Decentralized Federated Learning on the Edge: From the Perspective of Quantization and Graphical Topology
Zhigang Yan, Dong Li · IEEE Internet of Things Journal · 2024
Decentralized Federated Learning (DFL), a Federated Edge Learning (FEEL) framework without a server, can avoid the huge communication overhead of the server and the single point of failure within FEEL. Since there is no server, the convergence of DFL depends not only on the communication conditions and number of edge nodes, but also on the graph topology of the network over which edge nodes communicate. Moreover, in practical scenarios, considering the limited communication resources, such as the latency and energy costs, the convergence of DFL also suffers from quantization errors. In this paper, we consider the decentralized gradient descent (DGD), which is widely applied in DFL, and examine the influence of graph topology and quantization on the convergence of DGD in different wireless networks with cost constraints. According to the derived convergence bound, the maximum quantization error acceptable for the DFL convergence is obtained. Furthermore, motivated by the limited iterations caused by the impact of constrained communication costs on each edge node, we compare the convergence bounds between higher and lower connectivity topologies, which face different energy consumption in one round of communication. Based on this comparison, the impact of graph topologies with energy constraints on the convergence can be observed. Numerical simulations confirm the validity of our analyses, supporting the correctness of our theoretical findings.