Fig. 1. Concepts of artificial neural networks, and development timeline of optical neural networks. (a) Development timeline of optical neural networks; (b) features of structures and signals of biological neural networks^[18] and basic principle of artificial neural networks (x₁, x₂, and x₃ represent input signals, and w₁,w₂, and w_n represent calculation weights); (c) general categories of artificial neural networks

Fig. 2. Optical neural networks based on traditional optical devices. (a) Holography neural network^[32]; (b) optical neural network based on spatial light modulator^[34]

Fig. 3. Examples of nanophotonics neural networks based on micro/nano optical waveguides. (a) Prototype of optical neural network and experimental diagram^[37] (left: principle diagram; middle: experimental configuration; right: signal transmission image and scanning electron microscopic image of experimental setup); (b) structural design of a novel computing chip using optical signal as carrier^[38] (top-left: schematic diagram of artificial neural network, which includes input layers, hidden layers, and output layers; top-right: layered demonstration of artificial neural network; bottom: full optical integratable neural networks)

Fig. 4. Examples of nanophotonics neural networks based on diffractive elements. (a) Diffractive neural networks^[39](left: diffractive neural networks based on multi-layer transitive/reflective structures; right: experimental demonstration of training process using multi-layer diffractive neural networks); (b) schematic diagram of multi-layer photonic chip intergrated with CMOS fabricated by two-photon 3D laser etching technique^[45] ( middle: diagram of two photon laser processing using galvanometer scanning; right: inner structure of multi-layer nanoscale diffractive neural networks. The smallest resolution is 10 nm)

Fig. 5. Examples of inverse designs of topology optimization methods. (a) Inverse design of Z-shape photonic crystal waveguide by topology optimization method^[58]; (b) inverse design of wavelength splitter by gradient descent method^[59]; (c) inverse design of polarization beam splitter by direct search method^[60]

Fig. 6. Examples of design of plasmonic devices based on multi-layer perceptron. (a) Plasmonic nanostructures designed using deep learning model based on bidirectional multi-layer perceptron^[65]; (b) inverse design of chiral metamaterials designed using bidirectional multi-layer perceptron neural network^[66]

Fig. 7. Switch between long-term and short-term memory modes of photonic synapses under adjustment of number of pulses and frequency^[73]. (a) Change of number of pulses; (b) change of frequency

Fig. 8. Paired-pulse facilitation of photonic synapses^[73]. (a) Current change curves; (b) relationship between PPF efficiency (A₂/A₁) and pulse interval

Fig. 9. Learning-experience behavior of photonic synapses^[73]. (a) 1^st learning; (b) 1^st forgetting; (c) 2^nd learning; (d) 2^nd forgetting

Fig. 10. Plasticity of photonic synapses achieved based on excitation post synaptic currents and inhibition post synaptic currents (tuning applied electric voltage)^[73]

Fig. 11. Structures and performances of photonic synapses. (a) Photonic synapses based on indium gallium zinc oxides (left), and curve of electric current change under optical pulse excitation (right)^[83]; (b) principle of conductivity control of photonic synapses based on CH₃NH₃PbI₃ driven by photovoltalic effect^[89]; (c) structure of photonic synapses based on light induced grating control effect using carbon nanotubes (left) and electric current change curve (right)^[93]