Adaptive Gradient Quantization for Data-Parallel SGD

TL;DR

This paper introduces adaptive gradient quantization schemes, ALQ and AMQ, that dynamically adjust to changing gradient statistics during training, improving accuracy and robustness in low-cost communication scenarios.

Contribution

The paper proposes novel adaptive quantization methods that update compression schemes during training based on gradient statistics, unlike prior fixed schemes.

Findings

Improved validation accuracy by nearly 2% on CIFAR-10 and 1% on ImageNet.

Enhanced robustness to hyperparameter choices.

Effective in low-cost communication settings.

Abstract

Many communication-efficient variants of SGD use gradient quantization schemes. These schemes are often heuristic and fixed over the course of training. We empirically observe that the statistics of gradients of deep models change during the training. Motivated by this observation, we introduce two adaptive quantization schemes, ALQ and AMQ. In both schemes, processors update their compression schemes in parallel by efficiently computing sufficient statistics of a parametric distribution. We improve the validation accuracy by almost 2% on CIFAR-10 and 1% on ImageNet in challenging low-cost communication setups. Our adaptive methods are also significantly more robust to the choice of hyperparameters.