrent neural networks on sequence modeling, December 2014.<br />
<br />
[12] Weissenborn, D. and Rockta ̈schel, T. MuFuRU: The Multi-Function recurrent unit, June 2016.<br />
<br />
[13] Cho, K., Courville, A., and Bengio, Y. Describing multi- media content using attention-based Encoder–Decoder networks. IEEE Transactions on Multimedia, 17(11): 1875–1886, July 2015. ISSN 1520-9210.<br />
<br />
[14] Glorot, X. and Bengio, Y. Understanding the dif- ficulty of training deep feedforward neural net- works. In In Proceedings of the International Con- ference on Artificial Intelligence and Statistics (AIS- TATSaˆ10). Society for Artificial Intelligence and Statistics, 2010.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=stat946F18/Autoregressive_Convolutional_Neural_Networks_for_Asynchronous_Time_Series&diff=41752stat946F18/Autoregressive Convolutional Neural Networks for Asynchronous Time Series2018-11-28T19:50:24Z<p>Ka2khan: /* Gating and weighting mechanisms */</p>
<hr />
<div>This page is a summary of the paper "[http://proceedings.mlr.press/v80/binkowski18a/binkowski18a.pdf Autoregressive Convolutional Neural Networks for Asynchronous Time Series]" by Mikołaj Binkowski, Gautier Marti, Philippe Donnat. It was published at ICML in 2018. The code for this paper is provided [https://github.com/mbinkowski/nntimeseries here].<br />
<br />
=Introduction=<br />
In this paper, the authors propose a deep convolutional network architecture called Significance-Offset Convolutional Neural Network for regression of multivariate asynchronous time series. The model is inspired by standard autoregressive(AR) models and gating systems used in recurrent neural networks. The model is evaluated on various time series data including:<br />
# Hedge fund proprietary dataset of over 2 million quotes for a credit derivative index, <br />
# An artificially generated noisy auto-regressive series, <br />
# A UCI household electricity consumption dataset. <br />
<br />
This paper focuses on time series with multi-variate and noisy signals, especially financial data. Financial time series is challenging to predict due to their low signal-to-noise ratio and heavy-tailed distributions. For example, the same signal (e.g. price of a stock) is obtained from different sources (e.g. financial news, an investment bank, financial analyst etc.) asynchronously. Each source may have a different bias or noise. (Figure 1) The investment bank with more clients can update their information more precisely than the investment bank with fewer clients, then the significance of each past observations may depend on other factors that change in time. Therefore, the traditional econometric models such as AR, VAR, VARMA[1] might not be sufficient. However, their relatively good performance could allow us to combine such linear econometric models with deep neural networks that can learn highly nonlinear relationships. This model is inspired by the gating mechanism which is successful in RNNs and Highway Networks.<br />
<br />
The time series forecasting problem can be expressed as a conditional probability distribution below,<br />
<div style="text-align: center;"><math>p(X_{t+d}|X_t,X_{t-1},...) = f(X_t,X_{t-1},...)</math></div><br />
Thus, we focus on modeling the predictors of future values of time series given their past values. <br />
The predictability of financial dataset still remains an open problem and is discussed in various publications [2].<br />
<br />
[[File:Junyi1.png | 500px|thumb|center|Figure 1: Quotes from four different market participants (sources) for the same credit default swaps (CDS) throughout one day. Each trader displays from time to time the prices for which he offers to buy (bid) and sell (ask) the underlying CDS. The filled area marks the difference between the best sell and buy offers (spread) at each time.]]<br />
<br />
The paper also provides empirical evidence that their model which combines linear models with deep learning models could perform better than just DL models like CNN, LSTMs and Phased LSTMs.<br />
<br />
=Related Work=<br />
===Time series forecasting===<br />
From recent proceedings in main machine learning venues i.e. ICML, NIPS, AISTATS, UAI, we can notice that time series are often forecast using Gaussian processes[3,4], especially for irregularly sampled time series[5]. Though still largely independent, combined models have started to appear, for example, the Gaussian Copula Process Volatility model[6]. For this paper, the authors use coupling AR models and neural networks to achieve such combined models.<br />
<br />
Although deep neural networks have been applied into many fields and produced satisfactory results, there still is little literature on deep learning for time series forecasting. More recently, the papers include Sirignano (2016)[7] that used 4-layer perceptrons in modeling price change distributions in Limit Order Books, and Borovykh et al. (2017)[8] who applied more recent WaveNet architecture to several short univariate and bivariate time-series (including financial ones). Heaton et al. (2016)[9] claimed to use autoencoders with a single hidden layer to compress multivariate financial data. Neil et al. (2016)[10] presented augmentation of LSTM architecture suitable for asynchronous series, which stimulates learning dependencies of different frequencies through time gate. <br />
<br />
In this paper, the authors examine the capabilities of several architectures (CNN, residual network, multi-layer LSTM, and phase LSTM) on AR-like artificial asynchronous and noisy time series, household electricity consumption dataset, and on real financial data from the credit default swap market with some inefficiencies.<br />
<br />
====AR Model====<br />
<br />
An autoregressive (AR) model describes the next value in a time-series as a combination of previous values, scaling factors, a bias, and noise [https://onlinecourses.science.psu.edu/stat501/node/358/ (source)]. For a p-th order (relating the current state to the p last states), the equation of the model is:<br />
<br />
<math> X_t = c + \sum_{i=1}^p \varphi_i X_{t-i}+ \varepsilon_t \,</math> [https://en.wikipedia.org/wiki/Autoregressive_model#Definition (equation source)]<br />
<br />
With parameters/coefficients <math>\varphi_i</math>, constant <math>c</math>, and noise <math>\varepsilon_t</math> This can be extended to vector form to create the VAR model mentioned in the paper.<br />
<br />
===Gating and weighting mechanisms===<br />
