Moniqua: Modulo Quantized Communication in Decentralized SGD

TL;DR

Moniqua enables decentralized SGD to use highly efficient 1-bit quantized communication without sacrificing convergence speed or accuracy, reducing communication costs significantly.

Contribution

Moniqua introduces a provably efficient quantization method for decentralized SGD that requires no extra memory and works with 1-bit communication.

Findings

Moniqua converges faster in wall clock time than other quantized methods.

It maintains accuracy with 1-bit communication on ResNet models.

Theoretical bounds on communication bits per iteration are established.

Abstract

Running Stochastic Gradient Descent (SGD) in a decentralized fashion has shown promising results. In this paper we propose Moniqua, a technique that allows decentralized SGD to use quantized communication. We prove in theory that Moniqua communicates a provably bounded number of bits per iteration, while converging at the same asymptotic rate as the original algorithm does with full-precision communication. Moniqua improves upon prior works in that it (1) requires zero additional memory, (2) works with 1-bit quantization, and (3) is applicable to a variety of decentralized algorithms. We demonstrate empirically that Moniqua converges faster with respect to wall clock time than other quantized decentralized algorithms. We also show that Moniqua is robust to very low bit-budgets, allowing 1-bit-per-parameter communication without compromising validation accuracy when training ResNet20 and ResNet110 on CIFAR10.