The Architecture of Computing Power Network Towards Federated Learning: Paradigms and Perspectives

Abstract

Computing Power Network (CPN) is a new network paradigm for next generation communication systems. Meanwhile, Federated Learning (FL) has attracted more and more attention nowadays. However, there are few researches on the resource scheduling problem of federated learning in computing power network. There are a large number of heterogeneous computing resources available in the computing power network, so efficient utilization of resources in CPN for federated learning is very important. Therefore, our research focuses on the resource scheduling problem of federated learning in computing power networks to make up for the shortcomings of current related research. In this paper, we propose a framework and functional architecture combining CPN and federated learning for the purpose of resource optimization in federated learning. Besides, we show that task offloading using split learning can significantly improve the computational performance of federated learning, especially on local computing.