Gating mechanisms for neural networks has ability to overcome the problem of vanishing gradient, and can be expressed as <math display="inline">f(x)=c(x) \otimes \sigma(x)</math>, where <math>f</math> is the output function, <math>c</math> is a "candidate output" (a nonlinear function of <math>x</math>), <math>\otimes</math> is an element-wise matrix product, and <math>\sigma : \mathbb{R} \rightarrow [0,1] </math> is a sigmoid nonlinearity that controls the amount of output passed to the next layer. This composition of functions have proven to be an essential ingredient in popular recurrent architecture such as LSTM and GRU[11].<br />
<br />
The idea of the gating system is aimed to weight outputs of the intermediate layers within neural networks, and is most closely related to softmax gating used in MuFuRu(Multi-Function Recurrent Unit)[12], i.e.<br />
<math display="inline"> f(x) = \sum_{l=1}^L p^l(x) \otimes f^l(x), p(x)=softmax(\widehat{p}(x)), </math>, where <math>(f^l)_{l=1}^L </math>are candidate outputs(composition operators in MuFuRu), <math>(\widehat{p}^l)_{l=1}^L </math>are linear functions of inputs. <br />
<br />
This idea is also successfully used in attention networks[13] such as image captioning and machine translation. In this paper, the method is similar as this. The difference is that modeling the functions as multi-layer CNNs. Another difference is that not using recurrent layers, which can enable the network to remember the parts of the sentence/image already translated/described.<br />
<br />
=Motivation=<br />
There are mainly five motivations they stated in the paper:<br />
#The forecasting problem in this paper has done almost independently by econometrics and machine learning communities. Unlike in machine learning, research in econometrics are more likely to explain variables rather than improving out-of-sample prediction power. These models tend to 'over-fit' on financial time series, their parameters are unstable and have poor performance on out-of-sample prediction.<br />
#Although Gaussian processes provide a useful theoretical framework that is able to handle asynchronous data, they often follow heavy-tailed distribution for financial datasets.<br />
#Predictions of autoregressive time series may involve highly nonlinear functions if sampled irregularly. For AR time series with higher order and have more past observations, the expectation of it <math display="inline">\mathbb{E}[X(t)|{X(t-m), m=1,...,M}]</math> may involve more complicated functions that in general may not allow closed-form expression.<br />
#In practice, the dimensions of multivariate time series are often observed separately and asynchronously, such series at fixed frequency may lead to lose information or enlarge the dataset, which is shown in Figure 2(a). Therefore, the core of the proposed architecture SOCNN represents separate dimensions as a single one with dimension and duration indicators as additional features(Figure 2(b)).<br />
#Given a series of pairs of consecutive input values and corresponding durations, <math display="inline"> x_n = (X(t_n),t_n-t_{n-1}) </math>. One may expect that LSTM may memorize the input values in each step and weight them at the output according to the duration, but this approach may lead to an imbalance between the needs for memory and for linearity. The weights that are assigned to the memorized observations potentially require several layers of nonlinearity to be computed properly, while past observations might just need to be memorized as they are.<br />
<br />
[[File:Junyi2.png | 550px|thumb|center|Figure 2: (a) Fixed sampling frequency and its drawbacks; keep- ing all available information leads to much more datapoints. (b) Proposed data representation for the asynchronous series. Consecutive observations are stored together as a single value series, regardless of which series they belong to; this information, however, is stored in indicator features, alongside durations between observations.]]<br />
<br />
<br />
=Model Architecture=<br />
Suppose there's a multivariate time series <math display="inline">(x_n)_{n=0}^{\infty} \subset \mathbb{R}^d </math>, we want to predict the conditional future values of a subset of elements of <math>x_n</math><br />
<div style="text-align: center;"><math>y_n = \mathbb{E} [x_n^I | {x_{n-m}, m=1,2,...}], </math></div><br />
where <math> I=\{i_1,i_2,...i_{d_I}\} \subset \{1,2,...,d\} </math> is a subset of features of <math>x_n</math>.<br />
<br />
Let <math> \textbf{x}_n^{-M} = (x_{n-m})_{m=1}^M </math>. <br />
<br />
The estimator of <math>y_n</math> can be expressed as:<br />
<div style="text-align: center;"><math>\hat{y}_n = \sum_{m=1}^M [F(\textbf{x}_n^{-M}) \otimes \sigma(S(\textbf{x}_n^{-M}))].,_m ,</math></div><br />
The estimate is the summation of the columns of the matrix in bracket. Here<br />
#<math>F,S : \mathbb{R}^{d \times M} \rightarrow \mathbb{R}^{d_I \times M}</math> are neural networks. <br />
#* <math>S</math> is a fully convolutional network which is composed of convolutional layers only. <br />
#* <math display="inline">F(\textbf{x}_n^{-M}) = W \otimes [off(x_{n-m}) + x_{n-m}^I)]_{m=1}^M </math> <br />
#** <math> W \in \mathbb{R}^{d_I \times M}</math> <br />
#** <math> off: \mathbb{R}^d \rightarrow \mathbb{R}^{d_I} </math> is a multilayer perceptron.<br />
<br />
#<math>\sigma</math> is a normalized activation function independent at each row, i.e. <math display="inline"> \sigma ((a_1^T, ..., a_{d_I}^T)^T)=(\sigma(a_1)^T,..., \sigma(a_{d_I})^T)^T </math><br />
#* <math>a_{i} \in \mathbb{R}^{M}</math><br />
#* and <math>\sigma </math> is defined such that <math>\sigma(a)^{T} \mathbf{1}_{M}=1</math><br />
# <math>\otimes</math> is element-wise matrix multiplication.<br />
#<math>A.,_m</math> denotes the m-th column of a matrix A.<br />
<br />
Since <math>\sum_{m=1}^M W.,_m=W(1,1,...,1)^T</math> and <math>\sum_{m=1}^M S.,_m=S(1,1,...,1)^T</math>, we can express <math>\hat{y}_n</math> as:<br />
<div style="text-align: center;"><math>\hat{y}_n = \sum_{m=1}^M W.,_m \otimes (off(x_{n-m}) + x_{n-m}^I) \otimes \sigma(S.,_m(\textbf{x}_n^{-M}))</math></div><br />
This is the proposed network, Significance-Offset Convolutional Neural Network, <math>off</math> and <math>S</math> in the equation are corresponding to Offset and Significance in the name respectively.<br />
Figure 3 shows the scheme of network.<br />
<br />
[[File:Junyi3.png | 600px|thumb|center|Figure 3: A scheme of the proposed SOCNN architecture. The network preserves the time-dimension up to the top layer, while the number of features per timestep (filters) in the hidden layers is custom. The last convolutional layer, however, has the number of filters equal to dimension of the output. The Weighting frame shows how outputs from offset and significance networks are combined in accordance with Eq. of <math>\hat{y}_n</math>.]]<br />
<br />
