Prochain Meetup

Robust Machine Learning: theory&algorithms+Repas, Guillaume Lecué chercheur CNRS
Pour cette séance, rendez-vous dès 12h00 pour un petit repas pour prendre des forces avant l'exposé de Guillaume Lecué, chercheur au CNRS et membre du CREST, laboratoire de statistique. Dans cet exposé (résumé ci-dessous), Guillaume nous introduira une nouvelle technique en Machine Learning pour construire des algorithmes robustes aux 'outliers', ou données aberrantes. Il abordera ensuite les nombreuses applications possibles dans un contexte de classification ou d'estimation de distribution de probabilités ! ### Title: Robust machine learning via median-of-means: theory and algorithms Abstract: The aim of the talk is to construct procedures and algorithms that are robust to outliers. The Median-of-means (MOM) principle is the key tool to our approach. We introduce minmax MOM estimators and show that they achieve sub-Gaussian deviation bounds both in small and high-dimensional least-squares regression. In particular, these estimators are efficient under moments assumptions on data that may have been corrupted by outliers. Besides these theoretical guarantees, the definition of minmax MOM estimators suggests simple and systematic modifications of standard algorithms used to approximate least-squares estimators and their regularized versions. As a proof of concept, we perform an extensive simulation study of these algorithms for robust versions of the lasso. We also show that if one combines MOM principle with convex and Lipschitz loss functions then one can obtain statistical bounds (such as estimation and sharp oracle inequalities) which holds with exponentially large probability without any stochastic assumption. We therefore derive robust versions of classical procedures such as logistic LASSO, quantile regression, Huber regression kernel methods and there associated robust algorithms.

Hélioparc

2 Avenue du Président Pierre Angot · Pau

Prochains Meetups

Meetups précédents (23)