Reviews: GibbsNet: Iterative Adversarial Inference for Deep Graphical Models

This paper presents GibbsNet, a deep generative model formulated as transition operators. The transition operators are learned in an adversarial way, similar to that of the adversarially learned inference (ALI). However instead of using a fixed prior p(z), GibbsNet does not require the specification of a particular prior, but rather learn a prior implicitly. Training is done by unrolling the sampling process multiple times and doing adversarial learning to match the sampling distribution to the one clamped from data and doing posterior only once. When unrolling for only one step GibbsNet becomes equivalent to ALI.