Study On STDP-based Unsupervised Learning Algorithm In A Photonic Spiking Neural Network

In recent years,artificial intelligence technology has continued to develop,and more and more scientists are turning their attention to the direction of machine intelligence.Neural networks have important applications in the fields of machine learning and pattern recognition.The neural network based on hardware implementation is a research hotspot in recent years.We hope that the machine can think and work like the brain,greatly improving the convenience of human life and work.Neuronal cells are usually stimulated by the external stimuli to produce spikes that pass through the synapses between neurons,which is the basis of brain working.Therefore,neurons and synapses play very important roles in the brain's nervous system.Spike Timing Dependent Plasticity(STDP)is the neurobiological basis for nervous system development,learning,and memory.Therefore,in order to build a brain-like intelligent machine,the effective implementation of neurons and synaptic circuits is very critical.However,research on this area is very limited due to the lack of suitable devices to simulate the realization of neurons and synapses.But intelligent machines like the brain need more efficient,smaller devices to achieve its ultimate goal.The photonic neuromorphic system can simulate neuromorphic algorithms at speeds of millions to billions of times faster than biological brains.It is unmatched by other neuromorphic hardware systems and can handle more complex computational tasks than traditional digital or analog optical calculations,such as adaptive control,learning and memory,and sensory information processing.The unique adaptability,fault tolerance and spike signal mechanism of neural morphology calculation can avoid the bottleneck of optical computing development,such as chip integration of traditional digital light calculation and noise accumulation of analog light calculation.The photonic spiking neural network not only has a more similar working mechanism than the brain,but also combines the advantages of low power consumption and high bandwidth of photonic calculation.In conclusion,the photonic neuromorphic system has the advantages of fast speed,low energy consumption,high bandwidth and so on,which has important research significance.In this thesis,the Vertical-Cavity Surface-Emitting Laser with a Saturable Absorber (VCSEL)is studied as a photonic neuron,the Vertical-Cavity Semiconductor Optical Amplifier(VCSOA)as a key device to simulate the photonic synapse with STDP mechanism is proposed for the first time,and we set up a common photonic spiking neural network model and realize the STDP unsupervised learning algorithm based on photonic spiking neural network by using MATLAB simulation software.Postsynaptic neurons successfully identified the first spike that was issued by presynaptic neurons.Among them,the spike encoding based on VCSEL-SA,the STDP based on VCSOA,and the learning characteristics of photonic spiking neural network are studied in detail.These results are very interesting and valuable,and can provide theoretical basis and guidance for the field of photonic information processing of ultrafast photonic neuromorphic systems.