Performance-Oriented Design for Intelligent Reflecting Surface Assisted Federated Learning

To efficiently exploit the massive raw data that is pervading generated at mobile edge networks, federated learning (FL) has emerged as a promising distributed learning technique that was regarded as a substitute for centralized learning operations. By collaboratively training a shared learning model at edge devices, the raw data transmission and storage are bypassed via the local computed parameters/gradients exchange in FL. Hence, FL can overcome high communication latency and privacy issues. While the high dimensionality in iterative updates (millions of parameters/gradients may be included in the model training) still conflicts with the scarcity of communication resources. Over-the-air computation (AirComp) has come into the spotlight recently which profitably leverages the inherent superposition property of wireless channels to perform efficient model aggeration. However, the model aggregation accuracy is still severely damaged by the unfavorable wireless propagation channels. In this paper, we harness the intelligent reflecting surface (IRS) to program the wireless channel, thus acquiring a satisfying learning performance. Specifically, a performance-oriented design scheme that directly minimizes the optimality gap of the loss function is proposed to accelerate the convergence of AirComp based FL. Firstly, we analyze the convergence behavior of the FL procedure. Then, both offline and online design approaches are proposed based on the obtained optimality gap. We adopt the block coordinate descent (BCD) method to tackle the highly-intractable problem. Simulation results demonstrate that such a performance-oriented design strategy can achieve higher test accuracy than the conventional isolated mean square error (MSE) minimization approach in FL.