Implementing Neural Networks Over-the-Air via Reconfigurable Intelligent Surfaces
Meng Hua, Chenghong Bian, Haotian Wu, Deniz Gündüz · 2025
Read Implementing Neural Networks Over-the-Air via Reconfigurable Intelligent Surfaces by Meng Hua, Chenghong Bian, Haotian Wu, Deniz Gündüz (2025) 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-2508-01840.txt.
Introduction
Meng Hua and co-authors turn a wireless channel itself into a neural network layer. Their 2025 paper builds on reconfigurable intelligent surfaces, panels that steer radio waves by tuning many tiny reflectors. By adjusting the surface together with the transmitter and receiver, the over-the-air signal path can imitate the weighted connections of a fully connected layer.
The team derives an alternating algorithm that fits all three parts with mostly closed-form steps. They then study two training styles. In one, a single side trains with full channel knowledge and shares the result. In the other, both sides train by passing signals back and forth. Channel symmetry lets that second route skip costly measurement. A multi-surface version adds spatial diversity gains. Simulations reach satisfactory classification accuracy, suggesting future networks might compute while they communicate, rather than only carrying data to separate processors.
Catalog record
| CREATOR | Meng Hua, Chenghong Bian, Haotian Wu, Deniz Gündüz |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 63 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-01840
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Implementing Neural Networks Over-the-Air via Reconfigurable Intelligent Surfaces Meng Hua, Chenghong Bian, Haotian Wu, Deniz Gündüz Abstract By leveraging the superposition property, over-the-air computation (OAC) of waveforms enables computations to be performed in an analog fashion in wireless environments, leading to faster computation, lower latency, and reduced energy consumption. In this paper, we investigate reconfigurable intelligent surface (RIS)-aided multiple-input-multiple-output (MIMO) OAC systems designed to emulate the fully-connected (FC) layer of a neural network (NN) via analog OAC, where the RIS and the transceivers are jointly adjusted to engineer the ambient wireless propagation environment to emulate the weights of the target FC layer. We refer to this novel computational paradigm as AirFC . We first study the case in which the precoder, combiner, and RIS phase shift matrices are jointly optimized to minimize the mismatch between the OAC system and the target FC layer. To solve this non-convex optimization problem, we propose a low-complexity alternating optimization algorithm, where semi-closed-form/closed-form solutions for all optimization variables are derived. Next, we consider training of the system parameters using two distinct learning strategies, namely centralized training and distributed training . In the centralized training approach, training is performed at either the transmitter or the receiver, whichever possesses the channel state information (CSI), and the trained parameters are provided to the other terminal. In the distributed training approach, the transmitter and receiver iteratively update their parameters through back and forth transmissions by leveraging channel reciprocity, thereby avoiding CSI acquisition and significantly reducing computational complexity. Subsequently, we extend our analysis to a multi-RIS scenario by exploiting its spatial diversity gain to enhance the system performance, i.e., classification accuracy. Simulation results show that the AirFC system realized by the RIS-aided MIMO configuration achieves satisfactory classification accuracy. Furthermore, it is shown that the multi-RIS system brings significant improvement in terms of the classification accuracy, especially in line-of-sight (LoS)-dominated wireless environments. I Introduction We are witnessing a rapid growth of data traffic due to the increasing popularity of edge devices. Integrated with artificial intelligence, these devices enable ubiquitous computing and ambient intelligence [ 1 ] . However, processing massive amounts of data at the edge devices presents significant challenges due to their limited computational resources and battery-powered energy constraints. To address these challenges, offloading computation or inference tasks from edge devices to resource-rich edge servers via wireless channels is gaining increasing attention [ 2 , 3 , 4 , 5 ] . In the conventional transmit-then-compute approach, the edge device transmits signals to the edge server via the wireless channel, where the server decodes and processes the signals. However, this method often suffers from excessive latency and poor spectrum utilization. A promising alternative, termed over-the-air computation (OAC) , exploits the superposition property of the wireless multiple access channel to perform simple algebraic operations, such as weighted sums, arithmetic means, Euclidean norms, etc., inherently combining computation with communication [ 6 , 7 , 8 , 9 , 10 ] . Thanks to its spectral and energy efficiency, OAC has been employed in distributed training [ 11 , 12 ] as well as inference tasks [ 13 , 14 ] . Typically, machine learning (ML) models are trained in the digital domain using graphics-processing units, e.g., [ 15 , 16 , 17 ] , which are time-consuming and energy-intensive due to the separation of memory and processing inherent in von Neumann architectures. To address this limitation, one promising solution is to train neural networks (NNs) using physical hardware based on optics, mechanics, and electronics, known as physical neural networks (PNNs), operating in the analog domain [ 18 , 19 , 20 ] . This approach builds deep PNNs composed of layers of controllable physical systems, where the computation traditionally performed in the digital domain is approximately replaced by the transmission of waves or optics through physical media. Since the computation naturally occurs over the air, with signal velocities approaching the speed of light, training PNNs can significantly reduce both latency and power consumption. Inspired by this idea, MIMO OAC has been explored for realizing fully-connected (FC) layers with fast and energy-efficient computations due to their structural similarities. On one hand, both the MIMO OAC paradigm and FC layers in digital NNs resemble point-to-point MIMO systems. On the other hand, both primarily involve simple algebraic operations such as multiplication and addition. To date, there has been only limited work investigating the use of the OAC paradigm to implement NN functions [ 21 , 22 , 23 , 24 ] . For instance, [ 21 ] identified a mapping between wave physics dynamics and a prototype recurrent NN, demonstrating that acoustic signals, by leveraging wave physics properties, can perform vowel classification on raw audio signals. Similarly, [ 22 ] applied the stochastic gradient descent method to train an analog NN by adjusting the conductance of programmable resistive diodes. In [ 23 ] , a linear function in the FC layer was realized via the superposition of orthogonal frequency-division multiplexing (OFDM) signals emitted by multiple transmitters and received at a single receiver, i.e., a multiple-input single-output (MISO) system. This was subsequently extended to a MIMO system in [ 24 ] , where the precoder and combiner were jointly trained over the air, achieving satisfactory classification accuracy. However, in the aforementioned works, the waveform design must adapt to the channel, indicating that system performance is heavily influenced by channel conditions, such as fading and channel rank. The emerging technology of reconfigurable intelligent surface (RISs) offers a new paradigm for wireless communications [ 25 , 26 , 27 , 28 , 29 , 30 ] , attracting considerable attention and being considered a candidate technology for sixth-generation (6G) networks. An RIS is a planar metasurface composed of massive low-cost passive reflecting elements, such as PIN diodes and varactor diodes, each of which can be digitally controlled to independently adjust the amplitude and/or phase shift of incident signals, thereby reshaping wireless signal propagation. Due to this appealing feature, RIS has been extensively studied and incorporated into various wireless systems, such as mobile edge computing [ 31 ] , sensing [ 32 ] , and integrated sensing and communication [ 33 ] . Naturally, RIS is also appealing for analog computation [ 34 , 35 , 36 , 37 , 38 ] . For instance, [ 34 ] presented a multi-layer RIS experimental prototype that accomplished a range of deep learning tasks, including image classification, mobile communication coding-decoding, and real-time multi-beam focusing. Subsequently, RIS-based over-the-air semantic communication was studied in [ 35 ] . In [ 36 ] , the authors proposed utilizing RISs to realize 1D convolution operations by exploiting multipath delays, where each channel impulse response acts as an individual finite impulse response filter, convolving with the transmitted signal to emulate a digital convolutional NN. Building upon this idea, the work in [ 37 ] further extended the concept to 2D convolution, enabling RISs to emulate more complex convolutional neural network structures for spatial signal processing. The authors of [ 38 ] designed a complex-valued neural network inspired by ordinary differential equations (ODEs), termed the Air-ODE network, which uses the physics of RIS-based signal reflections to execute computations during transmission. However, these prior works focus either on