Séance Séminaire

Séminaire de Probabilités et Statistique

Uncertainty quantification for penalized M-estimators in high-dimensions (Webinaire)

The talk will focus on convex penalized M-estimators in high-dimensions in linear models when the ratio of the ambient dimension and the sample size are of the same order. A typical example covered by the theory is the combination of the Huber loss to fit the data and the Elastic-Net penalty to promote sparsity. I will present new means of uncertainty quantification for such penalized M-estimators for anisotropic Gaussian design and possibly heavy-tailed noise with two applications: (1) asymptotic normality and confidence intervals for single components of the true regression vector, and (2) estimation of the out-of-sample error of the penalized M-estimator. I will present applications to parameter tuning, where the practitioner can choose to minimize the length of the resulting confidence interval or the out-of-sample error.

Webinaire zoom uniquement (depuis le New Jersey) : https://umontpellier-fr.zoom.us/j/94087408185

Attention: à 14h