Parallel training of DNNs with Natural Gradient and Parameter Averaging

TL;DR

This paper presents a scalable neural network training framework that combines parameter averaging across multiple machines with an efficient Natural Gradient method, improving convergence and hardware utilization.

Contribution

It introduces a novel combination of parameter averaging and an approximate Natural Gradient method for effective multi-machine DNN training.

Findings

Parameter averaging enables multi-machine training with minimal network traffic.

The approximate Natural Gradient significantly improves convergence.

The combined method performs well in large-scale speech recognition tasks.

Abstract

We describe the neural-network training framework used in the Kaldi speech recognition toolkit, which is geared towards training DNNs with large amounts of training data using multiple GPU-equipped or multi-core machines. In order to be as hardware-agnostic as possible, we needed a way to use multiple machines without generating excessive network traffic. Our method is to average the neural network parameters periodically (typically every minute or two), and redistribute the averaged parameters to the machines for further training. Each machine sees different data. By itself, this method does not work very well. However, we have another method, an approximate and efficient implementation of Natural Gradient for Stochastic Gradient Descent (NG-SGD), which seems to allow our periodic-averaging method to work well, as well as substantially improving the convergence of SGD on a single machine.