M22: Rate-Distortion Inspired Gradient Compression

Abstract

In federated learning (FL), the communication constraint between the remote users and the Parameter Server (PS) is a crucial bottleneck. This paper proposes M22, a rate-distortion inspired approach to model update compression for distributed training of deep neural networks (DNNs). In particular, (i) we propose a family of distortion measures referred to as "M-magnitude weighted L<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf>" norm, and (ii) we assume that gradient updates follow an i.i.d. distribution with two degrees of freedom – generalized normal and Weibull distributions. To measure the gradient compression performance under a communication constraint, we define the per-bit accuracy as the optimal improvement in accuracy that a bit of communication brings to the centralized model over the training period. Using this performance measure, we systematically benchmark the choice of gradient distributions and the distortion measure. We provide substantial insights on the role of these choices and argue that significant performance improvements can be attained using such a rate-distortion inspired compressor.