⚡ electrical engineering

Relay-Assisted Activation-Integrated SIM for Wireless Physical Neural Networks

This paper proposes a relay-assisted wireless physical neural network architecture utilizing stacked intelligent metasurfaces with integrated activation layers to overcome the expressiveness limitations of linear-only implementations by enabling nonlinear analog processing and trainable multi-hop propagation for high-performance neural computation.

Original authors: Meng Hua, Deniz Gündüz

Published 2026-04-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Meng Hua, Deniz Gündüz

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you want to send a secret message (like a picture of a dog) across a noisy room to a friend. Usually, you would shout the message, your friend would hear it, write it down, and then try to figure out what it means. This process is slow and wastes energy because you have to translate the sound into words and back again.

Wireless Physical Neural Networks (WPNNs) are a new way of doing this. Instead of translating the signal into digital code (like 1s and 0s) and then processing it, they let the signal itself "think" as it travels through the air.

However, there's a problem with current versions of this technology: they are too "linear." Think of it like a straight hallway. If you walk down a straight hallway, you can only go forward. You can't turn corners or make complex decisions. To solve a hard puzzle (like recognizing a picture), you need to be able to twist, turn, and change direction. In math terms, you need non-linearity.

This paper proposes a clever solution to build a "thinking hallway" using three main ingredients: Mirrors, a Relay, and a "Smart" Amplifier.

The Cast of Characters

The Stacked Intelligent Mirrors (AI-SIMs):
Imagine a stack of transparent glass sheets hanging in the air. Each sheet is covered in tiny, programmable "pixels" (meta-atoms).
- The Passive Sheets: These are like standard mirrors. You can tilt them to change the direction of the light (the signal), but they can't change the brightness or the shape of the message. They only do simple, linear math.
- The "Smart" Sheets (Activation Layers): This is the paper's big innovation. These sheets have special sensors that can detect how loud the signal is and then squash or stretch it. Imagine a sheet that says, "If the signal is too quiet, I'll boost it; if it's too loud, I'll turn it down." This creates a curve, not a straight line. This "squashing" is what allows the system to make complex decisions, just like the brain does.
The Relay (The Middleman):
Usually, a relay is just a walkie-talkie that repeats what it hears. But in this system, the relay is a "smart repeater."
- It receives the signal, cleans it up a bit, and then amplifies it.
- Crucially, the relay's amplifier isn't perfect. When the signal gets too strong, the amplifier naturally "saturates" (it hits a limit and stops getting louder).
- The authors realized this imperfection is actually a feature! This natural "clipping" acts like a second "Smart Sheet," adding another layer of complex thinking to the system.

How It Works: The "Thinking" Journey

Here is the journey of the image (the "Dog?"):

The Launch: The image is turned into a radio wave and shot out from the transmitter.
The First Twist: The wave hits the first set of Smart Mirrors. They twist the wave's direction and "squash" parts of it to highlight important features (like the dog's ears).
The Relay Stop: The wave flies to the Relay. The Relay listens, cleans up the static noise, and then shouts it back out. Because the Relay's amplifier has a "volume limit," it naturally distorts the signal in a helpful way, adding another layer of complexity.
The Second Twist: The wave flies to the Receiver's Smart Mirrors. They twist and squash the signal again, refining the picture.
The Decision: Finally, the signal hits the receiver's antenna. By this point, the wave has been twisted, turned, and squashed so many times that the pattern of the wave is the answer. The receiver just looks at the final shape and says, "That looks like a dog!"

Why Is This a Big Deal?

The paper tested this system against older, "straight hallway" (linear) systems. Here is what they found:

The "Noisy Room" Test: In a very noisy environment (low signal), the old linear systems failed miserably. They couldn't tell the difference between a dog and a cat. The new system, with its "Smart Mirrors" and "Smart Relay," kept working perfectly. It's like having a friend who can hear you clearly even when a jackhammer is running next to you.
The "Size" Test: They tried making the mirrors bigger (more pixels). Interestingly, just making the mirrors bigger didn't help much once the system was already smart enough. The "thinking" power came from the non-linear squashing (the activation), not just the size of the mirror.
The "Depth" Test: They added more layers of mirrors. They found that even with just one layer of mirrors, the system was already deep enough to solve the problem because the Relay acted as a hidden "brain layer" in the middle.

The Takeaway

This paper shows that we don't need to convert radio waves into digital code to process them. We can build a "physical brain" out of mirrors and amplifiers that thinks while the signal is flying through the air.

By intentionally using the "imperfections" of hardware (like the relay's volume limit) and adding special "squashing" mirrors, they created a system that is faster, more energy-efficient, and much better at handling noise than current technology. It's like upgrading from a simple walkie-talkie to a team of smart, talking mirrors that can solve puzzles for you before the message even reaches your ear.

1. Problem Statement

Wireless Physical Neural Networks (WPNNs) represent a paradigm where neural computation is performed directly within the physical layer of wireless systems, offering low latency and high energy efficiency by avoiding analog-to-digital (ADC) and digital-to-analog (DAC) conversions. However, existing WPNN implementations face two critical limitations:

Lack of Nonlinearity: Most current systems rely on linear physical transformations (e.g., phase-only manipulation by Reconfigurable Intelligent Surfaces or RIS). This linearity fundamentally restricts the network's expressiveness and ability to solve complex classification tasks.
Hardware Constraints: Metasurface-based implementations are often constrained by unit-modulus phase-shift requirements, further limiting their neural representational capability.
Noise Sensitivity: Purely linear physical networks struggle significantly in low Signal-to-Noise Ratio (SNR) environments.

