Cold Posteriors through PAC-Bayes

Konstantinos Pitas, Julyan Arbel

We investigate the cold posterior effect through the lens of PAC-Bayes generalization bounds. We argue that in the non-asymptotic setting, when the number of training samples is (r…