Towards a Resource-Efficient Semi-Asynchronous Federated Learning for Heterogeneous Devices

Abstract

Our proposed resource-efficient semi-asynchronous federated learning (RE-SAFL) approach presents a comprehensive and effective solution for training large models such as Automatic Speech Recognition (ASR) models in a distributed and semi-asynchronous manner. In our research, we highlight the importance of employing a resource-efficient work allocation approach when deploying complex tasks such as ASR in real-time on edge devices such as mobile phones. To validate our approach, we conducted experiments on a real FL test-bed using Android-based mobile devices. By addressing the resource constraints of client devices and optimizing work allocation, our RE-SAFL framework opens up new possibilities for training large models in semi-asynchronous federated environments.