Topology Control for Wireless Decentralized Federated Learning via Network Coding

Federated learning is emerging as a new paradigm for joint training of machine learning models across multiple distributed devices. In contrast to many existing works that require a central server to facilitate the exchange of local parameters with a star topology, this work considers fully decentralized operations over a wireless ring network. By increasing the radio coverage of each device, the convergence time, as indicated by the second largest singular value of the weighted adjacency matrix, can be shortened, but the mutual interference will be increased causing larger communication delay. The tradeoff between learning and communication delays is characterized by mathematical analysis of the singular-value gap and by numerical experiments on a linear regression problem. By carefully designing the consensus coefficients of the learning algorithm, a network coding scheme is crafted to improve the entire tradeoff curve without consuming more radio resources. It points to a new direction of using network coding to speed up wireless decentralized learning.