Fast-Convergent Federated Learning

Federated learning has emerged recently as a promising solution for distributing machine tasks through modern networks of mobile devices. Recent studies have obtained lower bounds on the expected decrease in model loss that is achieved each round federated learning. However, convergence generally requires large number communication rounds, which induces delay training and costly terms network resources. In this paper, we propose fast-convergent algorithm, called $\mathsf {FOLB}$ , performs intelligent sampling devices to optimize speed. We first theoretically characterize bound improvement can be if are selected according their local models will provide current global model. Then, show obtains uniform by weighting device updates gradient information. able handle both computation heterogeneity adapting aggregations estimates device’s capabilities contributing updates. evaluate comparison with existing algorithms experimentally its trained accuracy, speed, and/or stability across various datasets.