Fig. 3: Handwritten digits classification.
From: Multilayer nonlinear diffraction neural networks with programmable and fast ReLU activation function
a Example images from the MNIST dataset, which comprises 10 digit classes (0–9). b Encoding binarized images into the transmission coefficients of the input metasurface. c Experimental test setup. d Schematic of the two network architectures: a nonlinear network with interleaved linear and nonlinear layers, and a linear network solely with linear layers. e t-SNE visualization of test set outputs for both networks, showing clearer class separation and tighter clustering in the nonlinear network. f Confusion matrices show 92.6% (nonlinear) and 88.5% (linear) accuracy on 5000 test images. g Sampled output field distributions generated by the linear and nonlinear networks for a simple image, demonstrating that both networks focus on the target region. h Sampled output field distributions for two complex images from both networks, in which the nonlinear network focuses on the correct regions while the linear network focuses on incorrect regions.
