A Journey through a Federated Valley to Painless Federated Learning

Title of the Talk:A Journey through a Federated Valley to Painless Federated Learning
Speaker: Dr.Bapi Chatterjee
Host Faculty: Prof. M.V Pandu Ranga Rao
Date: January 07, 2025
Time: 03:00 pm
Venue: LHC-11

Abstract In this talk, we will walk through the developments in Federated Learning over the last few years. We will discuss the challenges of training a machine learning model in the federated setting and the algorithms to address them. The talk will discuss the motivations and highlights of the algorithms FedAvg, FedProx, Scaffold, FedAdam, FedProto, and others. This journey will underline the additional tunable hyperparameters on the server to determine the scaling factor in some methods. To address that, two new algorithms are proposed. In the first method, we establish that a descent-ensuring step-size regime at the clients ensures descent for the server objective; thereby, it enables linear convergence for strongly convex federated objectives even with partial client participation. Our second method introduces an approach for linearization of the federated objective to compute the scaling factor. We will see that the empirical results show that the proposed methods perform at par or better than the popular federated learning algorithms for both convex and non-convex problems.

Speaker Profile:

Bapi Chatterjee is an Assistant Professor at IIIT Delhi. His research focuses on scalable algorithms for Distributed and Federated Machine Learning, Concurrent Data Structures, and Learned Data Structures. He has authored or co-authored publications that appeared at venues such as the PODC, AAAI, DISC, IPDPS, ICDCS, Theoretical Computer Science (TCS) Journal, etc. Before joining IIIT Delhi, he worked as a Postdoc Fellow at the Institute of Science and Technology Austria. Before that, he worked as a researcher at IBM India Research Lab. He earned a Ph.D. in Computer Science and Engineering from Chalmers University of Technology, Gothenburg, Sweden.