Research On Memristor-Based Elman Neural Network Circuit

With the development of artificial intelligence and in-depth research on biological neural networks,artificial neural networks have played an important role in more and more fields,to a great extent liberated the human brain and hands.Elman neural network has been widely used because of its unique network structure.However,at present,the realization of neural networks is mainly realized by software,which can not give full play to the advantages of large-scale parallel processing of Elman neural networks,which is not conducive to the extensive use of Elman neural networks in the actual industrial production process.Furthermore,in the process of hardware implementation Elman neural network,traditional basic circuit components also have shortcomings in such aspects as long-term storage and accurate adjustment of neural network circuit weight.Therefore,based on the study of structure principle of Elman neural network and the characteristics of memristor,an Elman neural network circuit based on memristor is proposed by combining the weight regulation of Elman neural network with the resistance variation characteristics of memristor.According to the different structure of the memristor module,the neural synaptic circuit in reverse series and the universal valued memristor synaptic circuit are designed.A neural synaptic circuit in reverse series can only represent the weight change of the neural network in a single direction and can not meet the needs of practical application.The universal valued memristor synaptic circuit can not only realize the weight adjustment change in the whole real number range,but also realize the weighted function of the input and weight of the neural network.Therefore,the universal valued memristor synaptic circuit is finally selected to realize the Elman neural network circuit.The neural network circuit is mainly composed of weight array module circuit,feedback layer decision module circuit,transfer function module circuit and V-I conversion module circuit.Among them,the weight array module circuit includes the hidden layer,the feedback layer and the output layer weight array module circuit,whose circuit structure is composed by the universal valued type memristor synaptic circuit,which not only serves as the input port of the whole neural network circuit,but also can adjust the weight direction and size in the neural network circuit.The feedback layer module circuit is mainly used to judge whether the feedback layer weight module circuit is in the working state in the neural network circuit,and according to the judgment result to determine the feedback layer decision module on and off.The transfer function module circuit is used to process some nonlinear input information and convert the nonlinear to the linear so that the output of the neural network has convergence.The V-I conversion module is mainly to convert the voltage signal output from the hidden layer transfer function into the current signal that can be recognized by the weight array module to make the circuit work normally.The performance of Elman neural network circuit based on memristor has digital image recognition experiment.The test results show that the neural network circuit can not only recognize digital image,but also have certain anti-interference performance.To simulate function in the experiment,using the uniform design method to choose the 6 set of test sample data,the experimental results show that the output voltage value of Elman neural network circuit based on memristor is compared with the expected value,and its relative error range is between 0%?13%,and the overall average relative error is only 4.5%.The circuit realizes the function of Elman neural network as a whole.