Chinese Journal of Lasers, Volume. 48, Issue 19, 1906001(2021)

Recent Progress in Photonic Reservoir Neural Network

Lingyan Yang and Lin Zhang*
Author Affiliations
  • School of Precision Instrument and Opto-Electronics Engineering, Tianjin University, Tianjin 300072, China
  • show less
    Figures & Tables(21)
    Schematic of general reservoir neural network
    Curves of nonlinear functions[19]. (a) Tanh function; (b) steady state power transfer function of SOA
    Swirl topology[20]
    Waterfall topology with feedback loop[22]
    16-node passive photonic reservoir[21]
    Reservoir based on single photonic crystal cavity[23]
    Diffractively coupled spatial-light reservoir[26]
    Spatial-light reservoir based on scattering[28]
    Schematic of general delay-based reservoir neural network[9]
    Optoelectronic hybrid delay-based reservoir[14]
    Delay-based all-optical reservoir[36]
    Delay-based photonic reservoir with semiconductor laser as node[38]
    Error rate versus total input power for different injection scenarios[50]
    Four-port architecture[51]
    Delay-based photonic reservoir based on SRL[52]
    Double-reservoir structure[54]
    Delay-based photonic reservoir with optical input layer used for channel equalization[57]
    Implementation scheme of all-optical readout layer[59]
    • Table 1. Performance comparison among parallel optical reservoirs

      View table

      Table 1. Performance comparison among parallel optical reservoirs

      InstitutePublicationtimeImplementationNode typeNonlinearity mechanismOptical delayReservoir sizeBenchmark testingInput signal powerSimulation/experiment
      Ghent University [19]2008All-opticalSOAsSOA gain saturationDelay between nodes is 6.25 ps25Signal classification of triangular wave and rectangular wave, error rate (ER) is 2.5%5 mWSimulation
      Ghent University [20]2011All-opticalSOAsSOA gain saturationDelay between nodes is 6.25 ps81Speech recognition,ER is 4.5%-Simulation
      University of Athens [22]2013All-opticalInGaAsP/InP MRRsNonlinearity in MRRsDelay between nodes is 50 fs to 4.5 ps25Digital word classification,ER is 0.1% for 3 bit signal,ER is 0.5% for 8 bit signal10 mWSimulation
      Ghent University [21,30]2014All-opticalWaveguidesNonlinear photoelectric conversion in detectorsDelay between nodes is 280 ps16Boolean XOR task,bit error rate (BER) is 0.01%;5-bit header recognition15 mWExperiment
      2019Nonlinear channel equalization in single-wavelength metro and high-speed short-reach links,BER is <0.2%Simulation
      Ghent University [31]2014All-opticalNanophotonic crystal cavities--200Signal generation task,normalized root-mean-square error (NRMSE) is 0.03-Simulation
      Ghent University [23]2018All-opticalSingle photonic crystal cavityNonlinear photoelectric conversion in detectorsQ-factor of photonic crystal cavity is 164006(output layer)Boolean tasks,BER is <0.1%;6-bit header recognition, BER is <0.1%1 mWSimulation
      Universite Paris Saclay & Universite de Lorraine [32]2018All-opticalSilicon MRRsNonlinearity in MRRs-16Boolean XOR task, BER is 0.025%2.4 mWSimulation
      University Bourgogne Franche-Comté [25]2018OptoelectronicPixels of SLM, interconnected by diffractionFiltering by polarizing beamsplitter-900(output layer)Mackey-Glass chaotic time serie prediction, NRMSE is 0.042-Experiment
      Sorbonne University [27]2020OptoelectronicPixels of SLM, interconnected by scatteringNonlinear photoelectric conversion in detectors-4096M-G chaotic time serie prediction-Experiment
      The Aerospace Corporation [29]2020OptoelectronicPixels of SLM, interconnected by mode interferenceNonlinear photoelectric conversion in detectors-200(output layer)Speaker identification, accuracy is 98.1%-Simulation
      Speaker identification, accuracy is 81.5%Experiment
      University of Athens [33]2021All-opticalMRRsNonlinear conversion of phase modulated inputs to amplitude modulated outputs in MRRsDelay between nodes is 3 ps to 8 ps30Dispersion compensation of PAM-4 signal, BER is <0.1%-Simulation
    • Table 2. Performance comparison among delay-based optical reservoirs

