Cold Posteriors through PAC-Bayes. (arXiv:2206.11173v1 [cs.LG])

Web: http://arxiv.org/abs/2206.11173

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 the cold
posterior effect should take into account that approximate Bayesian inference
does not readily provide guarantees of performance on out-of-sample data.
Instead, out-of-sample error is better described through a generalization
bound. In this context, we explore the connections between the ELBO objective
from variational inference and the PAC-Bayes …

arxiv bayes lg