Memristor-Based Affective Associative Memory And Hopfield Neural Network And Its Application In Image Segmentation

Artificial neural networks are dedicated to simulate the information processing mode of biological nervous system to solve complex practical problems.With the development of artificial intelligence,artificial neural network has achieved breakthroughs in many products and fields such as visual computing,intelligent robots,machine learning,intelligent transportation,etc.At present,neural networks are mostly implemented in software,leading to the inherent parallelism of neural networks reduced.However,it is well known that circuit operations can provide parallel high-speed computation for the implementation of neural networks.Therefore,it makes sense to study the hardware implementation of neural networks.Traditional neural network circuits mostly choose resistors or MOS devices to build electronic synapses.However,it is difficult to modify synaptic weights once these weights are determined,which limits the development of neural network hardware implementation.As a new type of two-terminal element,memristor possesses non-volatile and flexible resistance characteristics,making it a candidate device for imitating the function of biological synapses.In addition,memristor possesses dual functions of storage and computation,which makes it an ideal choice for building circuit systems that integrating storage and computing.It is also expected to break the bottleneck of separation of processor and memory in the traditional von-Neumann systems.Based on the idea of constructing a neural network through memristor and related hardware,this thesis proposes a memristive affective associative memory neural network circuit and a memristive competing Hopfield neural network circuit,and the following research results have been obtained.(1)A memristor-based affective associative memory neural network circuit is proposed,which includes the gradual learning,gradual forgetting and gradual transferring stages with variable emotional intensity.In the designed circuit,the memristors are utilized to define the synaptic weights.When the memristance decreases,the corresponding synaptic weight will increase and the synapse strength will be stronger.Making use of the leaky integrate-and-fire neuron model,the firing frequency of the output neurons is variable.By correlating the emotional intensity with the firing frequency of output neurons,the intensity of emotions can gradually change from strong to weak or from weak to strong.Finally,the Pspice simulation of the memristor-based affective associative memory neural network circuit is finished,and the simulation results verify the feasibility of this circuit.(2)A memristive competitive Hopfield neural network circuit is proposed to deal with the image segmentation problem.First,in order to realize the competition mechanism,the competition neuron and the competition control module are designed,whose function is to select an output neuron from the competitive layer as the winning node.Following,the parallel coding scheme is utilized to adjust the memristance of crossbar array,so as to solve the problem of weights storage in this neural network circuit.Then,a competitive Hopfield neural network circuit is proposed,which can complete the iteration of network and update of data.Finally,two image segmentation experiments are designed to verify the effectiveness of the proposed circuit.