Recent advancements in mobile networks, the proliferation of powerful edge devices, AI breakthroughs, and heightened data privacy concerns have spurred the adoption of distributed machine learning approaches like Federated Learning (FL) and Split Learning (SL), each with pros and cons. This paper introduces a novel training framework designed to match the accuracy of FL while minimizing edge device workload, edge server data traffic, and model usage latency to enhance user experience. The proposed architecture features a dual-layer setup, employs a heuristic clustering algorithm, and enables grouped edge devices to train segments of the model. This approach leverages device-to-device (D2D) communication and the Associated Learning (AL) model to address model partitioning. Furthermore, it streamlines communication by having only the primary device in each group liaise with the edge server, thereby alleviating server traffic. Through PyTorch and ns3 simulations, this study demonstrates its capability to improve accuracy, reduce latency, and enhance user experience, effectively lightening the load on edge devices and servers in specific scenarios.
AL-powered EdgeFL: Achieving Low Latency and High Accuracy in Federated Learning
2024-06-24
915489 byte
Conference paper
Electronic Resource
English
Latency Minimization for Split Federated Learning
IEEE | 2023
|High Accuracy Automated Drilling Processes for Achieving Laminar Flow
British Library Conference Proceedings | 2016
|High Accuracy Automated Drilling Processes for Achieving Laminar Flow
SAE Technical Papers | 2016
|