Séance Séminaire

Séminaire de Probabilités et Statistique

Bi-level optimization for hyperparameter optimization: application to structure discovery

Groupwise penalties are largely used in regularization-based optimization problems to enforce some group structure on the learned parameters. The most popular instance of such type of penalties is notably encountered in the Group Lasso problem where the $l_{1,2}$ norm imposes sparsity at the group level. However, most of existing methods require the group structure to be known a priori. In practice, this may be a too strong assumption, potentially hampering the prediction accuracy of the model. To circumvent this issue, we present a method to estimate the group structure by means of a continuous and differentiable bilevel optimization problem.

WEBINAIRE ouvert à toutes et tous : https://umontpellier-fr.zoom.us/j/85813807839