Model-heterogeneous federated learning in wireless networks

Abstract

The collaborative machine learning technique known as Federated Learning (FL) has been the subject of active research, allowing multiple devices to learn together. Recent research has focused on model-heterogeneous FL, which aims to address the computational differences between participating devices. However, these studies have assumed an ideal communication channel without noise, which is not reflective of real-world wireless communication environments. This paper presents an evaluation of the performance of benchmark algorithms FjORD and HeteroFL in a wireless communication environment. Additionally, we propose a more communication-efficient algorithm, W-FjORD, for model-heterogeneous FL. Numerical results indicate that the proposed scheme significantly improves performance compared to benchmark schemes.