DART-FL Optimizes Edge Federated Learning for Dynamic Inference

Edge intelligence systems face the dual challenge of performing both model training and online inference on resource-constrained devices, especially when inference demand fluctuates significantly across tasks. This paper introduces DART-FL (Demand-driven Adaptive Resource-aware Task-level Federated Learning), a framework designed to address these coupled challenges. DART-FL dynamically manages the split of computational resources between inference and training, ensuring service-level objectives (SLOs) for inference are met. Crucially, DART-FL also adapts task-level training emphasis. Using a queue-aware scheduler, it allocates remaining training capacity to tasks experiencing higher inference demand, effectively shifting learning progress towards their burst periods. This means frequently requested tasks improve faster, enhancing model accuracy precisely when it's most needed, without compromising long-term multitask performance. Evaluated with real-world workloads, DART-FL demonstrates its ability to adapt resource allocation and learning emphasis, making federated learning more efficient and responsive at the edge.