Wireless Federated Learning With Asynchronous and Quantized Updates

Federated learning (FL) is a framework of large-scale distributed learning with user privacy protection through local training and global aggregation. However, FL may suffer from asynchronous updates with quantized gradients in practice, which has not been well addressed in existing works. In this letter, we consider improving the convergence rate under the asynchronous FL system with performance guarantee and quantized. Specifically, our goal is to find the tradeoff between the quantization error (QE) and the staleness by taking the constraints on the quantization level and staleness into account. In order to make the optimization problem more efficient to solve, we derive an upper bound of QE and reformulate the optimization problem to obtain the optimal solutions. Simulation results quantify the impact of different quantization levels on the convergence rate, and demonstrate the performance improvement by achieving a tradeoff between the convergence rate and the QE.