Fig. 1. MZI-based optical neural network architectures. (a) 4-channel optical network using MZIs^[42]; (b) 6-channel optical network using MZIs^[43]; (c) hybrid interference‒diffraction neural network architecture^[46]; (d) 64 input‒output channel hybrid interference‒diffraction neural network architecture^[47]

Fig. 2. MRR weight matrix-based optical neural networks. (a) MRR weight configuration^[57]; (b) optical neural network with MRR weight configuration^[59]; (c) schematic of all-optical spiking neurosynaptic network^[60]; (d) optical neural network based on optical frequency comb and PCM cells^[61]

Fig. 3. Integrated diffractive optical neural networks. (a) Diffractive optical network based on subwavelength metal arrays^[71]; (b) optical neural network with three-layer subwavelength diffractive structures^[72]; (c) diffractive optical network based on cascaded multimode interference architecture^[77]

Fig. 4. Optical neural networks based on integrated lithium niobate. (a) 4×4 programmable photonic integrated chip with MZI interferometric structures^[82]; (b) reconfigurable optical neural network using diffractive architecture^[83]; (c) optical neural network implementing all-optical nonlinear activation function via periodically poled waveguide^[84]

Fig. 5. Time-domain photonic processors based on integrated lithium niobate. (a) Microwave photonic processor^[91]; (b) matrix-vector multiplication with cascaded lithium niobate modulators^[92]; (c) photonic processor with 1×16 beam splitting structure^[93]