Over-the-Air Edge Inference via End-to-End Metasurfaces-Integrated Artificial Neural Networks

TL;DR

The paper introduces MINNs, a framework that treats programmable metasurfaces (RIS/SIM) as end-to-end neural network layers to enable over-the-air edge inference. By training transceivers and metasurface controllers with backpropagation through fading channels, MINN achieves strong image classification performance under tight link budgets, even when channel state information is limited or unavailable. Key contributions include a detailed end-to-end system model, backpropagation-based training for channel-aware and channel-agnostic variants, and a comprehensive comparison against traditional and non-metasurface baselines, showing substantial power savings and robustness, especially with fixed metasurface configurations. This approach offers a practical path toward energy-efficient, low-complexity edge inference in dynamic wireless environments, leveraging the controllable propagation medium as a computational resource.

Abstract

In the Edge Inference (EI) paradigm, where a Deep Neural Network (DNN) is split across the transceivers to wirelessly communicate goal-defined features in solving a computational task, the wireless medium has been commonly treated as a source of noise. In this paper, motivated by the emerging technologies of Reconfigurable Intelligent Surfaces (RISs) and Stacked Intelligent Metasurfaces (SIM) that offer programmable propagation of wireless signals, either through controllable reflections or diffractions, we optimize the RIS/SIM-enabled smart wireless environment as a means of over-the-air computing, resembling the operations of DNN layers. We propose a framework of Metasurfaces-Integrated Neural Networks (MINNs) for EI, presenting its modeling, training through a backpropagation variation for fading channels, and deployment aspects. The overall end-to-end DNN architecture is general enough to admit RIS and SIM devices, through controllable reconfiguration before each transmission or fixed configurations after training, while both channel-aware and channel-agnostic transceivers are considered. Our numerical evaluation showcases metasurfaces to be instrumental in performing image classification under link budgets that impede conventional communications or metasurface-free systems. It is demonstrated that our MINN framework can significantly simplify EI requirements, achieving near-optimal performance with $50~$dB lower testing signal-to-noise ratio compared to training, even without transceiver channel knowledge.

Figures (8)

• Figure 1: A Metasurfaces-Integrated artificial Neural Network (MINN) performing Edge Inference (EI) by controlling the wireless propagation channel, which is treated as one or more hidden network layers.
• Figure 2: Block diagram and computation flow for the proposed framework where the metasurface-parametrizable channel acts as an intermediate component. Both the cases of reconfigurable and static metasurfaces are included, entailing different procedures during the forward and backward passes.
• Figure 3: implementation of the three modules of the proposed architecture for the considered MNIST classification problem. The channel matrices of $\mathbf{\mathcal{H}}(t)$ are flattened to vectors and are concatenated. The channel-aware branches are ignored when channel-agnostic transceivers are used.
• Figure 4: Comparison of achieved accuracy with different variations and the two adopted baselines, considering $N_t=4$ and channel-aware transceivers. Bars indicate the highest performance with each method across multiple training restarts, while the black horizontal lines represent the Q1-Q3 range, with each black vertical line representing the median performance.
• Figure 5: Similar to Fig. \ref{['fig:reconfigurable-vs-static-Nt2']} but for $N_t=12$.