The form of <math>\hat{y}_n</math> forced to separate the temporal dependence (obtained in weights <math>W_m</math>). S is determined by its filters which capture local dependencies and are independent of the relative position in time, the predictors <math>off(x_{n-m})</math> are completely independent of position in time. An adjusted single regressor for the target variable is provided by each past observation through the offset network. Since in asynchronous sampling procedure, consecutive values of x come from different signals and might be heterogeneous, therefore adjustment of offset network is important. In addition, significance network provides data-dependent weight for each regressor and sums them up in an autoregressive manner.<br />
<br />
===Relation to asynchronous data===<br />
One common problem of time series is that durations are varying between consecutive observations, the paper states two ways to solve this problem<br />
#Data preprocessing: aligning the observations at some fixed frequency e.g. duplicating and interpolating observations as shown in Figure 2(a). However, as mentioned in the figure, this approach will tend to loss of information and enlarge the size of the dataset and model complexity.<br />
#Add additional features: Treating the duration or time of the observations as additional features, it is the core of SOCNN, which is shown in Figure 2(b).<br />
<br />
===Loss function===<br />
The output of the offset network is series of separate predictors of changes between corresponding observations <math>x_{n-m}^I</math> and the target value<math>y_n</math>, this is the reason why we use auxiliary loss function, which equals to mean squared error of such intermediate predictions:<br />
<div style="text-align: center;"><math>L^{aux}(\textbf{x}_n^{-M}, y_n)=\frac{1}{M} \sum_{m=1}^M ||off(x_{n-m}) + x_{n-m}^I -y_n||^2 </math></div><br />
The total loss for the sample <math> \textbf{x}_n^{-M},y_n) </math> is then given by:<br />
<div style="text-align: center;"><math>L^{tot}(\textbf{x}_n^{-M}, y_n)=L^2(\widehat{y}_n, y_n)+\alpha L^{aux}(\textbf{x}_n^{-M}, y_n)</math></div><br />
where <math>\widehat{y}_n</math> was mentioned before, <math>\alpha \geq 0</math> is a constant.<br />
<br />
=Experiments=<br />
The paper evaluated SOCNN architecture on three datasets: artificially generated datasets, [https://archive.ics.uci.edu/ml/datasets/Individual+household+electric+power+consumption household electric power consumption dataset], and the financial dataset of bid/ask quotes sent by several market participants active in the credit derivatives market. Comparing its performance with simple CNN, single and multiplayer LSTM and 25-layer ResNet. Apart from the evaluation of the SOCNN architecture the paper also discusses the impact of network components such as: such as auxiliary<br />
loss and the depth of the offset sub-network. The code and datasets are available [https://github.com/mbinkowski/nntimeseries here]<br />
<br />
==Datasets==<br />
Artificial data: They generated 4 artificial series, <math> X_{K \times N}</math>, where <math>K \in \{16,64\} </math>. Therefore there is a synchronous and an asynchronous series for each K value.<br />
<br />
Electricity data: This UCI dataset contains 7 different features excluding date and time. The features include global active power, global reactive power, voltage, global intensity, sub-metering 1, sub-metering 2 and sub-metering 3, recorded every minute for 47 months. The data has been altered so that one observation contains only one value of 7 features, while durations between consecutive observations are ranged from 1 to 7 minutes. The goal is to predict all 7 features for the next time step.<br />
<br />
Non-anonymous quotes: The dataset contains 2.1 million quotes from 28 different sources from different market participants such as analysts, banks etc. Each quote is characterized by 31 features: the offered price, 28 indicators of the quoting source, the direction indicator (the quote refers to either a buy or a sell offer) and duration from the previous quote. For each source and direction, we want to predict the next quoted price from this given source and direction considering the last 60 quotes.<br />
<br />
==Training details==<br />
They applied grid search on some hyperparameters in order to get the significance of its components. The hyperparameters include the offset sub-network's depth and the auxiliary weight <math>\alpha</math>. For offset sub-network's depth, they use 1, 10,1 for artificial, electricity and quotes dataset respectively; and they compared the values of <math>\alpha</math> in {0,0.1,0.01}.<br />
<br />
They chose LeakyReLU as activation function for all networks:<br />
<div style="text-align: center;"><math>\sigma^{LeakyReLU}(x) = x</math> if <math>x\geq 0</math>, and <math>0.1x</math> otherwise </div><br />
They use the same number of layers, same stride and similar kernel size structure in CNN. In each trained CNN, they applied max pooling with the pool size of 2 every 2 convolutional layers.<br />
<br />
Table 1 presents the configuration of network hyperparameters used in comparison<br />
<br />
[[File:Junyi4.png | 400px|center|]]<br />
<br />
===Network Training===<br />
The training and validation data were sampled randomly from the first 80% of timesteps in each series, with ratio of 3 to 1. The remaining 20% of data was used as a test set.<br />
<br />
All models were trained using Adam optimizer because the authors found that its rate of convergence was much faster than standard Stochastic Gradient Descent in early tests.<br />
<br />
They used a batch size of 128 for artificial and electricity data, and 256 for quotes dataset, and applied batch normalization between each convolution and the following activation. <br />
<br />
At the beginning of each epoch, the training samples were randomly sampled. To prevent overfitting, they applied dropout and early stopping.<br />
<br />
Weights were initialized using the normalized uniform procedure proposed by Glorot & Bengio (2010).[14]<br />
<br />
The authors carried out the experiments on Tensorflow and Keras and used different GPU to optimize the model for different datasets.<br />
<br />
==Results==<br />
Table 2 shows all results performed from all datasets.<br />
[[File:Junyi5.png | 600px|center|]]<br />
We can see that SOCNN outperforms in all asynchronous artificial, electricity and quotes datasets. For synchronous data, LSTM might be slightly better, but SOCNN almost has the same results with LSTM. Phased LSTM and ResNet have performed really bad on artificial asynchronous dataset and quotes dataset respectively. Notice that having more than one layer of offset network would have negative impact on results. Also, the higher weights of auxiliary loss(<math>\alpha</math>considerably improved the test error on asynchronous dataset, see Table 3. However, for other datasets, its impact was negligible.<br />
