Rate-Convergence Tradeoff of Federated Learning over Wireless Channel

TL;DR

This paper analyzes how wireless channel conditions, coding rate, and packet errors affect federated learning convergence, proposing schemes to mitigate erasures and demonstrating the importance of memory in maintaining performance.

Contribution

It introduces two schemes for federated learning over wireless channels with packet erasures, analyzing the impact of coding rate and memory on convergence.

Findings

Memory significantly improves FL performance under packet erasures.

Coding rate influences convergence speed and accuracy.

Proposed schemes effectively mitigate packet erasure effects.

Abstract

In this paper, we consider a federated learning problem over wireless channel that takes into account the coding rate and packet transmission errors. Communication channels are modelled as packet erasure channels (PEC), where the erasure probability is determined by the block length, code rate, and signal-to-noise ratio (SNR). To lessen the effect of packet erasure on the FL performance, we propose two schemes in which the central node (CN) reuses either the past local updates or the previous global parameters in case of packet erasure. We investigate the impact of coding rate on the convergence of federated learning (FL) for both short packet and long packet communications considering erroneous transmissions. Our simulation results shows that even one unit of memory has considerable impact on the performance of FL in erroneous communication.