Université Paris 6
Pierre et Marie Curie
Université Paris 7
Denis Diderot

CNRS U.M.R. 7599
``Probabilités et Modèles Aléatoires''

Randomized estimators and empirical complexity for pattern recognition and least square regression

Auteur(s):

Code(s) de Classification MSC:

Résumé: We present an alternative to the penalized maximum likelihood approach to model selection. Instead of penalizing the likelihood, we consider its quantiles under some prior distribution on the parameter set, in order to derive non-asymptotic deviation bounds for some randomized estimators. This leads to a new measure of the complexity of models which, unlike Vapnik's entropy, is (in principle) computable from empirical observations.

Mots Clés: Model selection ; pattern recognition ; least square regression ; oracle inequalities ; deviation inequalities ; pseudo-Bayesian methods

Date: 2001-07-02

Prépublication numéro: PMA-677