Disaggregated Deep Learning via In-Physics Computing at Radio Frequency
This addresses energy efficiency for edge devices like cameras and drones, offering a more than 100x improvement over traditional methods, though it is incremental as it builds on existing wireless and analog computing concepts.
The paper tackles the problem of energy constraints in deep learning inference on resource-constrained edge devices by introducing WISE, a computing architecture that uses wireless broadcasting and in-physics computation at radio frequency, achieving 95.7% image classification accuracy with 6.0 fJ/MAC per client and a 165.8 TOPS/W efficiency.
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 autonomous navigation. However, deploying deep learning models directly on the often resource-constrained edge devices demands significant memory footprints and computational power for real-time inference using traditional digital computing architectures. In this paper, we present WISE, a novel computing architecture for wireless edge networks designed to overcome energy constraints in deep learning inference. WISE achieves this goal through two key innovations: disaggregated model access via wireless broadcasting and in-physics computation of general complex-valued matrix-vector multiplications directly at radio frequency. Using a software-defined radio platform with wirelessly broadcast model weights over the air, we demonstrate that WISE achieves 95.7% image classification accuracy with ultra-low operation power of 6.0 fJ/MAC per client, corresponding to a computation efficiency of 165.8 TOPS/W. This approach enables energy-efficient deep learning inference on wirelessly connected edge devices, achieving more than two orders of magnitude improvement in efficiency compared to traditional digital computing.