Topology-Optimal Multiple Gossip Steps for Decentralized Federated Learning via Gossip Tensor

This article proposes a new method called T-MGS (tensor-based multiple-gossip-steps) to improve decentralized federated learning (DFL, a system where many computers learn from data spread across multiple locations without sending all data to one place). The method reduces how many times computers need to communicate with each other by using gossip tensors (a mathematical structure for organizing multi-dimensional data) to control information flow between sites more efficiently, while maintaining the same model accuracy.