Title: Enabling Over-the-Air AI for Edge Computing via Metasurface-Driven Physical Neural Networks
Authors: Chao Feng, Xiaojiang Chen (Northwest University; Shaanxi Key Laboratory of Passive Internet of Things and Neural Computing); Shuo Liang (Northwest University; Shaanxi International Joint Research Centre for the Battery-Free Internet of Things); Chenghui Li (Northwest University; Shaanxi International Joint Research Centre for the Battery-Free Internet of Things); Gaogeng Zhao (Northwest University; Xi’an Advanced Battery-Free Sensing and Computing Technology International Science and Technology Cooperation Base); Beier Jing (Northwest University; Xi’an Key Laboratory of Advanced Computing and System Security); Yaxiong Xie (University at Buffalo SUNY)
Scribe: Ziyi Wang (XIamen University)
Introduction
Due to constrained computational capabilities and limited battery life, IoT devices often struggle with complex on-device intelligent tasks. While traditional “transmit-then-compute” edge AI architectures treat communication and computation as independent sequential processes, they inevitably face significant energy consumption and latency bottlenecks. Although existing Physical Neural Networks (PNN) leverage optical or radio frequency propagation for high-speed parallel computing, they still require complete transmission of raw data to servers, failing to truly break down the boundary between communication and computation. Early Over-the-air Computing approaches and systems like AirNN have explored the integration of communication and computing, yet their reliance on complex hardware architectures or specialized designs makes them difficult to deploy widely in standard wireless networks. How to seamlessly integrate neural network computing into standard wireless links without modifying existing communication protocols remains a critical challenge in the field of edge intelligence.
Key idea and contribution:
The paper proposes MetaAI, a novel wireless computing paradigm that directly integrates neural network computation into the process of wireless signal propagation. Its core innovation lies in utilizing programmable metasurfaces to transform the wireless channel itself into a configurable computational medium, thereby completing neural network computations directly during wireless signal transmission and achieving deep integration of communication and computation.
Key contributions of MetaAI include:
- Leveraging the properties of linear neural networks to decompose parallel neural computations into equivalent sequential operations, thereby resolving the fundamental mismatch between sequential wireless transmission and parallel neural computation.
- Implementing inference computations on a single reconfigurable metasurface while maintaining compatibility with standard wireless communication. The system supports not only single-modal inputs but also multi-modal, multi-sensor late-fusion via time-division multiplexing, enhancing robustness and accuracy for edge applications.
- Proposing a series of practical components, including multipath interference cancellation, a clock synchronization strategy (CDFA), and system noise mitigation, to ensure robustness in real-world wireless environments.
Evaluation
The authors implemented the MetaAI system using both dual-band (2.4/5 GHz) and single-band (3.5 GHz) metasurfaces, demonstrating its versatility across different wireless frequency bands. Extensive experiments show that even with a simple linear architecture, MetaAI achieves robust performance in various classification tasks, with an average recognition accuracy of 82.8% and a peak accuracy of 89.8%. Multi-sensor fusion further significantly enhances system performance: accuracy improves by up to 27.06% when combining heterogeneous sensors and by up to 25% when using multiple homogeneous sensors. MetaAI maintains its effectiveness even in challenging wireless environments with multipath effects and non-line-of-sight conditions. These results indicate that MetaAI does not merely trade accuracy for energy efficiency but instead pioneers a new Pareto-optimal point on the accuracy-energy efficiency curve.
Q&A
No
Personal Thoughts
This paper transforms the communication channel into a computational engine, materializing the vision of “transmission as computation.” The design is forward-thinking and particularly suitable for low-power IoT scenarios, while inherently offering privacy advantages. However, the paper also acknowledges limitations, such as support only for linear neural network models, computational accuracy constrained by the hardware resolution of metasurfaces, and potential latency overhead due to real-time reconfiguration requirements in response to device mobility-induced channel variations. Despite these constraints, its architectural innovation and conceptual inspiration far surpass current technical limitations.