Hyper-Sphere Quantization: Communication-Efficient SGD for Federated Learning

TL;DR

This paper introduces Hyper-Sphere Quantization (HSQ), a flexible gradient compression method for federated learning that reduces communication costs from $O(\sqrt{d} \log d)$ to $O(\log d)$ while maintaining convergence.

Contribution

HSQ provides a novel framework for gradient compression in federated learning, balancing communication efficiency and accuracy with theoretical convergence guarantees.

Findings

HSQ achieves up to $O(\log d)$ communication cost.

HSQ maintains convergence accuracy comparable to uncompressed training.

Experimental results confirm significant communication savings.

Abstract

The high cost of communicating gradients is a major bottleneck for federated learning, as the bandwidth of the participating user devices is limited. Existing gradient compression algorithms are mainly designed for data centers with high-speed network and achieve $O(\sqrt{d} \log d)$ per-iteration communication cost at best, where $d$ is the size of the model. We propose hyper-sphere quantization (HSQ), a general framework that can be configured to achieve a continuum of trade-offs between communication efficiency and gradient accuracy. In particular, at the high compression ratio end, HSQ provides a low per-iteration communication cost of $O(\log d)$, which is favorable for federated learning. We prove the convergence of HSQ theoretically and show by experiments that HSQ significantly reduces the communication cost of model training without hurting convergence accuracy.