Parallel Restarted SGD with Faster Convergence and Less Communication: Demystifying Why Model Averaging Works for Deep Learning

TLDR

Distributed training of deep neural networks commonly uses parallel minibatch SGD to accelerate learning, but its linear speed‑up is limited by increasing gradient communication costs, whereas model averaging reduces communication overhead yet its effectiveness remains theoretically unexplained. This paper aims to rigorously explain why model averaging can match the performance of parallel minibatch SGD while requiring significantly less communication. The authors conduct a theoretical analysis of model averaging, deriving convergence guarantees and communication complexity that show it achieves comparable speed‑up to parallel minibatch SGD. Their results confirm that, when the averaging interval is carefully controlled, model averaging attains a good training‑time speed‑up with reduced communication, matching the performance of parallel minibatch SGD.

Abstract

In distributed training of deep neural networks, parallel minibatch SGD is widely used to speed up the training process by using multiple workers. It uses multiple workers to sample local stochastic gradients in parallel, aggregates all gradients in a single server to obtain the average, and updates each worker’s local model using a SGD update with the averaged gradient. Ideally, parallel mini-batch SGD can achieve a linear speed-up of the training time (with respect to the number of workers) compared with SGD over a single worker. However, such linear scalability in practice is significantly limited by the growing demand for gradient communication as more workers are involved. Model averaging, which periodically averages individual models trained over parallel workers, is another common practice used for distributed training of deep neural networks since (Zinkevich et al. 2010) (McDonald, Hall, and Mann 2010). Compared with parallel mini-batch SGD, the communication overhead of model averaging is significantly reduced. Impressively, tremendous experimental works have verified that model averaging can still achieve a good speed-up of the training time as long as the averaging interval is carefully controlled. However, it remains a mystery in theory why such a simple heuristic works so well. This paper provides a thorough and rigorous theoretical study on why model averaging can work as well as parallel mini-batch SGD with significantly less communication overhead.