Machine Learning in Montpellier, Theory & Practice

In federated learning, multiple agents collaboratively train a machine learning model without exchanging local data. To achieve this, each agent locally updates a global model, and the updated models are periodically aggregated by a central server. In this talk, I will first focus on federated linear stochastic approximation (FedLSA), with an emphasis on agent heterogeneity. I will derive upper bounds on the sample and communication complexity of FedLSA and present a method to reduce communication costs using control variates. Special attention will be given to the « linear speed-up » phenomenon, demonstrating that the sample complexity scales inversely with the number of agents in both methods. Finally, I will discuss possible extensions of these results to non-linear federated learning and provide first-order expansions of the bias in federated averaging.

Map /Briefcase/RF-Inria_logo_signat.png" target="_blank" rel="noopener">Online