A Momentum-Based Wireless Federated Learning Acceleration With Distributed Principle Decomposition

Abstract

In the uplink period of wireless federated learning (WFL), multiple workers frequently upload uncoded training information to a server via orthogonal wireless channels. Due to the scarcity of wireless spectrum, the communication bottleneck appears during the uplink transmission. A one-shot distributed principle component analysis (PCA) method is leveraged to relieve the communication bottleneck by reducing the dimension of uploaded training information. Based on the low-dimensional training information, a Nesterov’s momentum accelerated WFL algorithm (i.e., PCA-AWFL) is proposed to reduce the communication rounds for the training of the federated learning system. For the non-convex loss functions, the finite-time convergence rate quantifies the impacts of system hyperparameters on the PCA-AWFL algorithm. Numerical results are used to demonstrate the performance improvement of the proposed PCA-AWFL algorithm over the benchmarks.