Séance Séminaire

Séminaire de Probabilités et Statistique

From multiple means estimation to personalized federated learning

I am currently working on personalized federated learning. The goal of these methods is to develop learning algorithms that train a model from data distributed across multiple agents, potentially exhibiting various forms of heterogeneity. Personalized approaches aim to leverage this multiplicity to improve the learning performance of a given agent’s model while accounting for its specificity, that is, avoiding bias induced by naive aggregation. We connect these questions to a high-dimensional multiple mean estimation problem, also known as multi-task averaging. These means are kernel mean embeddings, leading to the learning of a mixture over the different agents. I will present this connection, our proposed estimation methods, and the resulting federated learning algorithm and its statistical garantees. This talk will include different contributions, some in collaboration with Gilles Blanchard (Univ. Paris Saclay) and Hannah Marienwald (Technische Universität Berlin), and also ongoing works with Batiste Le Bars (Inria Lille) and Aurélien Bellet (Inria Montpellier).

Séminaire en salle 109, également retransmis sur zoom : https://umontpellier-fr.zoom.us/j/7156708132