Laser & Optoelectronics Progress, Volume. 62, Issue 17, 1739004(2025)

Research Progress on Training Methods for Photonic Neural Networks (Invited)

Mengting Yu1, Haowen Zhao1, Shuiying Xiang1,2、*, Xingxing Guo1,2, Changjian Xie1, and Yue Hao2
Author Affiliations
  • 1State Key Laboratory of Integrated Services Networks, School of Telecommunications Engineering, Xidian University, Xi'an 710071, Shaanxi , China
  • 2National Engineering Research Center of Wide Band-Gap Semiconductor, School of Microelectronics, Xidian University, Xi'an 710071, Shaanxi , China
  • show less
    Figures & Tables(12)
    STDP-based training frameworks. (a) BP-STDP algorithm[41]; (b) general-purpose SNN based on VCSELs and VCSOA[42]; (c) S-PSNN based on STDP[46]; (d) training framework combining STDP and Tempotron algorithm[47]
    Some gradient-based training methods. (a) Supervised learning based on backpropagation[55]; (b) training algorithm based on physically-aware analytic gradient[56]; (c) SPGD algorithm[57]; (d) full forward mode learning[58]
    Evolutionary algorithm training framework. (a) Neuro evolutionary training strategies based on GA and PSO algorithms[62]; (b) on-chip in-situ training based on GA[63]; (c) two-stage non-situ training of SGD and GA combination[51]; (d) optical channel distortion equalization method based on PhRC structure and PSO algorithm[65]
    Hardware-aware training framework. (a) Quantitative perceptual training framework[68]; (b) training framework for noise perception[72]; (c) training framework combining quantization and noise perception[76]
    Error correction method based on hardware architecture optimization. (a) Robust crossbar switch Xbar architecture[79]; (b) Bokun grid architecture[81]; (c) symmetric silicon microring resonator optical crossover switch array[82]
    Two error correction methods based on training algorithm optimization. (a) Local error correction algorithm[86]; (b) asymmetric training algorithm[96]
    Fully self-directed online learning scheme. (a) Online learning based on the adjoint variable method[48]; (b) online learning based on optical error backpropagation algorithm[50]; (c) online learning based on in-situ backpropagation techniques[53]
    Hardware-in-the-loop training scheme. (a) Online learning of silicon photonic neural networks[103]; (b) optoelectronic devices for adaptive training methods[15]; (c) silicon photonics online training architecture with direct feedback alignment algorithm[105]
    Python training framework for photonic neural networks. (a) Schematic diagram of the photoelectric activation function based on the Neuroptica framework[108]; (b) L²ight architecture[111]; (c) NeurOLight architecture[113]; (d) ADEPT architecture[114]; (e) ELight architecture [115]
    • Table 1. Main training methods of four types of photonic neural networks and their functions

      View table

      Table 1. Main training methods of four types of photonic neural networks and their functions

      TypeMain training methodFunction
      Optical feedforward neural networkGradient-based trainingSolve the gradient calculation problem in weight optimization and improve the convergence speed and classification accuracy of the model
      Quantization-aware trainingReduce the model’s dependence on high-precision optics and maintain performance at low bit quantization
      Hardware error correctionCompensate for optical component manufacturing errors (e.g., phase offset) and improve hardware adaptability
      Optical recurrent neural networkGradient-based trainingOptimize the weight parameters in the loop structure to enhance the modeling ability of time series
      Noise-aware trainingImprove the robustness of the model to the intrinsic noise of optoelectronic devices, such as shot noise
      Online learningAdapt to dynamic environmental changes in real time and process non-stationary time series data
      Optical convolutional neural networkGradient-based trainingRealize the efficient update of optical convolution kernel parameters and improve the ability of feature extraction
      Joint hardware-algorithm optimizationCollaborative design of optical hardware architectures and algorithms to reduce computational complexity and power consumption
      Hardware error correctionCalibrate static errors in optical systems (such as uneven optical path loss) to improve the stability of feature extraction
      Optical spiking neural networkSpike-timing dependent plasticitySimulate the pulse delivery mechanism of biological neurons to achieve low-power, high-parallel information processing
      Evolutionary algorithmsGlobal search for optimal network connection configurations for solving complex optimization problems
      Online learningAdjust synaptic weights in real-time with a continuous stream of data to adapt to dynamic environmental changes
    • Table 2. Comparison of network architecture and performance of different technical methods on Iris dataset

      View table

      Table 2. Comparison of network architecture and performance of different technical methods on Iris dataset

      Technical approachNetwork architectureClassification accuracy /%
      BP-STDP algorithm4‒30‒396.0
      Generalized supervised learning algorithm based on VCSEL-SA24‒20‒196.0±1.3
      Enhanced VCSEL-SA photonic spiking neural network4‒294.67
    • Table 3. Classification performance of four methods under Sinusoidal activation function

      View table

      Table 3. Classification performance of four methods under Sinusoidal activation function

      Technical approachClassification accuracy /%
      MNISTCIFAR10
      8 bit6 bit4 bit2 bit8 bit6 bit4 bit
      Robust QAT91.21±0.4991.04±0.5190.26±0.7767.63±2.2867.64±1.2466.56±1.3829.48±10.43
      Two-stage QAT91.65±0.1291.67±0.4090.68±1.4179.31±2.0680.08±0.6077.81±0.4376.77±0.58
      Normalized post-training quantization of Gaussian distribution82.34±0.3481.52±0.9770.07±1.42
      Mixed-precision QAT92.87±0.0992.19±0.2273.61±5.8570.01±0.7965.69±2.30
    Tools

    Get Citation

    Copy Citation Text

    Mengting Yu, Haowen Zhao, Shuiying Xiang, Xingxing Guo, Changjian Xie, Yue Hao. Research Progress on Training Methods for Photonic Neural Networks (Invited)[J]. Laser & Optoelectronics Progress, 2025, 62(17): 1739004

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: AI for Optics

    Received: Jun. 11, 2025

    Accepted: Jul. 10, 2025

    Published Online: Sep. 15, 2025

    The Author Email: Shuiying Xiang (syxiang@xidian.edu.cn)

    DOI:10.3788/LOP251421

    CSTR:32186.14.LOP251421

    Topics