WISE: Wireless Analog Computing at Radio Frequency for Disaggregated Deep Learning Inference

Zhihui Gao, Sri Krishna Vadlaman, Kfir Sulimany, Dirk Englund, Tingjun Chen

Keywords: Machine learning acceleration, in-physics computing, wireless network

Abstract:

The emerging deep learning enables various applications on today's edge devices, such as drones, smart wearables, and autonomous vehicles, while their energy- and memory-constrained nature demands efficient computing architectures to support real-time inference.
Hereby, we propose WISE, an analog computing architecture at radio frequency that performs deep learning inference between the remote model weights on the central radio, and inputs on the edge. Specifically, WISE is featured by two facts: (i) over-the-air model broadcasting enabling simultaneous inference across multiple edge devices, and (ii) analog computation of flexible and massive computing scales driven by a single frequency mixer. Extensive experiments on the software-defined testbed demonstrate that the deep learning based on WISE achieves 97.1\% classification accuracy on the MNIST dataset with an energy consumption of 31.01 fJ/MAC.

Submission Number: 24