Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AirCNN via Reconfigurable Intelligent Surfaces: Architecture Design and Implementation

Meng Hua, Haotian Wu, Deniz Gündüz

TL;DR

AirCNN addresses the challenge of implementing CNN layers directly in the wireless edge by shaping the ambient channel with RISs and performing over-the-air analog computation. The approach jointly trains transmitter precoders, RIS phase shifts, and receiver combiners to emulate Conv2d and ConvSD kernels, using both MISO and MIMO RIS-aided architectures. Key findings show that Conv2d with MISO generally outperforms MIMO, ConvSD benefits from low-power regimes, and deploying multiple RISs yields substantial gains in LoS-dominated channels. This framework demonstrates a viable path to low-latency, energy-efficient edge inference by performing neural computations over the wireless medium, leveraging the DoFs provided by RISs and OTA propagation.

Abstract

This paper introduces AirCNN, a novel paradigm for implementing convolutional neural networks (CNNs) via over-the-air (OTA) analog computation. By leveraging multiple reconfigurable intelligent surfaces (RISs) and transceiver designs, we engineer the ambient wireless propagation environment to emulate the operations of a CNN layer. To comprehensively evaluate AirCNN, we consider two types of CNNs, namely classic two-dimensional (2D) convolution (Conv2d) and light-weight convolution, i.e., depthwise separable convolution (ConvSD). For Conv2d realization via OTA computation, we propose and analyze two RIS-aided transmission architectures: multiple-input multiple-output (MIMO) and multiple-input single-output (MISO), balancing transmission overhead and emulation performance. We jointly optimize all parameters, including the transmitter precoder, receiver combiner, and RIS phase shifts, under practical constraints such as transmit power budget and unit-modulus phase shift requirements. We further extend the framework to ConvSD, which requires distinct transmission strategies for depthwise and pointwise convolutions. Simulation results demonstrate that the proposed AirCNN architectures can achieve satisfactory classification performance. Notably, Conv2d MISO consistently outperforms Conv2d MIMO across various settings, while for ConvSD, MISO is superior only under poor channel conditions. Moreover, employing multiple RISs significantly enhances performance compared to a single RIS, especially in line-of-sight (LoS)-dominated wireless environments.

AirCNN via Reconfigurable Intelligent Surfaces: Architecture Design and Implementation

TL;DR

AirCNN addresses the challenge of implementing CNN layers directly in the wireless edge by shaping the ambient channel with RISs and performing over-the-air analog computation. The approach jointly trains transmitter precoders, RIS phase shifts, and receiver combiners to emulate Conv2d and ConvSD kernels, using both MISO and MIMO RIS-aided architectures. Key findings show that Conv2d with MISO generally outperforms MIMO, ConvSD benefits from low-power regimes, and deploying multiple RISs yields substantial gains in LoS-dominated channels. This framework demonstrates a viable path to low-latency, energy-efficient edge inference by performing neural computations over the wireless medium, leveraging the DoFs provided by RISs and OTA propagation.

Abstract

This paper introduces AirCNN, a novel paradigm for implementing convolutional neural networks (CNNs) via over-the-air (OTA) analog computation. By leveraging multiple reconfigurable intelligent surfaces (RISs) and transceiver designs, we engineer the ambient wireless propagation environment to emulate the operations of a CNN layer. To comprehensively evaluate AirCNN, we consider two types of CNNs, namely classic two-dimensional (2D) convolution (Conv2d) and light-weight convolution, i.e., depthwise separable convolution (ConvSD). For Conv2d realization via OTA computation, we propose and analyze two RIS-aided transmission architectures: multiple-input multiple-output (MIMO) and multiple-input single-output (MISO), balancing transmission overhead and emulation performance. We jointly optimize all parameters, including the transmitter precoder, receiver combiner, and RIS phase shifts, under practical constraints such as transmit power budget and unit-modulus phase shift requirements. We further extend the framework to ConvSD, which requires distinct transmission strategies for depthwise and pointwise convolutions. Simulation results demonstrate that the proposed AirCNN architectures can achieve satisfactory classification performance. Notably, Conv2d MISO consistently outperforms Conv2d MIMO across various settings, while for ConvSD, MISO is superior only under poor channel conditions. Moreover, employing multiple RISs significantly enhances performance compared to a single RIS, especially in line-of-sight (LoS)-dominated wireless environments.

Paper Structure

This paper contains 10 sections, 9 equations, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. System model
  3. Conv2d-based Physical Neural Network
  4. Conv2d MISO
  5. Conv2d MIMO
  6. ConvSD-based Physical Neural Network
  7. ConvSD MISO
  8. ConvSD MIMO
  9. Numerical results
  10. Conclusion

Figures (7)

  • Figure 1: Illustration of conventional and OTA CNN architectures.
  • Figure 2: A toy example of transforming a convolutional operation to a matrix multiplication operation.
  • Figure 3: Multi-RIS aided OTA transmission neural network architecture.
  • Figure 4: Transmit power $P_{\rm max}$ versus classification accuracy.
  • Figure 5: Rician factor $K$ versus classification accuracy.
  • ...and 2 more figures