Optimal MIMO Combining for Blind Federated Edge Learning with Gradient Sparsification

TL;DR

This paper introduces an optimal MIMO receive combining method for federated learning that incorporates gradient sparsification, significantly enhancing performance in heterogeneous data scenarios.

Contribution

It presents a novel optimal combining strategy tailored for federated MIMO systems with gradient sparsification, addressing data heterogeneity.

Findings

Outperforms benchmark methods significantly

Effective in non-i.i.d. data scenarios

Improves federated learning efficiency

Abstract

We provide the optimal receive combining strategy for federated learning in multiple-input multiple-output (MIMO) systems. Our proposed algorithm allows the clients to perform individual gradient sparsification which greatly improves performance in scenarios with heterogeneous (non i.i.d.) training data. The proposed method beats the benchmark by a wide margin.