Learning Rate Optimization for Federated Learning Exploiting\n Over-the-air Computation

Abstract

Federated learning (FL) as a promising edge-learning framework can\neffectively address the latency and privacy issues by featuring distributed\nlearning at the devices and model aggregation in the central server. In order\nto enable efficient wireless data aggregation, over-the-air computation\n(AirComp) has recently been proposed and attracted immediate attention.\nHowever, fading of wireless channels can produce aggregate distortions in an\nAirComp-based FL scheme. To combat this effect, the concept of dynamic learning\nrate (DLR) is proposed in this work. We begin our discussion by considering\nmultiple-input-single-output (MISO) scenario, since the underlying optimization\nproblem is convex and has closed-form solution. We then extend our studies to\nmore general multiple-input-multiple-output (MIMO) case and an iterative method\nis derived. Extensive simulation results demonstrate the effectiveness of the\nproposed scheme in reducing the aggregate distortion and guaranteeing the\ntesting accuracy using the MNIST and CIFAR10 datasets. In addition, we present\nthe asymptotic analysis and give a near-optimal receive beamforming design\nsolution in closed form, which is verified by numerical simulations.\n