Why Cold Posteriors? On the Suboptimal Generalization of Optimal Bayes Estimates

Chen Zeno, Itay Golan, Ari Pakman, Daniel Soudry

Keywords: Cold posterior, Bayesian deep learning, Neural networks

TL;DR: We explore several possible reasons for the cold posterior effect.

Abstract: Recent works have shown that the predictive accuracy of Bayesian deep learning models exhibit substantial improvements when the posterior is raised to a 1/T power with T<1. In this work, we explore several possible reasons for this surprising behavior.