Efficient Decentralized Deep Learning by Dynamic Model Averaging

TL;DR

This paper introduces a decentralized deep learning protocol that significantly reduces communication costs while maintaining high predictive accuracy, adaptable to concept drifts and suitable for mobile and autonomous applications.

Contribution

It presents a dynamic model averaging protocol that reduces communication by an order of magnitude without sacrificing performance, with theoretical bounds and empirical validation.

Findings

Reduces communication by an order of magnitude compared to existing methods.

Maintains predictive performance and loss bounds similar to periodic averaging schemes.

Validates effectiveness through extensive empirical evaluation.

Abstract

We propose an efficient protocol for decentralized training of deep neural networks from distributed data sources. The proposed protocol allows to handle different phases of model training equally well and to quickly adapt to concept drifts. This leads to a reduction of communication by an order of magnitude compared to periodically communicating state-of-the-art approaches. Moreover, we derive a communication bound that scales well with the hardness of the serialized learning problem. The reduction in communication comes at almost no cost, as the predictive performance remains virtually unchanged. Indeed, the proposed protocol retains loss bounds of periodically averaging schemes. An extensive empirical evaluation validates major improvement of the trade-off between model performance and communication which could be beneficial for numerous decentralized learning applications, such as autonomous driving, or voice recognition and image classification on mobile phones.