NQFL: Nonuniform Quantization for Communication Efficient Federated Learning
Guojun Chen, Kaixuan Xie, Yuheng Tu, Tiecheng Song, Yinfei Xu, Jing Hu, Lun Xin
Abstract
Federated learning (FL), as a potential machine learning framework for privacy preservation, has gained significant attention. However, the considerable communication overhead associated with FL remains a prominent challenge. To mitigate this issue, a nonuniform quantization scheme based on Lloyd-Max algorithm is introduced in this letter. By employing this approach, less communication resources are consumed to achieve the same performance. Through performance analysis and numerical simulations, we verify the convergence and effectiveness of the proposed algorithm. It demonstrates the potential of our approach in reducing communication overhead while maintaining reliable performance in FL systems.