Integrated Sensing-Communication-Computation Based Online Federated Learning with Limited Cache

This paper investigates a cache-assisted scheme for training an over-the-air federated edge learning (Air-FEEL) system with integrated sensing, communication, and computing (ISCC) utilizing both cached and real-time sensory data. In each training round, a server coordinates multiple devices to perform wireless sensing and utilize the real-time and cached old sensory data samples for updating the AI model based on the gradient descent method. Subsequently, the technique of over-the-air computation (AirComp) is utilized to aggregate local gradient vectors from all devices at the server for updating the global model. Particularly, this work makes the first attempt to exploit the spare on-device cache to reuse the old sensory data obtained in the previous round for enhancing learning performance. The theoretical analysis of this Air-FEEL framework is analyzed, where the mathematical relation between the convergence rate and the number of cached and real-time sensory data samples, sensing signal-to-noise ratio (SNR), and AirComp SNR is unveiled. Based on this theoretical finding and targeting enhancing the learning performance, a cache-assisted ISCC scheme is proposed to tackle a non-convex and complicated convergence accelerating problem via joint sensing and communication power control, cache and sensory data size allocation, and computation frequency management. Experimental results based on human motion recognition tasks verify the theoretical convergence analysis and show that the proposed cache-assisted ISCC scheme outperforms existing ISCC-based FL schemes without utilizing the spare on-device cache.