Fig. 1. Schematic diagram of optical diffraction deep neural network

Fig. 2. Diffractive optical neural network architecture based on polarization multiplexing.(a) System structure and integrated architecture diagram of MDNN^[24]; (b) schematic diagram of a polarization-encoded diffraction neural network^[25]; (c) schematic diagram and theoretical architecture of M-D₂NNs^[26]

Fig. 3. Diffraction neural networks based on parallel wavelength multiplexing. (a) Architecture of optical multitask learning for multiwavelength D2NNs^[33]; (b) performance analysis of four-wavelength D2NNs; (c) multi-frequency-channel diffractive deep neural network^[34]; (d) training learning curves of the corresponding FMNIST and EMNIST

Fig. 4. Diffractive optical neural network based on mechanical reconstruction. （a) Conceptual diagram of multi-task P-DNN^[35]; (b) confusion matrix of MNISIT simulation and experiment for P-DNN^[35]; (c) system diagram of RubikONNs diffraction neural network^[36]; (d) RubikONNs using specific rotations to achieve four-task MTL image classification^[36]

Fig. 5. Diffraction neural networks based on angular momentum multiplexing.(a) Structure schematic of the polarization depth diffraction neural network^[44]; (b) rectangular columnar unit cell structure; (c) classification of polarized vortex beams for different topological charges, realization of the occurrence of different topological charges and classification results

Fig. 6. Diffraction neural network chip integration solution.(a) Schematic of the CMOS-integrated all-optical machine learning decoder^[45]; (b) schematic diagram of integrated diffraction deep neural network^[47]; (c) principle diagram and corresponding framework of the OENN^[46]; (d) schematic diagram of 3D photonic neural network based on VCSEL array^[48]

Fig. 7. Diffraction neural network applied to image recognition and classification. (a) Schematic illustration of the diffractive optical network for all-optical classification of overlapping phase objects^[52]; (b) architecture of the programmable D2NN platform^[53]

Fig. 8. Diffraction neural network applied to complex vision tasks. (a) Schematic illustration of the diffractive neural network for implementing optical logic operations^[54]; (b) full optical imaging achieved through a random diffuser based using the diffractive neural network^[56]; (c) conceptual schematic of the optical triangulation operator^[55]; (d) integration of metasurfaces on CMOS to realize spectral convolutional neural networks^[57]