SCAFFOLD: Stochastic Controlled Averaging for Federated Learning

arXiv1910.06378MaRDI QIDQ6327187

Ananda Theertha Suresh, Sashank J. Reddi, Mehryar Mohri, Sebastian U. Stich, Sai Praneeth Karimireddy, Satyen Kale

Publication date: 14 October 2019

Abstract: Federated Averaging (FedAvg) has emerged as the algorithm of choice for federated learning due to its simplicity and low communication cost. However, in spite of recent research efforts, its performance is not fully understood. We obtain tight convergence rates for FedAvg and prove that it suffers from `client-drift' when the data is heterogeneous (non-iid), resulting in unstable and slow convergence. As a solution, we propose a new algorithm (SCAFFOLD) which uses control variates (variance reduction) to correct for the `client-drift' in its local updates. We prove that SCAFFOLD requires significantly fewer communication rounds and is not affected by data heterogeneity or client sampling. Further, we show that (for quadratics) SCAFFOLD can take advantage of similarity in the client's data yielding even faster convergence. The latter is the first result to quantify the usefulness of local-steps in distributed optimization.

Has companion code repository: https://github.com/KarhouTam/FL-bench

Mathematics Subject Classification ID

Analysis of algorithms (68W40) Convex programming (90C25) Large-scale problems in mathematical programming (90C06) Distributed algorithms (68W15)

This page was built for publication: SCAFFOLD: Stochastic Controlled Averaging for Federated Learning