cardinality Restricted Boltzmann Machines

Overview

Restricted Boltzmann Machine (RBM) is a probabilistic model which is usually represented by a graphical model. RBM can encode a probability distribution and therefore can be used for density estimation. It can also be viewed as a generative stochastic model of a neural network (with one hidden layer). Although interesting by itself, popularity of RBM started from the time that Hinton and others used it as the building blocks of Deep Belief Networks (DBNs). From the computational point of view, the good characteristic of RBN is that the posterior over hidden variables is factorizable.

At least two notions of sparsity can be investigated for graphical models. The first notion is the sparsity of the graph (i.e., something related to the number of edges in the graph). In this sense, a sparse model is one that has a lot of independencies in its structure. However, this is not the focus of this paper. In fact, the structure of RBM is fixed.

The other notion of sparsity is the usual definition in the other approaches. In the neural network literature, it means that we do not want the entire neurons to be activated in the same time. For our RBM, we want that the values of hidden variables be sparse for (almost) all inputs.

Motivations and Challenges

Sparsity is usually regarded as a good property for modeling complex phenomenons. From theoretical point of view, statisticians sometimes relate it to the idea of shrinkage. Also, experimental analysis shows that it is certainly useful. In particular, it has been suggested that sparse deep architectures can be very efficient. Finally, sparsity increases the interpretability of data.

The problem is for RBM is that if we want to add sparsity constraints, then the hidden variables would be coupled together. Therefore, the posterior would not be factorizable anymore, which makes sparse RBM intractable.