NIPS 2017
Mon Dec 4th through Sat the 9th, 2017 at Long Beach Convention Center
Paper ID: 334 f-GANs in an Information Geometric Nutshell

### Reviewer 1

This paper considers the f-GAN principle generalized to f-divergences. Generalizing the connection between the KL divergence and regular exponential families by using the chi-logarithm, the authors prove the variational information geometric f-GAN (vig-f-GAN) identity, which provides an interpretation of the objective function of the f-GAN. The authors also prove a sufficient condition on activation functions to obtain factored generative distributions. The main theoretical result, vig-f-GAN identity is interesting and its applicability is promising. However, the meanings of main theorems are hard to follow since they are written in quite general forms. Section 6: It is hard to understand how the observations in experiments are related to theoretical results. p.7, l.258, For eq.(7), is means when C …’’: This clause has a wrong structure.