Investigation On Spike Time-dependent Plasticity With Weight-dependent Learning Window Based On VCSOA

Neural network is widely used in many fields such as pattern recognition,biomedicine and economy,which greatly promotes the development of artificial intelligence.Among them,the photonic spike neural network has certain advantages over the traditional neural network in information processing speed and energy consumption.As an important information processing unit in photonic spike neural networks,photonic synapses can simulate the spike time-dependent plasticity(STDP)learning mechanism related to learning and memory in human brain and become one of the hot research topics.In particular,the vertical cavity semiconductor optical amplifier(VCSOA)is an ideal photonic synapse device due to the advantages of easy integration and low power consumption.Therefore,it is great significance to study STDP learning mechanism based on VCSOA for the future construction of high performance photonic spike neural network.In this thesis,nonlinear dynamic characteristics of VCSOA are studied theoretically,and the influence of system parameters on the STDP learning mechanism based on VCSOA is discussed in detail.The results show that,with the increase of the spike injection power,the carrier consumption of the system also increases and the total recovery time of the carrier is almost the same.For optical STDP learning mechanism based on VCSOA,the height of STDP curve window increases with the increase of system bias current.For relative large cavity volume,the height of the STDP learning curve window will increase with the increase of the system internal loss.when the cavity volume decreases to a certain level,the fixed internal loss has little influence on the STDP learning curve window.For a certain volume of the cavity,the increase of the linear gain coefficient of the system will also induce to increase of the height of the STDP learning curve window,but when the cavity volume further decreases a certain level,the height and width of the STDP learning curve window will significantly decrease with the increase of the linear gain coefficient of the material.Moreover,we further introduced feedback signals into the system to realize STDP learning mechanism with weight-dependent learning window,which induces slower weight training process and the training time will be increase at a certain extent.The injection energy of the probe spike increases,the training time will be larger,but the increase of the injection energy of the trigger spike can well shorten the training time of the weight.In addition,for the probe spike and trigger spike,the weight training time will increase with the increase of wavelength detuning.For internal parameters of the system,when the bottom DBR reflectivity(R_b)increases,the weight training time increases slightly,but for the top DBR reflectivity(R_t),the weight training time decreases with the increase of R_t.in addition,the radius of the cavity active region(r)increases will increase the training time.However,with increase of the linear material gain coefficient(a),the fixed internal loss(?_l)and the internal quantum efficiency increases,the weight training time will decrease.