Full Speed Ahead: A Novel Selective Approach to Speed up Feature Extraction in Decentralized Federated Learning

The data generated by edge devices can be trained locally without the need to train in the cloud. To further utilize the resources of edge devices, we focus on how to accelerate edge training. However, there is one major challenge on the way towards efficient edge model training: complex models with high performance run slower than simple models, which means that the edge nodes need to communicate more times to ensure the reliability of the model performance. To address the challenge, we propose the granularity-based weight selection algorithm under multi-action control. In a nutshell, we continue to divide the models generated based on deep learning. We are then inspired by reinforcement learning to incorporate actions and rewards into the choice of aggregation methods. Extensive simulation results show that Our proposed method improves training efficiency by up to 43.75% to the benchmarks.