Relay-Assisted Activation-Integrated SIM for Wireless Physical Neural Networks This work was supported by the SNS JU Project 6G-GOALS under the EU’s Horizon Program with Grant 101139232.
Meng Hua and Deniz Gündüz · 2026
Read Relay-Assisted Activation-Integrated SIM for Wireless Physical Neural Networks This work was supported by the SNS JU Project 6G-GOALS under the EU’s Horizon Program with Grant 101139232. by Meng Hua and Deniz Gündüz (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-04212.txt.
Introduction
Meng Hua and Deniz Gündüz proposed a relay-assisted wireless physical neural network built on activation-integrated stacked intelligent metasurfaces. Each passive surface layer handling linear wave shaping is paired with a layer that performs nonlinear work in the analog domain. Relay boost and phase-shift matrices together serve as trainable network weights, while hardware-built activation functions supply the needed nonlinear twist. Simulations showed high classification accuracy, with hardware activation clearly beating purely linear designs.
Catalog record
| CREATOR | Meng Hua and Deniz Gündüz |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 26382 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.04212
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Relay-Assisted Activation-Integrated SIM for Wireless Physical Neural Networks This work was supported by the SNS JU Project 6G-GOALS under the EU’s Horizon Program with Grant 101139232. Meng Hua and Deniz Gündüz Abstract Wireless physical neural networks (WPNNs) have emerged as a promising paradigm for performing neural computation directly in the physical layer of wireless systems, offering low latency and high energy efficiency. However, most existing WPNN implementations primarily rely on linear physical transformations, which fundamentally limits their expressiveness. In this work, we propose a relay-assisted WPNN architecture based on activation-integrated stacked intelligent metasurfaces (AI-SIMs), where each passive metasurface layer enabling linear wave manipulation is cascaded with an activation metasurface layer that realizes nonlinear processing in the analog domain. By deliberately structuring multi-hop wireless propagation, the relay amplification matrix and the metasurface phase-shift matrices jointly act as trainable network weights, while hardware-implemented activation functions provide essential nonlinearity. Simulation results demonstrate that the proposed architecture achieves high classification accuracy, and that incorporating hardware-based activation functions significantly improves representational capability and performance compared with purely linear physical implementations. Email: {m.hua, d.gunduz}@imperial.ac.uk I Introduction Wireless physical neural networks (WPNNs) have recently emerged as a promising paradigm that exploits native wireless propagation and hardware characteristics to perform inference directly in the physical layer [ 5 , 14 , 7 ] . In conventional digital neural network pipelines, inference is executed in the digital domain on finite-precision data tensors, while wireless transmission is abstracted as a separate bit pipe that delivers representations. As a result, end-to-end systems typically rely on analog-to-digital and digital-to-analog conversions at the transceiver, introducing non-negligible additional latency and energy consumption. By contrast, WPNNs leverage the wireless channel together with reconfigurable hardware components as trainable physical transformations, thereby enabling inference to be carried out directly over the propagation medium through over-the-air (OTA) processing. Among various candidates, two of the most promising trainable physical hardware platforms have attracted significant attention, namely metasurfaces [ 6 , 12 , 11 , 10 ] and relays [ 4 , 1 , 2 , 3 , 13 ] . Metasurfaces consist of a large number of subwavelength meta-atoms whose electromagnetic responses can be dynamically tuned, enabling programmable manipulation of wireless wavefronts, such as phase, amplitude, and polarization. For example, work [ 6 ] adopted reconfigurable intelligent surfaces (RISs) and optimized phase shifts to engineer the ambient wireless propagation environment so as to emulate the operations of a convolutional neural network layer. On the other hand, multiple distributed relays actively process and forward received signals by optimizing amplification gains, behaving as a fully-connected (FC) layer. For example, [ 1 ] treated the relay network itself, comprising a large number of distributed relays, as a neural network to implement communication task by training relay gains. However, most existing WPNN implementations are dominated by linear physical transformations, e.g., phase-only wave manipulation by RISs. Such linearity fundamentally limits the expressive capability of physical networks. In addition, metasurface-based implementations are inherently constrained by the unit-modulus phase-shift requirement, which further limits their neural expressiveness. To address the above limitations, this paper proposes a relay-assisted WPNN architecture enabled by activation-integrated stacked intelligent metasurfaces (AI-SIMs). The proposed architecture integrates linear wave manipulation and hardware-implemented nonlinear activation within a unified OTA computational framework. Specifically, each AI-SIM