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 (relatively) small, discussions of