      View table

      Table 2. Performance comparison among delay-based optical reservoirs

      InstitutePublicationtimeImplementationNode typeNonlinearity mechanismOptical delayReservoir sizeBenchmark testingInput signal powerSimulation/experiment
      Universite Bourgogne Franche-Comte [18]2012OptoelectronicMZMElectro-optic nonlinearity of MZMDelay time of feedback loop is 20.9 μs400Spoken digit recognition, ER is < 0.5%; Santa Fe chaotic time serie prediction, normalized mean squared error (NMSE) is 0.124-Experiment
      Université Libre de Bruxelles [14]2012OptoelectronicMZMElectro-optic nonlinearity of MZMDelay time is 8.5 μs50Nonlinear channel equalization; signal classification of square wave and sine wave, NMSE is 0.0015-Simulation
      200Speech recognition, ER is 0.4%
      Université Libre de Bruxelles [36]2012All-opticalSOASOA gain saturationDelay time is 7.9 μs50Spoken digit recognition, ER is 3%-Experiment
      Campus Universitat de les Illes Balears [37]2013All-opticalSemiconductor laserNonlinearity of laser with optical injection and feedbackDelay time is 77.6 ns388Spoken digit recognition, ER is 0.014%; speaker identification, ER is 0.88%; Santa Fe chaotic time serie prediction, ER is 5.5%<15 μWExperiment
      Universite Libre de Bruxelles [46]2014All-opticalSemiconductor saturable absorber mirrorSaturable absorption of semiconductorDelay time is 8.0073 μs50Nonlinear channel equalization,signal prediction-Experiment
      200Speech recognition, ER is 2.6%
      Universite Libre de Bruxelles [47]2015All-opticalPassive fiber cavityNonlinear photoelectric conversion in detectorDelay time is 1.13 μs50Equalization of standardized nonlinear multipath RF channel,for SNRs of more than 28 dB,symbol error rate is zero; spoken digit recognition<0.57 mWExperiment
      University of Saitama [38]2016All-opticalSemiconductor laserNonlinearity of laser with optical injection and feedbackDelay time is 40.1 ns400Santa Fe chaotic time serie prediction,NMSE is 0.008-Simulation
      InstitutePublicationtimeImplementationNode typeNonlinearity mechanismOptical delayReservoir sizeBenchmark testingInput signal powerSimulation/experiment
      Dalian University of Technology [42]2017OptoelectronicMZMElectro-optic nonlinearity of MZMDelay time is 9.7 ns4006-bit header recognition, the lowest ER is 1.25%7 dBm, input voltage amplitude is < 2 VExperiment
      Universite Paris Saclay & Universite de Lorraine [39]2018All-opticalVCSELNonlinearity of laser with optical injection and feedback; polarization mode couplingDelay time is 0.64 ns32Nonlinear channel equalization,BER is 0.001%0.08 mWSimulation
      Delay time is 8 ns400Santa Fe chaotic time series prediction,NMSE is 10-30.7 mW
      Southwest University [43]2018All-opticalTwo mutually delay coupled semiconductor lasersNonlinearity of laser with optical injection and feedbackDelay time is 2 ns200Santa Fe chaotic time serie prediction,NMSE is 5.1×10-5; signal classification of sine wave and square wave,NMSE is 5.5×10-4-Simulation
      Xidian University [44]2019All-opticalSemiconductor nanolaserDelay time is 0.1 ns100Santa Fe chaotic time serie prediction, the lowest NMSE is 0.036-Simulation
      Université Libre de Bruxelles [48]2019All-opticalSingle quantum cascade laserNonlinearity of laser with optical feedbackDelay time is 4 ns200Santa Fe chaotic time serie prediction, NMSE is 0.006; nonlinear channel equalization, ER is 0.1%-Simulation
    • Table 3. Performance comparison among serial-parallel optical reservoirs

      View table

      Table 3. Performance comparison among serial-parallel optical reservoirs

      InstitutePublicationtimeImplementationNode typeMultiplexingNonlinearity mechanismOptical delayReservoir sizeBenchmark testingSimulation/experiment
      Vrije Universiteit Brussel [52]2015All-opticalSemiconductor ring laser (SRL)Two directional optical modes (CW and CCW) in SRLNonlinearity of laser with optical injection and feedback--Santa Fe chaotic time serie prediction, NMSE(CW) is 0.031, NMSE(CCW) is 0.040;equalization of a wireless communication channel, BER(CW) is 0.14%, BER(CCW) is 0.15%Simulation
      Xidian University [53]2019All-opticalVCSELTwo polarization-resolved modes in VCSELDelay time is 0.5 ns50Santa Fe chaotic time serie prediction;signal classification of sine wave, square wave and triangular waveSimulation
      Southwest University [54]2021OptoelectronicSemiconductor lasersTwo serial reservoirs are connected in parallel-100Santa Fe chaotic time serie prediction, NMSE is 0.007, processing speed is faster than that of single serial reservoirSimulation
      NTT [56]2021All-opticalPassive cavities on chipSeveral serial reservoirs are connected in parallelNonlinear photoelectric conversion in detectorsDelay time is (290+20)×l ps, l is cavity number512Santa Fe chaotic time serie prediction, NMSE is 0.06;image classification based on MNIST, accuracy is 91.3%Experiment
      Wavelength division multiplexingImage classification, accuracies are comparable to those of single wavelength experiments
      University of West Attica [15]2021All-opticalF-Plaser8 longitudinal modes in F-P laserDelay time is 1 ns50Nonlinear channel equalization of PAM-4 25 Gbaud signals, different tasks can be processed in parallelSimulation
    Tools

    Get Citation

    Copy Citation Text

    Lingyan Yang, Lin Zhang. Recent Progress in Photonic Reservoir Neural Network[J]. Chinese Journal of Lasers, 2021, 48(19): 1906001

    Download Citation

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

    Category: fiber optics and optical communications

    Received: Jul. 21, 2021

    Accepted: Sep. 2, 2021

    Published Online: Sep. 28, 2021

    The Author Email: Lin Zhang (lin_zhang@tju.edu.cn)

    DOI:10.3788/CJL202148.1906001

    Topics