[[File:Junyi6.png | 400px|center|]]<br />
In general, SOCNN has significantly lower variance of the test and validation errors, especially in the early stage of the training process and for quotes dataset. This effect can be seen in the learning curves for Asynchronous 64 artificial dataset presented in Figure 5.<br />
[[File:Junyi7.png | 500px|thumb|center|Figure 5: Learning curves with different auxiliary weights for SOCNN model trained on Asynchronous 64 dataset. The solid lines indicate the test error while the dashed lines indicate the training error.]]<br />
<br />
Finally, we want to test the robustness of the proposed model SOCNN, adding noise terms to asynchronous 16 dataset and check how these networks perform. The result is shown in Figure 6.<br />
[[File:Junyi8.png | 600px|thumb|center|Figure 6: Experiment comparing robustness of the considered networks for Asynchronous 16 dataset. The plots show how the error would change if an additional noise term was added to the input series. The dotted curves show the total significance and average absolute offset (not to scale) outputs for the noisy observations. Interestingly, the significance of the noisy observations increases with the magnitude of noise; i.e. noisy observations are far from being discarded by SOCNN.]]<br />
From Figure 6, the purple line and green line seems staying at the same position in training and testing process. SOCNN and single-layer LSTM are most robust compared to other networks, and least prone to overfitting.<br />
<br />
=Conclusion and Discussion=<br />
In this paper, the authors have proposed a new architecture called Significance-Offset Convolutional Neural Network, which combines AR-like weighting mechanism and convolutional neural network. This new architecture is designed for high-noise asynchronous time series and achieves outperformance in forecasting several asynchronous time series compared to popular convolutional and recurrent networks. <br />
<br />
The SOCNN can be extended further by adding intermediate weighting layers of the same type in the network structure. Another possible extension but needs further empirical studies is that we consider not just <math>1 \times 1</math> convolutional kernels on the offset sub-network. Also, this new architecture might be tested on other real-life datasets with relevant characteristics in the future, especially on econometric datasets and more generally for time series (stochastic processes) regression.<br />
<br />
=Critiques=<br />
#The paper is most likely an application paper, and the proposed new architecture shows improved performance over baselines in the asynchronous time series.<br />
#The quote data cannot be reached, only two datasets available.<br />
#The 'Significance' network was described as critical to the model in paper, but they did not show how the performance of SOCNN with respect to the significance network.<br />
#The transform of the original data to asynchronous data is not clear.<br />
#The experiments on the main application are not reproducible because the data is proprietary.<br />
#The way that train and test data were split is unclear. This could be important in the case of the financial data set.<br />
#Although the auxiliary loss function was mentioned as an important part, the advantages of it was not too clear in the paper. Maybe it is better that the paper describes a little more about its effectiveness.<br />
#It was not mentioned clearly in the paper whether the model training was done on a rolling basis for time series forecasting.<br />
#The noise term used in section 5's model robustness analysis uses evenly distributed noise (see Appendix B). While the analysis is a good start, analysis with different noise distributions would make the findings more generalizable.<br />
#The paper uses financial/economic data as one of its testing data set. Instead of comparing neural network models such as CNN which is known to work badly on time series data, it would be much better if the author compared to well-known econometric time series models such as GARCH and VAR.<br />
<br />
=References=<br />
[1] Hamilton, J. D. Time series analysis, volume 2. Princeton university press Princeton, 1994. <br />
<br />
[2] Fama, E. F. Efficient capital markets: A review of theory and empirical work. The journal of Finance, 25(2):383–417, 1970.<br />
<br />
[3] Petelin, D., Sˇindela ́ˇr, J., Pˇrikryl, J., and Kocijan, J. Financial modeling using gaussian process models. In Intelligent Data Acquisition and Advanced Computing Systems (IDAACS), 2011 IEEE 6th International Conference on, volume 2, pp. 672–677. IEEE, 2011.<br />
<br />
[4] Tobar, F., Bui, T. D., and Turner, R. E. Learning stationary time series using gaussian processes with nonparametric kernels. In Advances in Neural Information Processing Systems, pp. 3501–3509, 2015.<br />
<br />
[5] Hwang, Y., Tong, A., and Choi, J. Automatic construction of nonparametric relational regression models for multiple time series. In Proceedings of the 33rd International Conference on Machine Learning, 2016.<br />
<br />
[6] Wilson, A. and Ghahramani, Z. Copula processes. In Advances in Neural Information Processing Systems, pp. 2460–2468, 2010.<br />
<br />
[7] Sirignano, J. Extended abstract: Neural networks for limit order books, February 2016.<br />
<br />
[8] Borovykh, A., Bohte, S., and Oosterlee, C. W. Condi- tional time series forecasting with convolutional neural networks, March 2017.<br />
<br />
[9] Heaton, J. B., Polson, N. G., and Witte, J. H. Deep learn- ing in finance, February 2016.<br />
<br />
[10] Neil, D., Pfeiffer, M., and Liu, S.-C. Phased lstm: Acceler- ating recurrent network training for long or event-based sequences. In Advances In Neural Information Process- ing Systems, pp. 3882–3890, 2016.<br />
<br />
[11] Chung, J., Gulcehre, C., Cho, K., and Bengio, Y. Em- pirical evaluation of gated recurrent neural networks on sequence modeling, December 2014.<br />
<br />
[12] Weissenborn, D. and Rockta ̈schel, T. MuFuRU: The Multi-Function recurrent unit, June 2016.<br />
<br />
[13] Cho, K., Courville, A., and Bengio, Y. Describing multi- media content using attention-based Encoder–Decoder networks. IEEE Transactions on Multimedia, 17(11): 1875–1886, July 2015. ISSN 1520-9210.<br />
<br />
[14] Glorot, X. and Bengio, Y. Understanding the dif- ficulty of training deep feedforward neural net- works. In In Proceedings of the International Con- ference on Artificial Intelligence and Statistics (AIS- TATSaˆ10). Society for Artificial Intelligence and Statistics, 2010.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=stat946F18/Autoregressive_Convolutional_Neural_Networks_for_Asynchronous_Time_Series&diff=41751stat946F18/Autoregressive Convolutional Neural Networks for Asynchronous Time Series2018-11-28T19:48:46Z<p>Ka2khan: /* AR Model */</p>
