A Radio-Frequency-Based Fully Connected Layer using 1-Bit and 2-Bit Transmissive Intelligent Surfaces

This paper introduces a novel over-the-air computation method that utilizes low-complexity transmissive intelligent surfaces (TISs) for neural network inference. It is demonstrated that the signal propagation model through TIS closely resembles the fully connected layer of neural networks. And through training, the TIS phase shifts can be determined to perform a specific computation on radio-frequency (RF) signals. Considering the practical constraints of TIS designs with continuous phase shifts in millimeter-wave (mmWave) frequency, we propose a novel discretized complex-valued neural network structure and a training method suitable for low-complexity 1-bit and 2-bit TIS-based neural network layers. It is shown through simulation that the proposed method achieves high accuracy on an image classification task even for 1-bit or 2-bit TISs.