Memristor Electronic Synapse-based Neural Network And Its Application

Since the existence of memristor was proposed by Professor Chua,the interests in this nanoscale device have been increasing gradually in the last few decades.As an electronic device with many excellent characteristics,such as the nanometer geometries,nonvolatility,storage capacity and so forth,memristor has made important contributions to the development of the modern scientific.In particular,with the rapid development of neural network,multiple memristor combination circuits are commonly used to simulate the synapses in different neural networks.In recent years,some relevant work has been carried out by the researchers from industry and academe.However,there are still some room for improvement in simulating synapses in terms of the functions and characteristics.Therefore,in order to simulate the synapses more effectively,it is very important to design an improved synaptic circuit for the development of neural network.At the same time,based on the exploration of neural network,extensive neural network based applications have been conducted in real life,which leads to the fact that the further research on neural network becomes more valuable.In this dissertation,two kinds of classical memristor models are analyzed and compared in detail.Then,multiple memristor combination circuits are also investigated.Based on the analysis of memristor models and their combination circuits,the existing memristor-based synaptic circuits are summarized briefly.Furthermore,a new synaptic circuit and a simple neural network are designed.The application of neural network in real life is realized by corresponding learning algorithm to train neural network.The specific contents are as follows:(1)Two kinds of memristor models are introduced,namely,the Hewlett-Packard(HP)memristor and the spintronic memristor.The calculation methods of memristances about two memristors are derived,and their characteristics are simulated by MATLAB.Furthermore,the two models are analyzed and compared.Finally,composite characteristics of multiple memristor series and parallel circuits are also analyzed.(2)The existing memristor-based synapse circuits,that is,a single memristor based synapse,dual-memristor based synapse,and memristor bridge synapse are analyzed and summarized.The structures of several synaptic circuits are given,their functions are analyzed,and the correctness of the theoretical analysis is verified by simulation experiments.By comparing the number of memristors,weight,area and power consumption in several synaptic circuits,the advantage and disadvantage of each synapse is summarized,and the importance of further research on synaptic circuits is explained.(3)Based on the analysis of the existing memristor-based synaptic circuits,a new type memristor-based synaptic circuit is designed.The corresponding circuit structure is constructed,and the realization methods of weight processing and weight programming in simulating synapse are analyzed and discussed.Based on the synaptic circuit,the corresponding neuron structure is designed and the working process of the neuron is analyzed.Further,a simple neural network is constructed.At the same time,an improved learning algorithm is proposed to train the neural network,and the trained neural network is used to predict the data in real life.The experimental results show the effectiveness of this method.(4)A memristor-based synaptic circuit based on 4T2M(four transistors two memristors)is designed and analyzed,the hardware structure of multilayer neural network is further constructed,and a chip-in-the-loop learning method is introduced.The learning method is applied to train the constructed neural network,and the reconstruction of image super resolution is further realized.