Generative adversarial networks (GANs) are a powerful but poorly understood tool for unsupervised learning. We seek to build a principled understanding of why they work, and why they can sometimes fail.

Generative adversarial networks (GANs) are a very powerful framework for learning distributions given the ability to sample from them. GANs consist of two “dueling” neural networks: a “generator” which aims to generate a candidate distribution that is a good guess of the true distribution, and a “discriminator” whose goal is to distinguish between the true and candidate distribution. To train these two networks, one greedily refines the generator and then the discriminator in alternating steps. However, despite extensive research, GAN training is notoriously difficult in practice. Mostly, due to the underlying dynamics having extremely chaotic nature and being very poorly understood from a theoretical point of view. The goal of our work is to provide a principled framework for understanding GAN dynamics. Our overarching approach here is to study variants of GANs that are relatively simple but already exhibit many of the pathologies that real-world GANs face. We expect carefully analysis of these simpler variants to provide us with insight into these pathologies and lead to development of methods for dealing with them. In fact, this approach already enabled us to unveil a certain kind of systematic failure of GAN training that was not identified before. Ultimately, we want this line of research to make real-world GAN training a well-understood and reliable process.