Optimal Importance Sampling for Federated Learning

TL;DR

This paper develops optimal importance sampling strategies for federated learning, demonstrating that non-uniform sampling enhances the efficiency and performance of the FedAvg algorithm through theoretical analysis and empirical validation.

Contribution

It introduces the first optimal importance sampling methods for both agent and data selection in federated learning, improving upon uniform sampling approaches.

Findings

Non-uniform sampling without replacement outperforms uniform sampling.

Theoretical analysis confirms improved convergence rates.

Experimental results validate the effectiveness of the proposed sampling strategies.

Abstract

Federated learning involves a mixture of centralized and decentralized processing tasks, where a server regularly selects a sample of the agents and these in turn sample their local data to compute stochastic gradients for their learning updates. This process runs continually. The sampling of both agents and data is generally uniform; however, in this work we consider non-uniform sampling. We derive optimal importance sampling strategies for both agent and data selection and show that non-uniform sampling without replacement improves the performance of the original FedAvg algorithm. We run experiments on a regression and classification problem to illustrate the theoretical results.