layer consists of a linear metasurface layer with independently programmable phase shifts, followed by an activation metasurface layer that realizes analog-domain nonlinear processing via radio-frequency detection and analog circuits. This cascaded design enables energy-efficient linear transformations together with intrinsic hardware nonlinearity, thereby overcoming the expressiveness limitations of purely linear physical implementations. In addition, the relay is elevated from a communication facilitator to an integral computational layer of the WPNN. By explicitly accounting for hardware-imperfect relay power amplifiers, whose nonlinear input–output characteristics naturally resemble neural activation functions, the relay provides flexible gain control and effectively emulates FC layers. Through structured OTA propagation, the metasurface phase-shift matrices and the relay amplification matrix jointly serve as trainable network weights, while hardware-implemented nonlinearities supply essential expressive power. Numerical results demonstrate that the proposed architecture achieves substantial accuracy gains over linear physical implementations, particularly in low signal-to-noise ratio (SNR) regimes. Fig. 1: Relay-assisted AI-SIM architecture for implementing WPNN. II System Model As illustrated in Fig. 1 , we consider a relay-assisted WPNN system in which both the transmitter and the receiver are equipped with AI-SIMs. Each AI-SIM comprises L passive metasurface layers for linear electromagnetic wave manipulation, where each passive layer is immediately followed by an activation metasurface layer that implements nonlinear signal processing in the analog domain. At the transmitter side, the information-bearing signals are radiated through an antenna array and subsequently processed by the AI-SIM. Similarly, at the receiver side, an AI-SIM is placed before the antenna array to process the impinging electromagnetic waves. The direct wireless link between the transmitter and receiver is assumed to be blocked, resulting in a non-line-of-sight (NLoS) propagation condition. To enable reliable communication, a relay node is deployed to establish a multi-hop transmission link from the transmitter to the receiver. The relay forwards the received signals between the transmitter- and receiver-side AI-SIMs, and can be treated as a physical FC layer. The WPNN is employed for classification tasks. Specifically, an input image, denoted by {\bf S}\in{{\mathbb{R}}^{C\times H\times W}} , where C , H , and W represent the number of color channels, height, and width, respectively, is transmitted via a single relay. Let N_{\rm t} , N_{\rm s} , and N_{\rm r} denote the numbers of transmit antennas, relay antennas, and receive antennas, respectively. For notational simplicity, we assume that the number of meta-atoms on each passive and activation metasurface is identical and equal to M . II-A Passive and Activation-based Metasurface Let {\mathbf{\Theta}}_{l}^{\text{t}}={\text{diag}}\left({{e^{j\theta_{l,1}^{\text{t}}}},\ldots,{e^{j\theta_{l,M}^{\text{t}}}}}\right)\in{{\mathbb{C}}^{M\times M}} , l=\left\{{1,\ldots,L}\right\} , denote the phase shift matrix of the l -th passive metasurface at the transmitter-side AI-SIM, where \theta_{l,m}^{\text{t}}\in\left[{0,2\pi}\right) is the phase shift of its m -th meta-atom. Similarly, denote by {\mathbf{\Theta}}_{l}^{\text{r}}={\text{diag}}\left({{e^{j\theta_{l,1}^{\text{r}}}},\ldots,{e^{j\theta_{l,M}^{\text{r}}}}}\right)\in{{\mathbb{C}}^{M\times M}} , l=\left\{{1,\ldots,L}\right\} , the phase shift matrix of the l -th passive metasurface at the receiver-side AI-SIM. Each passive metasurface layer is followed by an activation-based metasurface layer with M nonlinear meta-atoms. According to [ 8 ] , the activation-based metasurface layer can integrate radio-frequency detectors, amplifiers, and voltage adders, such that each nonlinear meta-atom can implement tanh-like behavior depending on the circuit configuration. To capture the resulting nonlinearity, we define \bar{\pi}\left(\cdot\right) as the nonlinear operator associated with the activation metasurface layer. II-B Channel Model We assume that each metasurface is arranged as a uniform planar array (UPA), while the transmit and receive antennas are configured as uniform linear arrays (ULAs). Note that the metasurfaces are closely stacked, such that inter-layer propagation occurs in the near-field regime. According to Rayleigh-Sommerfeld diffraction theory, the transmission coefficient from the n -th meta-atom on the \left({l-1}\right) -th metasurface layer to the m -th meta-atom on the l -th metasurface layer at the transmitter-side AI-SIM, l=\left\{{2,3,\ldots,2L}\right\} , denoted by {\left[{{{\mathbf{W}}_{l}^{\rm t}}}\right]_{m,n}} , can be characterized as \displaystyle{\left[{{{\mathbf{W}}_{l}^{\rm t}}}\right]_{m,n}}=\frac{{{A}\cos\phi_{m,n}^{(l)}}}{{d_{m,n}^{(l)}}}\left({\frac{1}{{2\pi d_{m,n}^{(l)}}}-\frac{j}{{{\lambda}}}}\right)\exp\left({j\frac{{2\pi d_{m,n}^{(l)}}}{{{\lambda}}}}\right), (1) where {d_{m,n}^{(1)}} denotes the transmission distance, \lambda represents the wavelength, A is the area of each meta-atom, and {\phi_{m,n}^{(l)}} represents the angle between the propagation direction and the normal direction of the \left({l-1}\right) -th metasurface layer. Denote by {\mathbf{W}_{1}^{\rm t}}\in{{\mathbb{C}}^{M\times{N_{\text{t}}}}} the transmission