Diffractive Neural Network for Robust Handwritten Digit Recognition

The diffractive neural networks presented here contained 200 × 200 diffractive neurons on each layer with a pixel size (pitch) of 0.25 ${\lambda }_{\max }$ . During the training, each handwritten digit of the MNIST dataset was first upscaled from 28 × 28 pixels to 70 × 70 pixels using bilinear interpolation and then padded with zeros to cover 200 × 200 pixels. The broadband illumination was digitally modeled as multiple independently propagating monochrome plane waves; we used ${\lambda }_{\min }=0.6\mathrm{mm}$ and ${\lambda }_{\max }=1.2\mathrm{mm}$ based on the THz part of the spectrum. The propagation and wave modulation on each spectral channel were separately computed. Four different randomly selected MNIST images formed a training batch, providing amplitude-only modulation to the input broadband light. Each input object batch was propagated and disturbed by one randomly selected diffuser. The four distorted broadband fields were separately propagated through the diffractive network, and the loss value (Eq. (12)) was calculated accordingly. The resulting loss was back-propagated, and the pixel height values were updated using the Adam optimizer⁵⁶ with a learning rate of $1\times 10^{-3}$ . Our models were trained using Python (v3.7.3) and PyTorch (v1.11) for 100 epochs, which took 5 h to complete. A desktop computer with a GeForce RTX 3090 graphical processing unit (GPU, Nvidia Inc.), an Intel® Core ™ i9-7900X central processing unit (CPU, Intel Inc.), and 64 GB of RAM was used.