Adaptive Braking for Mitigating Gradient Delay

TL;DR

This paper introduces Adaptive Braking, a technique for momentum-based optimizers that dynamically scales gradients to mitigate delay effects in asynchronous neural network training, enabling stable and efficient training with large delays.

Contribution

The paper proposes Adaptive Braking, a novel method that improves asynchronous training stability and performance by dynamically adjusting gradients based on their alignment with velocity.

Findings

Enabling training ResNets on CIFAR-10 and ImageNet with delays ≥ 32 steps.

Minimal drop in final test accuracy with Adaptive Braking.

Improved stability and acceleration in asynchronous training.

Abstract

Neural network training is commonly accelerated by using multiple synchronized workers to compute gradient updates in parallel. Asynchronous methods remove synchronization overheads and improve hardware utilization at the cost of introducing gradient delay, which impedes optimization and can lead to lower final model performance. We introduce Adaptive Braking (AB), a modification for momentum-based optimizers that mitigates the effects of gradient delay. AB dynamically scales the gradient based on the alignment of the gradient and the velocity. This can dampen oscillations along high curvature directions of the loss surface, stabilizing and accelerating asynchronous training. We show that applying AB on top of SGD with momentum enables training ResNets on CIFAR-10 and ImageNet-1k with delays $D \geq$ 32 update steps with minimal drop in final test accuracy.