Abstract

Modern edge devices, such as cameras, drones, and Internet-of-Things nodes,rely on deep learning to enable a wide range of intelligent applications,including object recognition, environment perception, and autonomousnavigation. However, deploying deep learning models directly on the oftenresource-constrained edge devices demands significant memory footprints andcomputational power for real-time inference using traditional digital computingarchitectures. In this paper, we present WISE, a novel computing architecturefor wireless edge networks designed to overcome energy constraints in deeplearning inference. WISE achieves this goal through two key innovations:disaggregated model access via wireless broadcasting and in-physics computationof general complex-valued matrix-vector multiplications directly at radiofrequency. Using a software-defined radio platform with wirelessly broadcastmodel weights over the air, we demonstrate that WISE achieves 95.7% imageclassification accuracy with ultra-low operation power of 6.0 fJ/MAC perclient, corresponding to a computation efficiency of 165.8 TOPS/W. Thisapproach enables energy-efficient deep learning inference on wirelesslyconnected edge devices, achieving more than two orders of magnitude improvementin efficiency compared to traditional digital computing.