<hr />
<div>This page is a summary of the paper "[http://proceedings.mlr.press/v80/binkowski18a/binkowski18a.pdf Autoregressive Convolutional Neural Networks for Asynchronous Time Series]" by Mikołaj Binkowski, Gautier Marti, Philippe Donnat. It was published at ICML in 2018. The code for this paper is provided [https://github.com/mbinkowski/nntimeseries here].<br />
<br />
=Introduction=<br />
In this paper, the authors propose a deep convolutional network architecture called Significance-Offset Convolutional Neural Network for regression of multivariate asynchronous time series. The model is inspired by standard autoregressive(AR) models and gating systems used in recurrent neural networks. The model is evaluated on various time series data including:<br />
# Hedge fund proprietary dataset of over 2 million quotes for a credit derivative index, <br />
# An artificially generated noisy auto-regressive series, <br />
# A UCI household electricity consumption dataset. <br />
<br />
This paper focuses on time series with multi-variate and noisy signals, especially financial data. Financial time series is challenging to predict due to their low signal-to-noise ratio and heavy-tailed distributions. For example, the same signal (e.g. price of a stock) is obtained from different sources (e.g. financial news, an investment bank, financial analyst etc.) asynchronously. Each source may have a different bias or noise. (Figure 1) The investment bank with more clients can update their information more precisely than the investment bank with fewer clients, then the significance of each past observations may depend on other factors that change in time. Therefore, the traditional econometric models such as AR, VAR, VARMA[1] might not be sufficient. However, their relatively good performance could allow us to combine such linear econometric models with deep neural networks that can learn highly nonlinear relationships. This model is inspired by the gating mechanism which is successful in RNNs and Highway Networks.<br />
<br />
The time series forecasting problem can be expressed as a conditional probability distribution below,<br />
<div style="text-align: center;"><math>p(X_{t+d}|X_t,X_{t-1},...) = f(X_t,X_{t-1},...)</math></div><br />
Thus, we focus on modeling the predictors of future values of time series given their past values. <br />
The predictability of financial dataset still remains an open problem and is discussed in various publications [2].<br />
<br />
[[File:Junyi1.png | 500px|thumb|center|Figure 1: Quotes from four different market participants (sources) for the same credit default swaps (CDS) throughout one day. Each trader displays from time to time the prices for which he offers to buy (bid) and sell (ask) the underlying CDS. The filled area marks the difference between the best sell and buy offers (spread) at each time.]]<br />
<br />
The paper also provides empirical evidence that their model which combines linear models with deep learning models could perform better than just DL models like CNN, LSTMs and Phased LSTMs.<br />
<br />
=Related Work=<br />
===Time series forecasting===<br />
From recent proceedings in main machine learning venues i.e. ICML, NIPS, AISTATS, UAI, we can notice that time series are often forecast using Gaussian processes[3,4], especially for irregularly sampled time series[5]. Though still largely independent, combined models have started to appear, for example, the Gaussian Copula Process Volatility model[6]. For this paper, the authors use coupling AR models and neural networks to achieve such combined models.<br />
<br />
Although deep neural networks have been applied into many fields and produced satisfactory results, there still is little literature on deep learning for time series forecasting. More recently, the papers include Sirignano (2016)[7] that used 4-layer perceptrons in modeling price change distributions in Limit Order Books, and Borovykh et al. (2017)[8] who applied more recent WaveNet architecture to several short univariate and bivariate time-series (including financial ones). Heaton et al. (2016)[9] claimed to use autoencoders with a single hidden layer to compress multivariate financial data. Neil et al. (2016)[10] presented augmentation of LSTM architecture suitable for asynchronous series, which stimulates learning dependencies of different frequencies through time gate. <br />
<br />
In this paper, the authors examine the capabilities of several architectures (CNN, residual network, multi-layer LSTM, and phase LSTM) on AR-like artificial asynchronous and noisy time series, household electricity consumption dataset, and on real financial data from the credit default swap market with some inefficiencies.<br />
<br />
====AR Model====<br />
<br />
An autoregressive (AR) model describes the next value in a time-series as a combination of previous values, scaling factors, a bias, and noise [https://onlinecourses.science.psu.edu/stat501/node/358/ (source)]. For a p-th order (relating the current state to the p last states), the equation of the model is:<br />
<br />
<math> X_t = c + \sum_{i=1}^p \varphi_i X_{t-i}+ \varepsilon_t \,</math> [https://en.wikipedia.org/wiki/Autoregressive_model#Definition (equation source)]<br />
<br />
With parameters/coefficients <math>\varphi_i</math>, constant <math>c</math>, and noise <math>\varepsilon_t</math> This can be extended to vector form to create the VAR model mentioned in the paper.<br />
<br />
===Gating and weighting mechanisms===<br />
Gating mechanisms for neural networks has ability to overcome the problem of vanishing gradient, and can be expressed as <math display="inline">f(x)=c(x) \otimes \sigma(x)</math>, where <math>f</math> is the output function, <math>c</math> is a "candidate output" (a nonlinear function of <math>x</math>), <math>\otimes</math> is an element-wise matrix product, and <math>\sigma : \mathbb{R} \rightarrow [0,1] </math> is a sigmoid nonlinearity that controls the amount of output passed to the next layer. This composition of functions may lead to popular recurrent architecture such as LSTM and GRU[11].<br />
<br />
The idea of the gating system is aimed to weight outputs of the intermediate layers within neural networks, and is most closely related to softmax gating used in MuFuRu(Multi-Function Recurrent Unit)[12], i.e.<br />
<math display="inline"> f(x) = \sum_{l=1}^L p^l(x) \otimes f^l(x), p(x)=softmax(\widehat{p}(x)), </math>, where <math>(f^l)_{l=1}^L </math>are candidate outputs(composition operators in MuFuRu), <math>(\widehat{p}^l)_{l=1}^L </math>are linear functions of inputs. <br />
<br />
