AIDRC Seminar Series - Dr Maxime Vono

Title:

Bayesian Federated Learning

Abstract:

In the modern big data era, it has become commonplace to acquire and process a large amount of data at the edge nodes of a network. These nodes are typically devices (coined clients), such as IoT sensors or mobiles, coordinated by a central server. This rapid progress in data acquisition and storage has contributed to bringing out new paradigms regarding access to the data and their use in machine learning. Among these, we can cite partial device participation, limited upload bandwidth, data privacy issues, and heterogeneous data. Federated learning (FL) is a particular branch of machine learning which aims at addressing these new challenges.

In this talk, I will tackle the FL paradigm through the Bayesian lens in order to quantify epistemic uncertainty. I will present existing research routes to meet this goal before focusing on two competing but complementary simulation-based approaches based on Langevin dynamics. I will provide non-asymptotic convergence guarantees for these approaches and highlight the impact of communication overhead, data heterogeneity, and partial device participation on convergence. I will finally illustrate the benefits of these approaches on several FL benchmarks.