Ming Dong, Professor, Ontario Tech University

Yafeng Liu, Associate Professor

Downlink-Uplink Beamforming Design for Wireless Federated Learning

16:00-17:00 April 25 (Thursday), S625

Federated learning (FL) is an effective distributed machine learning technique that allows multiple devices to collaboratively learn a global model using their local datasets. In a wireless environment, frequent exchange of the model information can stress the limited communication resources, and the fluctuation of the wireless link and noisy reception leads to degraded FL performance. This necessitates efficient communication design to effectively support FL.

In this talk, I will present our recent work on the beamforming design to improve wireless FL training performance. We first consider joint downlink-uplink beamforming design for wireless FL with a multi-antenna base station to train a global model. I will describe how we capture the downlink and uplink beamforming and receiver noise effect in the expected global loss function by deriving the global model update expression over communication rounds and using the bounding technique. We then propose a low-complexity joint beamforming algorithm to minimize the upper bound on the expected global loss, which only requires closed-form gradient updates. We show that the proposed joint beamforming design solution can substantially outperform the conventional separate-link design approach and nearly attain the performance of ideal FL with error-free communication links. We then consider simultaneously training multiple machine learning models using wireless FL, where we consider round-robin device-model assignment and downlink beamforming for concurrent multiple model updates. To maximize the multi-model training convergence rate, we bound the gap to optimality of the global model update, which captures the noisy transmission and inter-model interference. We show that minimizing this upper bound leads to a multi-group multicast beamforming design to minimize the sum of inverse received signal-to-interference-plus-noise ratios. Simulation shows that our proposed multi-model FL solution outperforms conventional single-model sequential training and multi-model zero-forcing beamforming.