This idea is also successfully used in attention networks[13] such as image captioning and machine translation. In this paper, the method is similar as this. The difference is that modeling the functions as multi-layer CNNs. Another difference is that not using recurrent layers, which can enable the network to remember the parts of the sentence/image already translated/described.<br />
<br />
=Motivation=<br />
There are mainly five motivations they stated in the paper:<br />
#The forecasting problem in this paper has done almost independently by econometrics and machine learning communities. Unlike in machine learning, research in econometrics are more likely to explain variables rather than improving out-of-sample prediction power. These models tend to 'over-fit' on financial time series, their parameters are unstable and have poor performance on out-of-sample prediction.<br />
#Although Gaussian processes provide a useful theoretical framework that is able to handle asynchronous data, they often follow heavy-tailed distribution for financial datasets.<br />
#Predictions of autoregressive time series may involve highly nonlinear functions if sampled irregularly. For AR time series with higher order and have more past observations, the expectation of it <math display="inline">\mathbb{E}[X(t)|{X(t-m), m=1,...,M}]</math> may involve more complicated functions that in general may not allow closed-form expression.<br />
#In practice, the dimensions of multivariate time series are often observed separately and asynchronously, such series at fixed frequency may lead to lose information or enlarge the dataset, which is shown in Figure 2(a). Therefore, the core of the proposed architecture SOCNN represents separate dimensions as a single one with dimension and duration indicators as additional features(Figure 2(b)).<br />
#Given a series of pairs of consecutive input values and corresponding durations, <math display="inline"> x_n = (X(t_n),t_n-t_{n-1}) </math>. One may expect that LSTM may memorize the input values in each step and weight them at the output according to the duration, but this approach may lead to an imbalance between the needs for memory and for linearity. The weights that are assigned to the memorized observations potentially require several layers of nonlinearity to be computed properly, while past observations might just need to be memorized as they are.<br />
<br />
[[File:Junyi2.png | 550px|thumb|center|Figure 2: (a) Fixed sampling frequency and its drawbacks; keep- ing all available information leads to much more datapoints. (b) Proposed data representation for the asynchronous series. Consecutive observations are stored together as a single value series, regardless of which series they belong to; this information, however, is stored in indicator features, alongside durations between observations.]]<br />
<br />
<br />
=Model Architecture=<br />
Suppose there's a multivariate time series <math display="inline">(x_n)_{n=0}^{\infty} \subset \mathbb{R}^d </math>, we want to predict the conditional future values of a subset of elements of <math>x_n</math><br />
<div style="text-align: center;"><math>y_n = \mathbb{E} [x_n^I | {x_{n-m}, m=1,2,...}], </math></div><br />
where <math> I=\{i_1,i_2,...i_{d_I}\} \subset \{1,2,...,d\} </math> is a subset of features of <math>x_n</math>.<br />
<br />
Let <math> \textbf{x}_n^{-M} = (x_{n-m})_{m=1}^M </math>. <br />
<br />
The estimator of <math>y_n</math> can be expressed as:<br />
<div style="text-align: center;"><math>\hat{y}_n = \sum_{m=1}^M [F(\textbf{x}_n^{-M}) \otimes \sigma(S(\textbf{x}_n^{-M}))].,_m ,</math></div><br />
The estimate is the summation of the columns of the matrix in bracket. Here<br />
#<math>F,S : \mathbb{R}^{d \times M} \rightarrow \mathbb{R}^{d_I \times M}</math> are neural networks. <br />
#* <math>S</math> is a fully convolutional network which is composed of convolutional layers only. <br />
#* <math display="inline">F(\textbf{x}_n^{-M}) = W \otimes [off(x_{n-m}) + x_{n-m}^I)]_{m=1}^M </math> <br />
#** <math> W \in \mathbb{R}^{d_I \times M}</math> <br />
#** <math> off: \mathbb{R}^d \rightarrow \mathbb{R}^{d_I} </math> is a multilayer perceptron.<br />
<br />
#<math>\sigma</math> is a normalized activation function independent at each row, i.e. <math display="inline"> \sigma ((a_1^T, ..., a_{d_I}^T)^T)=(\sigma(a_1)^T,..., \sigma(a_{d_I})^T)^T </math><br />
#* <math>a_{i} \in \mathbb{R}^{M}</math><br />
#* and <math>\sigma </math> is defined such that <math>\sigma(a)^{T} \mathbf{1}_{M}=1</math><br />
# <math>\otimes</math> is element-wise matrix multiplication.<br />
#<math>A.,_m</math> denotes the m-th column of a matrix A.<br />
<br />
Since <math>\sum_{m=1}^M W.,_m=W(1,1,...,1)^T</math> and <math>\sum_{m=1}^M S.,_m=S(1,1,...,1)^T</math>, we can express <math>\hat{y}_n</math> as:<br />
<div style="text-align: center;"><math>\hat{y}_n = \sum_{m=1}^M W.,_m \otimes (off(x_{n-m}) + x_{n-m}^I) \otimes \sigma(S.,_m(\textbf{x}_n^{-M}))</math></div><br />
This is the proposed network, Significance-Offset Convolutional Neural Network, <math>off</math> and <math>S</math> in the equation are corresponding to Offset and Significance in the name respectively.<br />
Figure 3 shows the scheme of network.<br />
<br />
[[File:Junyi3.png | 600px|thumb|center|Figure 3: A scheme of the proposed SOCNN architecture. The network preserves the time-dimension up to the top layer, while the number of features per timestep (filters) in the hidden layers is custom. The last convolutional layer, however, has the number of filters equal to dimension of the output. The Weighting frame shows how outputs from offset and significance networks are combined in accordance with Eq. of <math>\hat{y}_n</math>.]]<br />
<br />
The form of <math>\hat{y}_n</math> forced to separate the temporal dependence (obtained in weights <math>W_m</math>). S is determined by its filters which capture local dependencies and are independent of the relative position in time, the predictors <math>off(x_{n-m})</math> are completely independent of position in time. An adjusted single regressor for the target variable is provided by each past observation through the offset network. Since in asynchronous sampling procedure, consecutive values of x come from different signals and might be heterogeneous, therefore adjustment of offset network is important. In addition, significance network provides data-dependent weight for each regressor and sums them up in an autoregressive manner.<br />
<br />
===Relation to asynchronous data===<br />
One common problem of time series is that durations are varying between consecutive observations, the paper states two ways to solve this problem<br />
