Adaptive Gradient Quantization for Data-Parallel SGD
This work addresses communication efficiency for distributed deep learning training, offering an incremental improvement over fixed quantization methods.
The paper tackles the problem of communication bottlenecks in data-parallel SGD by introducing adaptive gradient quantization schemes (ALQ and AMQ) that adjust to changing gradient statistics during training, improving validation accuracy by almost 2% on CIFAR-10 and 1% on ImageNet in low-cost communication setups.
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.