Expanding the Reach of Federated Learning by Reducing Client Resource Requirements

Communication on heterogeneous edge networks is a fundamental bottleneck in Learning (FL), restricting both model capacity and user. To address this issue, we introduce two novel strategies to communication costs: (1) the use of lossy compression on the global sent server-to-client; and (2) Federated Dropout, which allows users to train locally on smaller subsets of the global model and also a reduction in both client-to-server communication and local. We empirically show that these strategies, combined with existing approaches for client-to-server communication, collectively provide to a 14\ reduction in server-to-client communication, a 1. 7\ in local computation, and a 28\ reduction in upload, all without degrading the quality of the final model. We thus reduce FL's impact on client device resources, allowing higher models to be trained, and a more diverse set of users to be reached.