Tram-FL: Reducing Communication and Computation Costs through Sequential Model Circulation in Decentralized Federated Learning
Tram-FL proposes a new decentralized federated learning method called Tram-FL to reduce computational and communication costs. Instead of each client keeping its own model, Tram-FL trains one single global model by circulating it among nodes sequentially. It optimizes the training schedule and uses quantized momentum to achieve high accuracy even with non-IID data, significantly cutting communication traffic.