Investigation Of Optical Neurons And Learning Mechanism

With the advent of the era of big data,the amount of data processed by computers is becoming larger,and the types of problems that need to be solved are more abundant.However,the traditional electronic computer based on Von Neumann structure is restricted to the channel between storage and calculator,and there are bottlenecks in the improvement of computing power of traditional electronic computer;hence,it appears to be helpless in the face of pattern recognition,automatic control,learning etc.Neuromorphic computing aims at mimicking the behavior of human brain using non-Von Neumann structure.It has the same advantages as the human brain,including parallelism,redundancy;thus,it has the natural advantage in solving the above problems.Photonies,characterized by its high bandwidth,high parallelism and low energy consumption,is an ideal way to realize neuromorphic computing.The key of photonics neuromorphic computing lies in the implementation of single optical neuron,optical neural network learning mechanism and large-scale optical neural network architecture.In the Past decade,photonics neuromorphic computing has become a research hotspot,but it is still in its infancy.In this dissertation,the research of optical neurons and learning mechanism is carried out under the support of National Natural Science Foundation of China(NSFC).We deeply studied several different optical neuron models,including continuous optoelectronic sigmoid neuron and optical LIF(leaky-integrate-fire)neuron based on bulk devices,integrator spiking neuron and resonator spiking neuron based on integrated optical devices.And we studied the learning mechanism of the optical spiking neuron.This research is crucial to the development of photonics neuromorphic computing.The main works of this dissertation are as follows.1.Based on the nonlinear polarization rotation effect(NPR)of semiconductor optical amplifier(SOA),a continuous optoelectronic sigmoid neuron is demonstrated and can realize sigmoid transfer function.Excitatory and inhibitory stimuli to the proposed artificial neuron can be implemented by optical injection and electrical modulation,respectively.We experimentally studied the effect of SOA's bias current,probe beam's power and initial polarization state on the transfer function of the proposed neuron.2.Based on the NPR effect of SOA,an optical LIF neuron is proposed.When the initial polarization state of the probe beam is correctly set,the excitatory and inhibitory neuron can be mimicked.And based on the cross-gain modulation effect of SOA,the optical spike-timing-dependent plasticity(STDP)circuit is proposed.A variety of STDP learing mechanisms can be implemented.The effects of SOA bias current,optical power and pulse-width on learning window and window height were studied experimentally.3.Based on the Q-switching effect of VCSEL-SA structure,an optical spiking neuron is proposed using coherent and incoherent beam as excitatory stimuli.The bifurcation features of VCSEL-SA spiking neuron with different driving current of gain region and SA region,as well as excitatory working conditions,were studied.Then we numerically studied the spiking characteristics of VCSEL-SA spiking neuron under single pulse or two pulses injection,including threshold,response time,response spike's amplitude and refractory period.4.By covering the phase change material VO2 on the top of micro-ring resonator,the threshold plasticity for spiking rings is proposed.We studied the bifurcation features of spiking rings and established the PI hysteresis model of the VO2.Combined with the coupled mode theory of the micro-ring resonator,the effects of initial operating temperature,VO2 cover length and perturbation amplitude on the threshold value of the micro-ring pulse neurons are studied.