Dynamics And Neural-Like Computing Of Photonic Spiking Neural Network

Compared with the traditional von Neumann computing architecture,human brain shows strong parallel computing capability and low power consumption.Based on the physical observation of human brain,the efficiency of human brain may come from the complex connected structure,the structural-functional organization hierarchy and the complex regulation mechanism of neurons.The neuromorphic computing paradigm inspired by the basic computing mechanism of human brain provides a new way to overcome the "memory wall" and "power bottleneck" of Moore’s Law.Among them,spiking neural network(SNN),as the third generation neural network,has attracted much attention because of its biological plausibility,spatio-temporal coding dimension,low power consumption and event-driven characteristics.Computing machines of SNN based on microelectronics have surpassed the energy and dimensional efficiency wall of classical platforms,but still encounter the "electronic bottleneck" of bandwidth,speed,power consumption and crosstalk,etc.Because of the natural advantages in speed,bandwidth and power consumption,optical computing has been widely used in the field of information processing as a hardware acceleration platform.However,due to the difficulties of controlling and storing optical signals,the research of SNN computing based on optical platforms and devices is still in its infancy.Neural computing mainly concerns on the linear/nonlinear computing capabilities and the potential for large-scale integration.The vertical-cavity surface-emitting laser(VCSEL)has been widely concerned in the study of optical nonlinear computing because of its low power consumption,easy large-scale integration and rich dynamics.With the support of the national key research and development project and the National Natural Science Foundation of China,this paper focuses on the international hot spots of optical neural computing.Based on the theoretical model of two-section VCSEL with an embedded saturable absorber region(VCSEL-SA),the spiking dynamics and encoding mechanism of the photonic spiking neuron were firstly studied.Then the research was extended to coupled nodes,focusing on the collective dynamics and computational model of neural networks.The basic theory,computational architecture and application of photonic SNN were systematically investigated by combining theoretical analysis,numerical simulation and experimental demonstrations.This work promotes the development of theory and application of optical neural computing.The main contents and innovations include:1)Based on the theoretical model of VCSEL-SA,the neuron-like dynamic characteristics were firstly introduced,and the dynamic mechanisms of temporal coding and frequency coding were analyzed in depth;In addition,based on the self-pulsation characteristics of photonic spiking neurons,the collective dynamics in both small network motif and largescale networks were respectively studied by numerical models,demonstrating the decisive role of network topology on the dynamic behavior of spiking neurons.Through theoretical analysis and numerical simulation,the relationship between network symmetry and spatiotemporal synchronous spiking of photonic spiking neurons was revealed.Besides,based on four different large-scale networks such as global coupling network,non-local coupling network,small world network and sparse random network,the partial synchronous and partial asynchronous dynamic behavior was successfully demonstrated.Finally,based on the collective synchronization dynamics of pulse-coupled photonic spiking neurons,the feasibility of image segmentation based on photonic pulse-coupled neural network(PCNN)was discussed.2)Based on the temporal coding mechanism of VCSEL-SA neurons,the Hebbian-like algorithm and application of bio-inspired photonic SNN were studied,combined with the photonic synaptic plasticity based on the vertical cavity semiconductor optical amplifier(VCSOA).Several methods were proposed to optimize the performance of photonic SNN,and the algorithm-hardware collaborative computing was demonstrated through experimental platform.Firstly,based on the plasticity of synaptic afferent delay and efficacy,an efficient temporal-coding supervised learning algorithm based on delay-weight coadjustment was utilized to realize spike sequence learning and pattern classification.The adopted network is a feedforward SNN with a single output neuron.In addition,inspired by the existence of multiple synaptic links in biological neural networks,the influence of multiple synaptic links on the performance of photonic spiking neural networks was analyzed.At the same time,considering the limited classification performance of a single output node,the network was extended to multiple output dimensions,and the corresponding coding mechanism and training algorithm were also introduced.Finally,based on the complex topology structure of biological neural network,the photonic liquid state machine(LSM)calculation was realized for the first time in a randomly connected network,and achieved high precision pattern classification.3)Based on the modulation effect of pump current and external optical injection on the self-pulsation frequency of VCSEL-SA-based photonic spiking neurons,the frequency coding characteristics of photonic spiking neurons were studied,and an photonic SNN conversion method based on frequency coding was proposed.Firstly,the fitted optical activation function was used to train an artificial neural network(ANN).During the training process,the optical constraints were taken into account via limiting the range of input characteristics and weights.Then,the pre-trained data was mapped to the injected optical power or current to realize the conversion of photonic SNN.Based on several benchmark datasets,it can be proved that the conversion method can effectively improve the performance of photonic SNNs.In addition,cosidering hardware implementation,the impact of parameter mismatch and deviation on conversion performance was considered.Finally,the classification of Iris data-set was verified based on an photonic spiking neuron chip and the optoelectronic platform.4)Considering the hybrid-integration problem of Ⅲ-Ⅴ devices and the commercially mature complementary metal oxide semiconductor(CMOS)platform,the photonic SNN architecture based on silicon platform was studied.Firstly,the principle and defects of reconfigurable weight configuration based on the thermo-optical effect of microring resonators(MRR)were introduced.Then,based on the nonlinear effects of silicon MRRs,an photonic spiking neuron model based on an add-drop MRR was proposed,and its spiking dynamics and quasi-threshold characteristics were analyzed.In addition,a synaptic response mechanism based on an add-drop MRR was realized based on theoretical analysis.Then,an photonic SNN architecture based on MRRs was constructed,and the spike sequence learning task was realized based on the proposed delay-weight co-learning algorithm,which reveals the importance of the algorithm in the photonic SNN.In summary,this thesis focuses on the optical neural computing architecture,and studies the dynamics of spiking neural networks and neuromorphic computing based on the theoretical model of photonic spiking neurons.A variety of computational architectures based on photonic spiking neurons are proposed to acheive tasks such as spiking sequence recognition and pattern classification,and the collaboration between algorithms and hardware is demonstrated,which is of great significance to the development of photonic spiking neural computing.