NeurIPS 2020

Liberty or Depth: Deep Bayesian Neural Nets Do Not Need Complex Weight Posterior Approximations

Meta Review

The authors analyze the effects of MFVI on weight posteriors. In particular, the authors analyze whether mean field posteriors can represent richer posteriors in function space through adding depth than by utilizing correlated variational families. Strengths: -understanding the induced posterior correlations in function space is timely and interesting -most of the analysis is clear and easy to follow Weaknesses: - Theorems should be checked for rigorousness and be renamed as "propositions"