stat441w18/e-gan

Presented by

1. Yufeng Yue

2. Shanzi Wang

3. Yumeng Li

4. Yuyang Bao

5. Yun Shi

6. Wan Feng Cai

7. Ki Beom Lee

8. Qian Xiang

Introduction

In recent years, Generative Adversarial Networks (GAN) have emerged as a powerful method for learning generative models with real-world data. However, existing GANs (GAN and its variants) is experiencing training problems such as instability and mode collapse.

In this paper, authors proposed a novel GAN framework called Evolutionary Generative Adversarial Networks (E-GAN) for stable GAN training and improved generative performance. Compared with existing GANs, which employ a pre-defined adversarial objective function alternately training a generator and a discriminator, E-GAN uses different adversarial training objectives as mutation operations and evolves a population of generators to adapt to the environment (i.e., the discriminator). E-GAN also uses an evaluation mechanism to measure the quality and diversity of generated samples, such that only well-performing generator(s) are preserved and used for further training so that E-GAN would be able to overcome the limitations of an individual adversarial training objective and always preserves the best offspring, contributing to progress in and the success of GANs.

In a later section of the paper, experiment results on several datasets show that E-GAN achieves convincing generative performance and reduces the training problems inherent in existing GANs.

Method

Evaluation

First, we simply feed generator produced images into the discriminator D and observe the average value of the output, which we name the quality fitness score:

Then, we employ a similar principle to evaluate generator optimization stability and generative diversity as: Based on the aforementioned two fitness scores, the final evaluation (or fitness) function is:

where gamma >= 0. Overall, a relatively high fitness score F, leads to higher training efficiency and better generative performance.