Towards Serverless Semi-Decentralized Federated Learning with Heterogeneous Optimizers
Researchers have introduced a novel serverless semi-decentralized federated learning (SSD-FL) approach designed to operate without persistent server infrastructure. This method addresses the complexities of cluster formation in decentralized environments by employing a one-time device-to-device initialization phase for clustering. The SSD-FL framework segments training into intra-cluster and inter-cluster phases, ensuring global convergence through unique "effective loss functions" that integrate diverse local optimizers and network regularization. AI
IMPACT This research introduces a novel serverless approach to federated learning, potentially improving efficiency and scalability for distributed AI model training.