New perspective from Blackwell's "comparisons of experiments" on generative adversarial networks

We bring the tools from Blackwell's seminal result on comparing two stochastic experiments, to shine new lights on the modern  applications of great interest: generative adversarial networks (GAN). Binary hypothesis testing is at the center of GANs, and we propose new data processing inequalities that allows us to discover new algorithms to combat mode collapse, provide sharper analyses, and provide simpler proofs. This leads to a new architecture 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 analyze the new architecture and show that it encourages generators with no mode collapse. Experimental results show that the proposed architecture can learn to generate samples with diversity that is orders of magnitude better than competing approaches, while being simpler. For this talk, I will assume no prior background on GANs.