The paper addresses the need for a WPNN architecture that integrates hardware-implemented nonlinear activation functions with linear wave manipulation to enhance computational expressiveness and robustness.

2. Methodology

The authors propose a Relay-Assisted Activation-Integrated Stacked Intelligent Metasurfaces (AI-SIM) architecture. The system operates entirely in the analog domain over the air (OTA).

System Architecture

Transmitter & Receiver: Both ends are equipped with AI-SIMs. Each AI-SIM consists of $L$ $L$ cascaded layers. Each layer comprises:
1. A Passive Metasurface Layer: Performs linear wave manipulation via programmable phase shifts ( $\Theta$ ).
2. An Activation Metasurface Layer: Immediately follows the passive layer and implements nonlinear processing in the analog domain using RF detectors and analog circuits (realizing a $\tanh$ -like function).
Relay Node: A relay is deployed between the transmitter and receiver to establish a multi-hop link (blocking the direct line-of-sight). The relay acts as a fully-connected (FC) layer.
- Crucially, the relay utilizes the intrinsic nonlinearity of its power amplifier (modeled by the Rapp model) to provide additional nonlinear activation, rather than treating it as a linear amplifier.
Signal Flow:
1. Input images are mapped to complex-valued signals.
2. Signals pass through the Transmitter AI-SIM (Linear $\to$ Nonlinear $\to$ Linear $\to$ Nonlinear...).
3. Signals propagate through the channel to the Relay.
4. The Relay performs Least Squares (LS) channel estimation to recover distorted signals, applies amplification (with inherent nonlinearity), and forwards them.
5. Signals pass through the Receiver AI-SIM (similar cascaded structure).
6. The final output is converted to real values, flattened, and fed into a standard digital FC layer with Softmax for classification.

Training Strategy

The entire system (metasurface phase shifts, relay amplification matrix, and bias vectors) is trained end-to-end using a Cross-Entropy Loss function. The trainable parameters include:

Phase shifts of passive metasurfaces ( $\Theta$ ).
Bias vectors injected at activation layers ( $b$ ).
Relay amplification matrix ( $Z$ ).
The system learns to optimize these physical parameters to minimize classification error.

3. Key Contributions

Activation-Integrated SIM (AI-SIM): The paper introduces a novel hardware architecture where passive linear metasurfaces are cascaded with active metasurfaces capable of analog-domain nonlinear activation. This overcomes the expressiveness limits of purely linear RIS.
Relay as a Computational Layer: The relay is elevated from a simple communication facilitator to an integral computational layer. By leveraging the nonlinear input-output characteristics of hardware power amplifiers, the relay effectively emulates a fully-connected layer with activation, adding depth and nonlinearity to the network.
Unified OTA Framework: The work demonstrates a unified framework where trainable weights are distributed across the channel (relay gains) and the hardware (metasurface phases), enabling end-to-end inference directly over the propagation medium.
Robustness in Low SNR: The architecture is specifically designed to maintain performance in noisy environments where linear physical networks fail.

4. Numerical Results

The authors evaluated the system using the Fashion-MNIST dataset (10 classes, 28x28 grayscale images) under various conditions:

Impact of Nonlinearity:
- At SNR = -20 dB, the proposed "Relay, nonlinear" scheme achieved 68.67% accuracy.
- In contrast, the "Relay, linear" scheme (without activation) dropped to 10.11% accuracy, proving that hardware-based nonlinearity is essential for low-SNR robustness.
Network Width (Meta-atoms $M$ ):
- The proposed scheme outperformed benchmarks when $M \le 36$ .
- Performance peaked at $M=16$ (accuracy $\approx 78\%$ ) and slightly degraded with larger widths, suggesting a trade-off between dimensionality gain and noise accumulation in OTA systems.
Network Depth (Passive Layers $L$ ):
- Increasing $L$ (depth) did not significantly improve accuracy beyond $L=1$ . The authors note that even with $L=1$ , the system effectively has 4 FC layers (2 from AI-SIM, 1 from Relay, 1 digital), which is sufficient for the task.
Comparison with Ideal Baselines:
- The proposed OTA scheme achieved performance comparable to an idealized "No OTA" baseline (direct connection without channel noise), and even outperformed it in very low SNR regimes due to the implicit robustness of the relay-assisted processing.

5. Significance

This work represents a significant step forward in 6G and physical layer machine learning.

Breaking Linearity: It solves the fundamental bottleneck of linear-only physical neural networks by integrating hardware-based nonlinearity directly into the wave propagation medium.
Energy Efficiency: By performing computation in the analog domain without ADC/DAC conversions at intermediate nodes, the system offers a path toward ultra-low-energy AI inference.
Hardware-Software Co-design: It demonstrates that hardware imperfections (like power amplifier saturation) can be repurposed as useful computational features (activation functions) when the system is trained end-to-end.
Practical Viability: The results confirm that WPNNs can operate reliably in challenging, noisy wireless environments, making them a viable candidate for future distributed and edge AI applications.