Wireless Networks Act as Deep Physical Neural Networks.

This research explores the concept of Wireless Physical Neural Networks (WPNNs), which embed neural computations directly into analog wireless hardware, promising significant reductions in energy consumption and latency compared to traditional digital AI implementations. The paper introduces a deep WPNN architecture built upon a multi-hop Multiple-Input Multiple-Output (MIMO) relay network. In this design, each relay functions as a trainable complex linear gain and bias unit, with the crucial innovation being the use of the power amplifier's inherent nonlinearity as the activation function. This cascade of relays effectively creates an over-the-air fully connected network whose parameters can be trained end-to-end. The researchers developed two transceiver designs for different channel state information (CSI) scenarios, demonstrating accurate over-the-air inference for image classification and highlighting the benefits of exploiting hardware nonlinearity for enhanced computational capabilities.