The power of two samples in Generative Adversarial Networks (GAN).

We bring the tools from Blackwell's seminal result in 1958 on comparing two stochastic experiments, to shine new lights on a modern  applications of great interest: generative adversarial networks (GAN). Binary hypothesis testing is at the center of this application, and we propose new data processing inequalities that allows us to discover new algorithms, provide sharper analyses, and provide simpler proofs. This leads to a new framework to handle one of the major challenges in GAN known as ``mode collapse''; the lack of diversity in the samples generated by the learned generators.

The hypothesis testing view of GAN allows us to make a fundamental connection between our proposed idea of "packing" and mode collapse, suggesting that packing is the right framework to deal with mode collapse, when training GANs. For this talk, I will assume no prior background on GAN.