#Data preprocessing: aligning the observations at some fixed frequency e.g. duplicating and interpolating observations as shown in Figure 2(a). However, as mentioned in the figure, this approach will tend to loss of information and enlarge the size of the dataset and model complexity.<br />
#Add additional features: Treating the duration or time of the observations as additional features, it is the core of SOCNN, which is shown in Figure 2(b).<br />
<br />
===Loss function===<br />
The output of the offset network is series of separate predictors of changes between corresponding observations <math>x_{n-m}^I</math> and the target value<math>y_n</math>, this is the reason why we use auxiliary loss function, which equals to mean squared error of such intermediate predictions:<br />
<div style="text-align: center;"><math>L^{aux}(\textbf{x}_n^{-M}, y_n)=\frac{1}{M} \sum_{m=1}^M ||off(x_{n-m}) + x_{n-m}^I -y_n||^2 </math></div><br />
The total loss for the sample <math> \textbf{x}_n^{-M},y_n) </math> is then given by:<br />
<div style="text-align: center;"><math>L^{tot}(\textbf{x}_n^{-M}, y_n)=L^2(\widehat{y}_n, y_n)+\alpha L^{aux}(\textbf{x}_n^{-M}, y_n)</math></div><br />
where <math>\widehat{y}_n</math> was mentioned before, <math>\alpha \geq 0</math> is a constant.<br />
<br />
=Experiments=<br />
The paper evaluated SOCNN architecture on three datasets: artificially generated datasets, [https://archive.ics.uci.edu/ml/datasets/Individual+household+electric+power+consumption household electric power consumption dataset], and the financial dataset of bid/ask quotes sent by several market participants active in the credit derivatives market. Comparing its performance with simple CNN, single and multiplayer LSTM and 25-layer ResNet. Apart from the evaluation of the SOCNN architecture the paper also discusses the impact of network components such as: such as auxiliary<br />
loss and the depth of the offset sub-network. The code and datasets are available [https://github.com/mbinkowski/nntimeseries here]<br />
<br />
==Datasets==<br />
Artificial data: They generated 4 artificial series, <math> X_{K \times N}</math>, where <math>K \in \{16,64\} </math>. Therefore there is a synchronous and an asynchronous series for each K value.<br />
<br />
Electricity data: This UCI dataset contains 7 different features excluding date and time. The features include global active power, global reactive power, voltage, global intensity, sub-metering 1, sub-metering 2 and sub-metering 3, recorded every minute for 47 months. The data has been altered so that one observation contains only one value of 7 features, while durations between consecutive observations are ranged from 1 to 7 minutes. The goal is to predict all 7 features for the next time step.<br />
<br />
Non-anonymous quotes: The dataset contains 2.1 million quotes from 28 different sources from different market participants such as analysts, banks etc. Each quote is characterized by 31 features: the offered price, 28 indicators of the quoting source, the direction indicator (the quote refers to either a buy or a sell offer) and duration from the previous quote. For each source and direction, we want to predict the next quoted price from this given source and direction considering the last 60 quotes.<br />
<br />
==Training details==<br />
They applied grid search on some hyperparameters in order to get the significance of its components. The hyperparameters include the offset sub-network's depth and the auxiliary weight <math>\alpha</math>. For offset sub-network's depth, they use 1, 10,1 for artificial, electricity and quotes dataset respectively; and they compared the values of <math>\alpha</math> in {0,0.1,0.01}.<br />
<br />
They chose LeakyReLU as activation function for all networks:<br />
<div style="text-align: center;"><math>\sigma^{LeakyReLU}(x) = x</math> if <math>x\geq 0</math>, and <math>0.1x</math> otherwise </div><br />
They use the same number of layers, same stride and similar kernel size structure in CNN. In each trained CNN, they applied max pooling with the pool size of 2 every 2 convolutional layers.<br />
<br />
Table 1 presents the configuration of network hyperparameters used in comparison<br />
<br />
[[File:Junyi4.png | 400px|center|]]<br />
<br />
===Network Training===<br />
The training and validation data were sampled randomly from the first 80% of timesteps in each series, with ratio of 3 to 1. The remaining 20% of data was used as a test set.<br />
<br />
All models were trained using Adam optimizer because the authors found that its rate of convergence was much faster than standard Stochastic Gradient Descent in early tests.<br />
<br />
They used a batch size of 128 for artificial and electricity data, and 256 for quotes dataset, and applied batch normalization between each convolution and the following activation. <br />
<br />
At the beginning of each epoch, the training samples were randomly sampled. To prevent overfitting, they applied dropout and early stopping.<br />
<br />
Weights were initialized using the normalized uniform procedure proposed by Glorot & Bengio (2010).[14]<br />
<br />
The authors carried out the experiments on Tensorflow and Keras and used different GPU to optimize the model for different datasets.<br />
<br />
==Results==<br />
Table 2 shows all results performed from all datasets.<br />
[[File:Junyi5.png | 600px|center|]]<br />
We can see that SOCNN outperforms in all asynchronous artificial, electricity and quotes datasets. For synchronous data, LSTM might be slightly better, but SOCNN almost has the same results with LSTM. Phased LSTM and ResNet have performed really bad on artificial asynchronous dataset and quotes dataset respectively. Notice that having more than one layer of offset network would have negative impact on results. Also, the higher weights of auxiliary loss(<math>\alpha</math>considerably improved the test error on asynchronous dataset, see Table 3. However, for other datasets, its impact was negligible.<br />
[[File:Junyi6.png | 400px|center|]]<br />
In general, SOCNN has significantly lower variance of the test and validation errors, especially in the early stage of the training process and for quotes dataset. This effect can be seen in the learning curves for Asynchronous 64 artificial dataset presented in Figure 5.<br />
[[File:Junyi7.png | 500px|thumb|center|Figure 5: Learning curves with different auxiliary weights for SOCNN model trained on Asynchronous 64 dataset. The solid lines indicate the test error while the dashed lines indicate the training error.]]<br />
<br />
Finally, we want to test the robustness of the proposed model SOCNN, adding noise terms to asynchronous 16 dataset and check how these networks perform. The result is shown in Figure 6.<br />
