Numerical And Experimental Investigation Of Neuron-like Information Processing For Photonic Neuromorphic Systems

In recent years,with the advancement of science and technology and the gradual increase in data exchanges,the data such as images and voices that computers need to process has increased rapidly.However,computers developed based on the Moore's Law have encountered the problems of"memory wall"and"power bottleneck",and gradually cannot meet the huge data processing needs.In contrast,brain-like computing has shown great advantages such as strong processing performance,simultaneous processing of multiple signals,and low power consumption.According to research,the adult brain can perform 10¹⁶ operations per second on average,and the energy required to perform these operations is only about 20 watts.Although the information processing technology based on microelectronics spiking neural network has made great achievements,it has encountered bottlenecks in energy consumption and speed.The optical platform has great advantages in the field of information processing due to its unique advantages such as fast speed,large bandwidth,and low power consumption.Therefore,neural-like information processing based on the photonic spiking neural network has gradually become a hot spot in recent years,but the current research is still in the early stage of exploration.This thesis aims at international frontier hot-spots and carries out research on neural information processing of photonic spiking neural network with the support of the National Natural Science Foundation of China.Neuronal-like properties have been investigated numerically and experimentally based on vertical-cavity surface-emitting lasers(VCSELs)and VCSELs with an embedded saturable absorber(VCSEL-SA).Besides,synaptic-like properties have been investigated numerically and experimentally based on the vertical-cavity semiconductor optical amplifier(VCSOA).In addition,the information processing has been investigated numerically and experimentally based on the small photonic spiking neural network.Moreover,the photonic spiking neural network algorithm has been investigated.The research results are of great significance to the development of photonic spiking neural networks.The research content and innovations of the thesis are as follows:1.The photonic neuron is one of the important units of the photonic spiking neural network.In investigating of the properties of the photonic neuron is the foundation of the investigation of information processing in photonic neural network.For the excitability of photonic neurons,a scheme based on VCSEL-SA with external optical injection was proposed,which realized single-channel and dual-channel excitatory response.The effects of intensity and duration time of external optical stimulation and the parameters of the laser on excitatory properties were analyzed.For the dendritic stimulation response of pyramidal neurons,the schemes for realizing the dendritic stimulation response of pyramidal neurons based on injection-locked state,spiking state,and relaxation oscillation state of VCSEL were proposed.The effects of external optical injection and laser parameters on dendritic stimulation response of pyramidal neurons were analyzed.The investigation of the properties of neurons based on the VCSEL is the foundation of the information processing in the photonic spiking neural network.2.Since there is no negative pulse in the optical domain,it is difficult to achieve neuronal-like inhibitory response.In this thesis,the inhibitory response based on the VCSEL-SA polarization mode competition mechanism was discovered for the first time,which solved the problem of achieving inhibitory response in the optical domain.The effects of external optical injection and the parameters of laser on inhibitory properties were analyzed carefully.Based on the inhibition,a scheme of achieving the winner-take-all(WTA)mechanism was further proposed.The effects of the parameters of the laser on the WTA mechanism were analyzed to provide theoretical support for the competitive learning mechanism.In addition,schemes of pattern recognition and max-pooling operation were proposed based on the WTA mechanism.The realization of pattern recognition promotes the information processing in the all-optical spiking neural network;the realization of the max-pooling operation promotes the development of the all-optical convolutional neural network.3.Because of the lack of commercial devices with synaptic-like properties,a scheme based on self-feedback VCSOA to achieve synaptic-like properties was proposed.Based on the numerical and experimental investigation,a non-volatile synaptic response was achieved.The effects of external optical injection and laser parameters on synaptic-like properties were analyzed.In addition,based on the proposed synaptic-like properties scheme,the self-learning of the photonic spiking neural network was achieved,and the firing timing of first spike in the pattern was recognized,which promotes the development of online learning in all-optical spiking neural network.4.Because of the bottleneck of energy consumption and speed limit in microelectronics,the schemes for many types of information processing in photonic spiking neural networks were proposed,including the achieving schemes of information storage,binary convolution,exclusive OR operation,etc.,which promote the development of all-optical spiking neural networks.For information storage,the effects of key parameters in the mutual coupling system on information storage were analyzed;For all-optical binary convolution,an implementation scheme based on a single commercial VCSEL was proposed,which is simple and energy-saving;on this all-optical binary convolution scheme,the image gradient value was calculated;For the exclusive OR operation,a solution based on the dendrites stimulus pyramidal neuron response was proposed based on a single commercial VCSEL.The system has advantages,such as simple,energy-saving,and fast.5.The algorithm research based on the photonic spiking neural network included a modified supervised learning algorithm,and the collaborative application of the algorithm and hardware.Because of the time information is difficult to extract in the photonic spiking neural network,a modified supervised learning algorithm was improved.The proposed supervised learning algorithm does not depend on the time information of the photonic spiking neural network.In addition,based on this algorithm,Arabic numeral recognition was realized in the photonic spiking neural network.In addition,this thesis investigated the application of Tempotron learning algorithm in the photonic spiking neural network,and achieving the spiking train recognition,which promotes the coordinated development of the hardware,algorithm,and the all-optical spiking neural network.