Fig. 1. Biological neuron、artificial neuron and common activation functions. (a) Schematic diagram of biological neuron; (b) schematic illustration of artificial neuron; (c) schematic representation of fully connected neural network; (d) common activation function profiles and their mathematical expressions

Fig. 2. Optical fully connected neural network based on cascaded MZI. (a) Two-layer cascaded MZI optical fully connected neural network incorporating OIU and ONU^[72]; (b) complex-valued optical fully connected neural network architecture^[73]; (c) packaging architecture and hardware-software collaborative processing workflow of Envise^[75]

Fig. 3. Optical fully connected neural network based on MRM arrays. (a) Neural network models and the optical implementation method^[77]; (b) core structure and working principle of complex-valued MVM based on MRM arrays^[78]; (c) structure and working principle of 4×4 cross-bar MRM array^[79]

Fig. 4. Optical fully connected neural network based on diffractive metasurfaces. (a) Schematic of conventional ANN and proposed optical neural network^[81]; (b) neural network architecture integrating diffraction units with MZI arrays^[83]; (c) structure schematic of on-chip diffractive optical neural network^[85]; (d) computational process of execution unit and large-scale clustering in Taichi chip’s distributed computing architecture^[86]

Fig. 5. Optical convolutional neural network based on MRM arrays. (a) Conceptual schematic of PTFP chip^[89]; (b) schematic diagram of parallel edge extraction unit^[92]; (c) integrated photonic tensor core for parallel convolutional processing^[93]; (d) experimental setup and processing workflow of photonic convolution accelerator for image processing^[94]

Fig. 6. Multiple implementation approaches for optical convolutional neural network. (a) Optical convolutional neural network based on PIN current-controlled attenuators^[97]; (b) structure diagram of compact optical convolutional processing unit^[98]; (c) schematic representation and working principle of PCNC matrix core^[99]; (d) one-dimensional convolution window sliding based on AWGR^[100]; (e) diffraction-driven multi-kernel optical convolution unit chip^[101]

Fig. 7. In-situ training schemes for optical neural network. (a) Schematic diagram of SPGD in-situ training experimental setup and optical processing chip^[107]; (b) principle and experimental scheme of in-situ intensity measurement-based gradient calculation^[108]; (c) optical and electrical components for in-situ trainable diffractive neural networks^[109]; (d) architecture of fully integrated coherent optical neural network, including input/output, nonlinear activation, and in-situ training process^[110]

Fig. 8. Electro-optical activation functions. (a) Experimental setup diagram of MRM-based O-E-O activation function^[112]; (b) optoelectronic neurons and absorption curves of different materials based on EAM^[113]; (c) core components of electro-optic nonlinear activation function^[115]; (d) device structure comprising Ge-Si photodetecter with MZI/MRM-type EO switch^[116]

Fig. 9. All-optical activation functions. (a) Schematic of device comprising MRM loaded on MZI and realized activation function profiles^[119]; (b) SiGe photodetector integrated with nonlinear activation and in-situ monitoring functions^[121]; (c) three-dimensional schematic of graphene/silicon heterojunction device and working principle^[123]; (d) schematic of thermally-controlled photodetector and activation function profiles at different temperatures^[124]

Fig. 10. Common architectures and applications of optical neural networks^{[72-73,76-77,85,96,99,107,109,114,121-122,126-129]}