Fig. 11

(a) Schematics of optical diffractive deep neural networks (D²NN). Multiple diffractive layers can work as a classifier (left for handwritten digits and fashion products) and an imager (right) (b) Error back-propagation algorithms for in situ optical training of diffractive ONNs. During each iteration, layer coefficients are constantly upgraded by four-step procedures marked as a sequence of forward propagation, error calculation, backward propagation, and gradient update. (c) A diagram of Fourier-space diffractive neural network (F-D²NN) worked for salient objective detection. The F-D²NN is implemented by placing a D²NN and optical nonlinearity (introduced from ferroelectric thin films SBN:60) at the Fourier plane. (d) Optical logic operations allowed by a diffractive neural network. Top: the layout of diffractive neural network including an input, hidden layers, and an output. Metasurface-based hidden layers could directionally scatter specially coded incident light into one of two regions (red labels of ‘1’ and ‘0’) at the output layer. Bottom: experimental intensity distributions at the output layer for NOT, OR and AND logic operations, respectively. Figures reproduced with permission: (a) Ref. [28], Copyright 2018 AAAS; (b) Ref. [123], Copyright 2020 OSA; (c) Ref. [124], Copyright 2019 APS; (d) Ref. [125], Copyright 2020 Nature Publishing Group