Research On Convolutional Neural Network Based On Memristor

With the decades years' development,artificial intelligence has been applied in many fields such as autopilot,face recognition and intelligent medical.Convolutional neural network,as a neural network algorithm that has received a lot of attention,shows unique advantages in the field of computer vision.However,with the rapid development of neural network algorithms,the requirements for computer processing capabilities get higher and higher.In traditional computer architectures,computing and storage are separate.As the amount of network computing increases,resource consumption also increases.The existing hardware structure is gradually unable to meet the computing needs of neural networks.The appearance of memristor is expected to solve this problem.The memristor with learning and memory features can be used to store the weight of the neural network,also it can be used to network computing by its analog characteristics,which is very suitable for the hardware implementation of neural network.Therefore,this paper uses a simple linear memristor model combined with CNN algorithm to design a CNN hardware model based on memristor in MATLAB for recognizing handwritten digital images.The CNN model's training method is ex-situ training which can obtain network weights and bias parameters.According to the storage principle of the memristor,a corresponding weight loading module is designed to implement the weight read and write operations on memristor after analyzing the relationship between weights and resistance values.And the module is verified by image reading and writing experiments.The weight and bias parameters of the neural network are converted into the corresponding resistance values according to the conversion formula.And the resistance value conversion of memristor is achieved by the weight loading module.The design of the convolution layer model and the fully connected layer model is completed based on memristor array model.The pooling layer model is realized by using four memristors combined with an integrator module.The integrator is designed to restore the weight information converted into resistance values and process the output of the convolution layer and the pooling layer.The integrator module and the activation function model are used to implement the neuron function of processing infor mation.Based on the principle of the Softmax function,the final output layer is designed to process the network output results,finsh the final output.After the module design of the network model is completed,the overall MATLAB model can be completed.The CNN model based on memristor is designed to have good performance.The power consumption of the memristor array part is about 0.2mw.The speed of recognizing images is fast,and the time required is 1us.The recognition accuracy of randomly extracted test pictures reaches 90% with experimental verification.