Research On Memristor Based Neural Network

Memristor is the fourth circuit element in addition to resistor,inductor,capacitor,the essence of it is a non-linear resistance,the resistance value depends on the magnitudes of the current flowing through,with a memory function to maintain resistance after power-down.Neural network is an important research topic in modern intelligent control and information processing.However,the traditional neuron circuit is not able to accurately adjust and store the neuronal synapse's weight,which makes the design of neuron's weight extremely limited,as a result,its development has been greatly limited.The emergence of memristors has provided the possibility for artificial neural networks to simulate neuronal synapses on the circuit.First,memory characteristics of memristors efficiently mimic neuronal synaptic plasticity.Second,memristor can relize a hign density distribution of human brain due to its nanometer-scale size and low power consumption,and the characteristics of memristor STDP(Plasticity based on pulse dependence),and a variety of forgetting characteristics make artificial neural network possessing more human brain function.Therefore,the memristor-based neural network research will reach a new height in the field of artificial intelligence.According to the basic mathematical models and extended models of memristor,this paper analyzes the basic characteristics of memristor,studies the STDP characteristics of memristor,focuses on the simulation of biological synapses,and the establishment and application of memristive neural network,The research contains the following aspects:(1)From the perspectives of ideal memristor and generalized memristor,the theoretical basis of memristor is elaborated.The difference between charge-controlled memristor and magneto-memuistor is analyzed,and provides the specific mathematical expressions;This paper introduces the Ti O2 model of HP lab,and introduces the concept of memory window function and the forgotten Schottky model for comparison;in order to verify the theoretical characteristics of memristor,building the HP memristor model in environment of the SPICE,and the experiment and performance analysis is simulated under sinusoidal excitation.(2)Based on the basic theory of neural network and STDP characteristics of neurons,STDP characteristics of memristor were verified by experimental simulation.Based on the improved magnetron model of memristor,the memristor model was transformed and derived by using the theory of calculus and the basic circuit theory,and the relationship between the conductance change of memristor and the adjustment of biological synaptic weight was obtained.The simulation of synaptic weight adjustment provides a foundation for the construction of memristive neural network.The key point of this study is the realization of the memristor neural network based on synaptic weight adjustment.(3)Compared with the common image encryption algorithms and ideas,this paper proposes an image encryption neural network model based on the memristive synaptic and the image encryption experiments.The security of image encryption is analyzed from five aspects: histogram,correlation,image information entropy,difference key and key sensitivity.(4)Based on memristor mathematic model and image recognition algorithm,a model of memristor neural network for image recognition is designed.By using the change of conductance of memristor to represent the difference of pixel gray value of image,the coding function is introduced.The concept of weight update and aggregation function,the design of the three-layer memristive neural network with input transformation layer,memristor learning recognition layer and decision output layer was carried out.The simulation experiment is carried out in MATLAB to verify whether the edge extraction results have continuity and the accuracy of recognition.