Efficient federated learning via the alternating direction method of multipliers

Next Thursday April 25, 2024:

Organized by: FAU DCN-AvH, Chair for Dynamics, Control, Machine Learning and Numerics – Alexander von Humboldt Professorship at FAU, Friedrich-Alexander-Universität Erlangen-Nürnberg (Germany)
Title: Efficient federated learning via the alternating direction method of multipliers
Speaker: Ziqi Wang
Affiliation: PhD student at FAU DCN-AvH Chair for Dynamics, Control, Machine Learning and Numerics – Alexander von Humboldt Professorship.

Abstract. Federated Learning (FL) has emerged as a promising framework for learning from distributed data while maintaining privacy. The development of efficient FL algorithms encounters various challenges, including heterogeneous data and systems, limited communication capacities, and constrained local computational resources. In this talk, we introduce a novel approach leveraging the alternating direction method of multipliers (ADMM) to address these challenges. Our inexact and self-adaptive FedADMM algorithm mitigates the necessity for empirical hyperparameter settings. The convergence of the proposed algorithm using the inexactness criterion is analyzed, and numerical experiments demonstrate the enhanced performance of our algorithm.