CIM Seminar

Abstract: A surprising characteristic of many overparametrized models in machine learning is that their performance improves by increasing their complexity, a phenomenon known as double descent. We review the recent progress made in describing this phenomenon from the statistical viewpoint. In particular, we present a novel analysis of feature selection in linear models by the l1-norm regularization. This framework is known by the LASSO and Basis pursuit optimization problems. We provide novel precise expressions for their performance in the asymptotically large setups with random features. We show that a double descent is achieved in the presence of a combination of strong and week features and discuss different possible scenarios.

The seminar will be held via Zoom. A link to the Zoom meeting will be posted at the morning of the event on the CIM mail list and when registering to the event.

Registration to the event.