Fiche de COSTANTINI Marina

Titre : Optimization methods over networks for popular content delivery and distributed machine learning
Date : 26 juin 2023
Heure: 14:00
Résumé : The great number of devices connected to the Internet requires designing algorithms that make wise use of the network resources. We present efficient algorithms for two timely applications: network edge caching and distributed model training. For network edge caching, we propose an algorithm that uses recommendation engines to promote contents that are cached. For distributed model training, we present algorithms that exploit inter-agent communication to train machine learning models that account for the data of all agents without requiring to share the data. First, we propose a family of algorithms with minimum synchronization that can reduce convergence time by letting the agents design their communication pattern online. Second, we analyze how inter-agent communication can improve the convergence of federated learning, a setting where in principle agents can only communicate with a central server. Our algorithms achieve superior performance entailing minimal or no structural changes
Lieu : EURECOM