Accelerating Federated Edge Learning via Topology Optimization

Federated edge learning (FEEL) is envisioned as a promising paradigm to achieve privacy-preserving distributed learning. However, it consumes excessive time due the existence of straggler devices. In this article, novel topology-optimized FEEL (TOFEL) scheme proposed tackle heterogeneity issue in federated and improve communication-and-computation efficiency. Specifically, problem jointly optimizing aggregation topology computing speed formulated minimize weighted summation energy consumption latency. To solve mixed-integer nonlinear problem, we propose solution method penalty-based successive convex approximation (SCA), which converges stationary point primal under mild conditions. facilitate real-time decision making, an imitation-learning-based developed, where deep neural networks (DNNs) are trained offline mimic method, imitation DNNs deployed at devices for online inference. Thereby, efficient approach seamlessly integrated into TOFEL framework. Simulation results demonstrate that accelerates process achieves higher Moreover, apply 3-D object detection with multivehicle cloud data sets CARLA simulator. The confirm superior performance over conventional designs same resource deadline constraints.