In the context of assessing the generalization abilities of a randomized model or learning algorithm, PAC-Bayes and Differential Privacy (DP) theories are the state-of-the-art tools. For this reason, in this paper, we will develop tight DP-based generalization bounds, which improve over the current state-of-the-art ones both in terms of constants and rate of convergence. Moreover, we will also prove that some old and new randomized algorithm, show better generalization performances with respect to their non private counterpart, if the DP is exploited for assessing their generalization ability. Results on a series of algorithms and real world problems show the practical validity of the achieved theoretical results.

Randomized learning: Generalization performance of old and new theoretically grounded algorithms

Oneto, Luca;Cipollini, Francesca;Ridella, Sandro;Anguita, Davide
2018

Abstract

In the context of assessing the generalization abilities of a randomized model or learning algorithm, PAC-Bayes and Differential Privacy (DP) theories are the state-of-the-art tools. For this reason, in this paper, we will develop tight DP-based generalization bounds, which improve over the current state-of-the-art ones both in terms of constants and rate of convergence. Moreover, we will also prove that some old and new randomized algorithm, show better generalization performances with respect to their non private counterpart, if the DP is exploited for assessing their generalization ability. Results on a series of algorithms and real world problems show the practical validity of the achieved theoretical results.
File in questo prodotto:
File Dimensione Formato  
J029 - NEUCOM.pdf

accesso chiuso

Tipologia: Documento in versione editoriale
Dimensione 1.11 MB
Formato Adobe PDF
1.11 MB Adobe PDF   Visualizza/Apri   Richiedi una copia

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11567/914696
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 4
  • ???jsp.display-item.citation.isi??? 4
social impact