Reviews: PixelGAN Autoencoders

Update after rebuttal: I believe the paper can be accepted as a poster. I advise the authors to polish the writing to better highlight their contributions, motivation and design choices. This could make the work attractive and rememberable, not "yet another hybrid generative model". The authors provide a theoretical justification of the approach based on a decomposition of variational evidence lower bound (ELBO). The authors provide qualitative results with different priors on the hidden distribution, and quantitative results on semi-supervised learning on MNIST, SVHN and NORB.