Aller au contenu. | Aller à la navigation

Outils personnels


UMR 5672

logo de l'ENS de Lyon
logo du CNRS
Vous êtes ici : Accueil / Séminaires / Machine Learning and Signal Processing / Camille Castera (University of Tübingen)

Camille Castera (University of Tübingen)

Near-optimal Closed-loop Method via Lyapunov Damping for Convex Optimization
Quand ? Le 21/11/2023,
de 13:00 à 14:00
Où ? M7 101
Participants Camille Castera
Ajouter un événement au calendrier vCal

Speaker: Camille Castera

Title: Near-optimal Closed-loop Method via Lyapunov Damping for Convex Optimization

Abstract: Nesterov's algorithm remains mysterious in many ways, such as its intriguing so-called damping coefficient $(k-1)/(k+2)$, crucial for its efficiency. This coefficient is said to be ""open-loop" as it depends on the iteration index $k$, making the initial iteration index an hyper-parameter that affects the performance of the method (unlike most other popular algorithms). To overcome this issue, we introduce a new continuous-time system with "closed-loop" damping. We do so by replacing the open-loop coefficient by the square root of a Lyapunov function, hence coupling the damping with the speed of convergence of the system. We show that our system is the first closed-loop strategy that achieves a convergence rate arbitrarily close to the optimal one for first-order methods on smooth convex functions. We then derive a practical first-order algorithm, via explicit discretization of the system and present numerical experiments supporting our theoretical findings. This is joint work with S. Maier and P. Ochs.

In Room M7 101