Encouraging GAN diversity via evolutionary computing

Abstract

Generative Adversarial Networks(GANs) have become very popular for their use in generating high quality images. Unfortunately, GANs also suffer from training instability, making them hard to use in practice[5]. In this thesis, we investigate a specific form of instability called mode collapse, where the model only learns a portion of the distribution. We augment standard GANs with approaches from evolutionary computing and find the augmentation does improve diversity substantially. Additionally, we develop new evolutionary models that further encourage diversity, along with an accompanying modular framework.