Helios: Heterogeneity-Aware Federated Learning with Dynamically Balanced Collaboration

TL;DR

Helios is a federated learning framework that dynamically adapts to device heterogeneity by compressing models for stragglers, improving training efficiency without sacrificing convergence.

Contribution

It introduces a heterogeneity-aware approach with a soft-training method to address stragglers in federated learning, enhancing efficiency and convergence.

Findings

Effective acceleration of stragglers in federated learning.

Maintains convergence despite model compression.

Improves overall training efficiency.

Abstract

In this paper, we propose Helios, a heterogeneity-aware FL framework to tackle the straggler issue. Helios identifies individual devices' heterogeneous training capability, and therefore the expected neural network model training volumes regarding the collaborative training pace. For straggling devices, a "soft-training" method is proposed to dynamically compress the original identical training model into the expected volume through a rotating neuron training approach. With extensive algorithm analysis and optimization schemes, the stragglers can be accelerated while retaining the convergence for local training as well as federated collaboration.