Marc Teboulle, Tel Aviv University

Non-Euclidean Proximal Based Methods
Date
Apr 8, 2022, 11:00 am12:00 pm
Event Description

We consider nonsmooth composite minimization problems, where the differentiable part of the objective fails to satisfy the standard global Lipschitz gradient property, a common and restrictive assumption used in almost all first order methods (FOM). To better capture the geometry of the given problem, we introduce the class of smooth adaptable functions which naturally translates into a Non-Euclidean sufficient descent property, and leads to non-Euclidean proximal schemes based on Bregman distances. This general framework lays the ground to significantly extend the scope of FOM, both in theory and applications, and this talk will present some of these recent developments.