NeurIPS 2020

Learning Latent Space Energy-Based Prior Model


Meta Review

This paper introduces a variational EM approach to latent variable models with an energy-based models (EBM) for the latent distribution. The work is a nice contribution in a line of work that revisits the use of EBMs. The reviewers praised the technical novelty and the technical justification of the training algorithm. This work is relevant to the NeurIPS and likely to be of particular interest this year. Some reviewers noted that the current manuscript does not sufficiently discuss related work and some comparisons are lacking, noting in particular the DVAE line of work. Most reviewers agreed that the DVAE and related work concerns are not severe enough to recommend rejection, so I am recommending acceptance. It is critical that the authors take into account the reviewer comments and make a particular effort to address the related work concerns in the revised paper.