NeurIPS 2020

Hierarchical Gaussian Process Priors for Bayesian Neural Network Weights

Meta Review

This work presents new and interesting ways to define a prior for Bayesian NNs by injecting a priori correlation structure in the weights. While there were reviewer questions about the motivation of such a technique (as opposed to using GPs in lieu of a neural network) the author response has convincingly addressed these concerns. There were also some concerns about interpretation of these functional priors, and how to control or tailor the inductive bias to a particular problem setting.