Fig. 1: Multi-task processing with MN-DNN.
From: Multilayer nonlinear diffraction neural networks with programmable and fast ReLU activation function
a Schematic of the MN-DNN operation. The network is able to perform both dataset-based image recognition and real-time human posture classification. Incident EM waves, encoded with input information, are processed by the MN-DNN comprising three linear and three nonlinear layers. Each nonlinear unit integrates an RF detector, an amplifier, and a voltage adder. The classification result corresponds to the location of the maximum energy on the output plane. b Input-output response of the nonlinear metasurface unit, exhibiting a ReLU-like characteristic with a dynamically adjustable threshold and slope. The output is negligible for weak inputs and increases linearly beyond the threshold. c Experimentally measured time delay of a single nonlinear metasurface layer (17.7 ns). Experimental demonstration of classification capabilities on the MNIST (d), Fashion-MNIST (e) and human postures (f) tasks.
