Saved in:
Bibliographic Details
Main Authors: Picard-Weibel, Antoine, Clerico, Eugenio, Moscoviz, Roman, Guedj, Benjamin
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2503.08231
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915191231348736
author Picard-Weibel, Antoine
Clerico, Eugenio
Moscoviz, Roman
Guedj, Benjamin
author_facet Picard-Weibel, Antoine
Clerico, Eugenio
Moscoviz, Roman
Guedj, Benjamin
contents We discuss necessary conditions for a PAC-Bayes bound to provide a meaningful generalisation guarantee. Our analysis reveals that the optimal generalisation guarantee depends solely on the distribution of the risk induced by the prior distribution. In particular, achieving a target generalisation level is only achievable if the prior places sufficient mass on high-performing predictors. We relate these requirements to the prevalent practice of using data-dependent priors in deep learning PAC-Bayes applications, and discuss the implications for the claim that PAC-Bayes ``explains'' generalisation.
format Preprint
id arxiv_https___arxiv_org_abs_2503_08231
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle How good is PAC-Bayes at explaining generalisation?
Picard-Weibel, Antoine
Clerico, Eugenio
Moscoviz, Roman
Guedj, Benjamin
Machine Learning
We discuss necessary conditions for a PAC-Bayes bound to provide a meaningful generalisation guarantee. Our analysis reveals that the optimal generalisation guarantee depends solely on the distribution of the risk induced by the prior distribution. In particular, achieving a target generalisation level is only achievable if the prior places sufficient mass on high-performing predictors. We relate these requirements to the prevalent practice of using data-dependent priors in deep learning PAC-Bayes applications, and discuss the implications for the claim that PAC-Bayes ``explains'' generalisation.
title How good is PAC-Bayes at explaining generalisation?
topic Machine Learning
url https://arxiv.org/abs/2503.08231