Device Selection and Resource Allocation for Layerwise Federated Learning in Wireless Networks

In this article, we study device selection and resource allocation (DSRA) for layerwise federated learning (FL) in wireless networks. For effective learning, DSRA should be carefully determined considering the characteristics of both layerwise FL and wireless networks. To address this, we propose a DSRA algorithm for layerwise FL, called LAFLAS, that maximizes the total average number of the shallow and entire parameter transmissions over time in FL while guaranteeing the ratio between their numbers. Through experiments, we show that the model trained by our LAFLAS outperforms those by state-of-the-art algorithms, which demonstrates the effectiveness of DSRA by LAFLAS.