Accumulated Gradient Normalization

TL;DR

This paper introduces a novel distributed optimizer called Accumulated Gradient Normalization that improves asynchronous data parallel optimization by normalizing gradients, reducing staleness effects, and enhancing convergence rates.

Contribution

It proposes a new optimizer that normalizes gradients to better align worker contributions and mitigate implicit momentum, improving asynchronous training stability and efficiency.

Findings

Achieves better convergence rates than EASGD and DynSGD.

Effectively mitigates parameter staleness in asynchronous optimization.

Empirically outperforms existing optimizers in distributed settings.

Abstract

This work addresses the instability in asynchronous data parallel optimization. It does so by introducing a novel distributed optimizer which is able to efficiently optimize a centralized model under communication constraints. The optimizer achieves this by pushing a normalized sequence of first-order gradients to a parameter server. This implies that the magnitude of a worker delta is smaller compared to an accumulated gradient, and provides a better direction towards a minimum compared to first-order gradients, which in turn also forces possible implicit momentum fluctuations to be more aligned since we make the assumption that all workers contribute towards a single minima. As a result, our approach mitigates the parameter staleness problem more effectively since staleness in asynchrony induces (implicit) momentum, and achieves a better convergence rate compared to other optimizers such as asynchronous EASGD and DynSGD, which we show empirically.