Research On Spiking Neural Network Based On Memristor Synapse And Its Application

Nowadays,with the development of information technologies such as artificial intelligence and brain-like computing,and the construction of infrastructure such as the Internet of Things and communication networks,these emerging industries have brought convenience to people's production and life,and at the same time,they have also produced massive amounts of data for us to process.Therefore,there are higher requirements for the processing speed and the response ability of equipment and networks.As a nano-level component,the memristor not only has the advantages of variable resistance,non-volatility,and synaptic-like characteristics,but also has higher integration,lower power consumption,and faster reading and writing speed,it is expected to greatly improve data processing capabilities and machine learning capabilities in the fields of storage circuits,logic computing,and brain-like computing.It is of great significance for building a new generation of intelligent information processing systems.Spiking neural network is a more reasonable simulation of real biological neural network.It encodes real-world information into spike sequences to realize its functions.It has the advantages of low power consumption,high timeliness,and real-time performance.It has huge application potential in associative memory,pattern recognition,image processing,etc.The combination of the two provides a extremely feasible solution for the construction of the intelligent system that integrates storage and computing and the realization of large-scale data processing.Based on the research results of memristor and spiking neural network,this paper combines the two to explore the realization of the spiking neural network based on the memristor.The details are as follows:(1)Three commonly used memristor models are introduced in detail,namely HP memristor,spintronic memristor,and voltage threshold adaptive memristor.The change mechanism of memristance in each memristor model is described through mathematical expressions.The memristive characteristics are studied through MATLAB simulation experiments,which provide basic support for the following research based on memristor.(2)Based on the HP memristor model,the advantages and disadvantages of its classic window function model are summarized.On this basis,we propose a new general window function,and give the value range of adjustable parameters in the model.Furthermore,simulation experiments are carried out on the memristor model using the new window function,we observe the changes in some important physical quantities such as its memristance,state variable,and verify its effectiveness.And the general window function is compared with the classical window function model to clearly show its Existing advantages.(3)Based on the structural similarity between the memristor and the synapse,the relationship between the change of the memristance and the change of the synapse weight is established through mathematical formulas,and the series-parallel circuit of the memristor is discussed.On this basis,a comprehensive summary of the existing memristive synaptic circuits is carried out by means of circuit analysis and numerical simulation,laying a foundation for the next step of building a neural network.(4)Combining the physiological mechanism of spike neurons,the three widely used neuron models-HH model,LIF model and IZ model are discussed in detail and effective simulations.Furthermore we realize the spiking neural network based on the memristor,including the improvement of the LIF model,the elaboration of the neural network structure and the mathematical derivation of the training algorithm.Finally,the standard handwritten digit recognition data set(MNIST)is used to fully verify the feasibility and effectiveness of the overall scheme.