Accelerating Federated Learning with Adaptive Test-Driven Quantization at Edge Network
Yutao Liu, Chenyu Fan, Xiaoning Zhang · 2024
With the rapid development of 6G wireless communication and edge computing in recent years, mobile devices and terminals have collected vast amounts of data to train Machine Learning (ML) models. Since communication and computing resource constraints and data privacy exist for ML training at edge networks, the concept of Federated Learning (FL) is proposed to solve the above problems as a promising distributed ML paradigm. In FL, each participating client trains the local model and sends the local model upgrade to the central parameter server until the training process is completed. However, FL encounters a low bandwidth problem for parameter aggregation at network edges. To this end, in this paper, we propose the Test-Driven Adaptive Quantization (TAQ) method for FL. In TAQ, each client dynamically adopts a quantization level according to the training accuracy on a public testing dataset for the purpose of balancing communication overhead and training accuracy. In addition, we conduct a theoretical analysis of the convergence of TAQ. Detailed experimental results indicate that TAQ outperforms the state-of-the-art techniques, achieving a significant improvement in convergence speed ranging from 21% to 53%.