Fig. 1. Optical neural networks (ONNs) based on optical matrix multiplication. (a) Optical matrix multiplier based on lens and cylindrical lens^[9]; (b) implementation of Hopfield optical neural network based on optical matrix multiplication^[11]; (c) optical neural networks based on volume hologram^[13]; (d) optical matrix multiplication based on MRR array^[19]; (e) optical matrix multiplication based on MRR and PCM units^[23]; (f) optical matrix multiplication based on microcombs and PCMs^[25]; (g) ONNs based on Mach-Zehnder interferometer (MZI)^[33]; (h) MZI-based ONNs supporting in situ training^[35]; (i) optical matrix multiplication based on time and wavelength interleaving^[38]

Fig. 2. Optical neural networks based on light diffraction. (a) Diffractive deep neural networks (D2NN)^[43]; (b) Fourier-space D2NN in visible regime^[44]; (c) optoelectronic hybrid neural networks with reconfigurable diffractive unit^[46]; (d) optoelectronic analog chip based on diffractive neural network^[50]; (e) multiscale U-Net diffractive neural network^[52]; (f) multi-task diffractive neural network based on wavelength multiplexing^[54]

Fig. 3. Optical convolutional neural networks (CNNs). (a) Receptive field mechanism in biological vision^[62]; (b) initial convolutional neural network Neocognitron^[64]; (c) optical convolutional layer based on 4f system^[69]; (d) high-speed convolutional computing based on 4f system and DMD^[70]; (e) all-optical convolutional neural network^[71]

Fig. 4. Optical neural networks fabricated by two-photon laser 3D printing. (a) Photon chips fabricated based on three-dimensional two-photon laser printing technology^[45]; (b) image encryption and decryption based on photonic chips fabricated by printing on CMOS surfaces^[76]; (c) photonic chip integrated on CMOS surface performing pupil phase retrieval^[77]; (d) scattering image reconstruction based on diffraction neural network fabricated at optical fiber end face^[78]

Fig. 5. Optical neural networks based on metasurface. (a) Optical logic operation unit based on metasuface^[81]; (b) reconfigurable and programmable metasurface^[82]; (c) super-resolution detection based on diffraction neural network^[83]; (d) waveform-multiplexed metasurface optical diffraction neural network^[57]; (e) 3D depth structure perception based on encoding and decoding metasurfaces^[85]; (f) optical 1D on-chip diffractive neural networks^[91]