Transceiver Design in Over-the-Air Federated Learning over MIMO Channels

In this paper, we investigate communication-efficient wireless federated learning (FL) in MIMO environments, where multiple edge devices with heterogeneous datasets undergoes local iterations during every round and subsequently transmit the gradients simultaneously to the server for aggregation. We derive an upper bound on the time-averaged gradient norm to characterize the convergence of over-the-air computation FL. Guided by the convergence analysis, we minimize the performance upper bound via an optimization problem solved with an alternating algorithm. Simulation results demonstrate that the proposed algorithm achieves similar performance compared to baseline method.