[[File:Junyi8.png | 600px|thumb|center|Figure 6: Experiment comparing robustness of the considered networks for Asynchronous 16 dataset. The plots show how the error would change if an additional noise term was added to the input series. The dotted curves show the total significance and average absolute offset (not to scale) outputs for the noisy observations. Interestingly, the significance of the noisy observations increases with the magnitude of noise; i.e. noisy observations are far from being discarded by SOCNN.]]<br />
From Figure 6, the purple line and green line seems staying at the same position in training and testing process. SOCNN and single-layer LSTM are most robust compared to other networks, and least prone to overfitting.<br />
<br />
=Conclusion and Discussion=<br />
In this paper, the authors have proposed a new architecture called Significance-Offset Convolutional Neural Network, which combines AR-like weighting mechanism and convolutional neural network. This new architecture is designed for high-noise asynchronous time series and achieves outperformance in forecasting several asynchronous time series compared to popular convolutional and recurrent networks. <br />
<br />
The SOCNN can be extended further by adding intermediate weighting layers of the same type in the network structure. Another possible extension but needs further empirical studies is that we consider not just <math>1 \times 1</math> convolutional kernels on the offset sub-network. Also, this new architecture might be tested on other real-life datasets with relevant characteristics in the future, especially on econometric datasets and more generally for time series (stochastic processes) regression.<br />
<br />
=Critiques=<br />
#The paper is most likely an application paper, and the proposed new architecture shows improved performance over baselines in the asynchronous time series.<br />
#The quote data cannot be reached, only two datasets available.<br />
#The 'Significance' network was described as critical to the model in paper, but they did not show how the performance of SOCNN with respect to the significance network.<br />
#The transform of the original data to asynchronous data is not clear.<br />
#The experiments on the main application are not reproducible because the data is proprietary.<br />
#The way that train and test data were split is unclear. This could be important in the case of the financial data set.<br />
#Although the auxiliary loss function was mentioned as an important part, the advantages of it was not too clear in the paper. Maybe it is better that the paper describes a little more about its effectiveness.<br />
#It was not mentioned clearly in the paper whether the model training was done on a rolling basis for time series forecasting.<br />
#The noise term used in section 5's model robustness analysis uses evenly distributed noise (see Appendix B). While the analysis is a good start, analysis with different noise distributions would make the findings more generalizable.<br />
#The paper uses financial/economic data as one of its testing data set. Instead of comparing neural network models such as CNN which is known to work badly on time series data, it would be much better if the author compared to well-known econometric time series models such as GARCH and VAR.<br />
<br />
=References=<br />
[1] Hamilton, J. D. Time series analysis, volume 2. Princeton university press Princeton, 1994. <br />
<br />
[2] Fama, E. F. Efficient capital markets: A review of theory and empirical work. The journal of Finance, 25(2):383–417, 1970.<br />
<br />
[3] Petelin, D., Sˇindela ́ˇr, J., Pˇrikryl, J., and Kocijan, J. Financial modeling using gaussian process models. In Intelligent Data Acquisition and Advanced Computing Systems (IDAACS), 2011 IEEE 6th International Conference on, volume 2, pp. 672–677. IEEE, 2011.<br />
<br />
[4] Tobar, F., Bui, T. D., and Turner, R. E. Learning stationary time series using gaussian processes with nonparametric kernels. In Advances in Neural Information Processing Systems, pp. 3501–3509, 2015.<br />
<br />
[5] Hwang, Y., Tong, A., and Choi, J. Automatic construction of nonparametric relational regression models for multiple time series. In Proceedings of the 33rd International Conference on Machine Learning, 2016.<br />
<br />
[6] Wilson, A. and Ghahramani, Z. Copula processes. In Advances in Neural Information Processing Systems, pp. 2460–2468, 2010.<br />
<br />
[7] Sirignano, J. Extended abstract: Neural networks for limit order books, February 2016.<br />
<br />
[8] Borovykh, A., Bohte, S., and Oosterlee, C. W. Condi- tional time series forecasting with convolutional neural networks, March 2017.<br />
<br />
[9] Heaton, J. B., Polson, N. G., and Witte, J. H. Deep learn- ing in finance, February 2016.<br />
<br />
[10] Neil, D., Pfeiffer, M., and Liu, S.-C. Phased lstm: Acceler- ating recurrent network training for long or event-based sequences. In Advances In Neural Information Process- ing Systems, pp. 3882–3890, 2016.<br />
<br />
[11] Chung, J., Gulcehre, C., Cho, K., and Bengio, Y. Em- pirical evaluation of gated recurrent neural networks on sequence modeling, December 2014.<br />
<br />
[12] Weissenborn, D. and Rockta ̈schel, T. MuFuRU: The Multi-Function recurrent unit, June 2016.<br />
<br />
[13] Cho, K., Courville, A., and Bengio, Y. Describing multi- media content using attention-based Encoder–Decoder networks. IEEE Transactions on Multimedia, 17(11): 1875–1886, July 2015. ISSN 1520-9210.<br />
<br />
[14] Glorot, X. and Bengio, Y. Understanding the dif- ficulty of training deep feedforward neural net- works. In In Proceedings of the International Con- ference on Artificial Intelligence and Statistics (AIS- TATSaˆ10). Society for Artificial Intelligence and Statistics, 2010.</div>Ka2khanhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=policy_optimization_with_demonstrations&diff=41466policy optimization with demonstrations2018-11-26T22:15:51Z<p>Ka2khan: /* Optimization */</p>
<hr />
<div>= Introduction =<br />
<br />
The reinforcement learning (RL) method has made significant progress in a variety of applications, but the exploration problems regarding how to gain more experience from novel policies to improve long-term performance are still challenges, especially in environments where reward signals are sparse and rare. There are currently two ways to solve such exploration problems in RL: 1) Guide the agent to explore states that have never been seen. 2) Guide the agent to imitate a demonstration trajectory sampled from an expert policy to learn. When guiding the agent to imitate the expert behavior for learning, there are also two methods: putting the demonstration directly into the replay memory [1] [2] [3] or using the demonstration trajectory to pre-train the policy in a supervised manner [4]. However, neither of th