How good is PAC-Bayes at explaining ...
Type de document :
Pré-publication ou Document de travail
URL permanente :
Titre :
How good is PAC-Bayes at explaining generalisation?
Auteur(s) :
Picard-Weibel, Antoine [Auteur]
MOdel for Data Analysis and Learning [MODAL]
Inria Lille - Nord Europe
Centre International de Recherche Sur l'Eau et l'Environnement [Suez] [CIRSEE]
Clerico, Eugenio [Auteur]
Universitat Pompeu Fabra [Barcelona] [UPF]
Moscoviz, Roman [Auteur]
Centre International de Recherche Sur l'Eau et l'Environnement [Suez] [CIRSEE]
Guedj, Benjamin [Auteur]
University College of London [London] [UCL]
Inria Lille - Nord Europe
MOdel for Data Analysis and Learning [MODAL]
Inria Lille - Nord Europe
Centre International de Recherche Sur l'Eau et l'Environnement [Suez] [CIRSEE]
Clerico, Eugenio [Auteur]
Universitat Pompeu Fabra [Barcelona] [UPF]
Moscoviz, Roman [Auteur]
Centre International de Recherche Sur l'Eau et l'Environnement [Suez] [CIRSEE]
Guedj, Benjamin [Auteur]

University College of London [London] [UCL]
Inria Lille - Nord Europe
Date de publication :
2025
Mot(s)-clé(s) en anglais :
Machine Learning (stat.ML)
Machine Learning (cs.LG)
FOS: Computer and information sciences
Machine Learning (cs.LG)
FOS: Computer and information sciences
Discipline(s) HAL :
Informatique [cs]/Apprentissage [cs.LG]
Mathématiques [math]/Statistiques [math.ST]
Mathématiques [math]/Statistiques [math.ST]
Résumé en anglais : [en]
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 ...
Lire la suite >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.Lire moins >
Lire la suite >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.Lire moins >
Langue :
Anglais
Collections :
Source :
Date de dépôt :
2025-03-19T07:28:47Z
Fichiers
- 2503.08231
- Accès libre
- Accéder au document