Analog Gradient Aggregation for Federated Learning Over Wireless Networks: Customized Design and Convergence Analysis

摘要: This article investigates the analog gradient aggregation (AGA) solution to overcome communication bottleneck for wireless federated learning applications by exploiting idea of over-the-air transmission. Despite various advantages, this special transmission also brings new challenges both transceiver design and algorithm due nonstationary local gradients time-varying channels in different rounds. To address these issues, we propose a novel AGA solution. In particular, parameters are optimized with consideration nonstationarity based on simple feedback variable. Moreover, rate is proposed stochastic descent algorithm, which adaptive quality estimation. Theoretical analyses provided convergence Finally, effectiveness confirmed two separate experiments linear regression shallow neural network. The simulation results verify that outperforms state-of-the-art baseline schemes much faster speed.