Stochastic-Memristor-Based Memristive Neural Networks And Its Application In Image Classification

Today,a variety of artificial intelligence applications are emerging.Among them,artificial neural network(ANN)is one of the most effective methods.The training of the artificial neural network is compute-intensive.It is difficult to meet the low power consumption and small size scenarios when ANN is implemented by software.But the conditions can be satisfied by using memristors.Memristors are very suitable for implementing a large number of matrix-vector multiplication in artificial neural networks due to their nanoscale,non-volatility and low power consumption.However,the influence of the stochastic memristor on artificial neural networks is unknown.The influence of stochasticity of memductance adjustment on neural networks' performance is studied in this thesis.Firstly,the causes of memristive stochasticity are analyzed,and a suitable stochastic memristor model is established.Then the circuit structure of a multi-layer neural network based on two transistors and one memristor cells is introduced.Through the simulations,it is found that the multi-layer neural network algorithm based on stochastic memristors has higher classification accuracy on the Iris dataset when the dataset contains no noise or the dataset noise is low.In addition,compared with the deterministic memristors,when the noise degree of the dataset is low,the multi-layer neural network based on stochastic memristors usually gets higher classification accuracy when the degree of stochasticity is low.Finally,the circuit structure of a convolutional neural network based on stochastic memristor is introduced.The impact of stochasticity of memristors on the performance of convolutional neural networks is analyzed on the CIFAR-10 dataset.Four situations are considered: the original image,the original image with multiplicative noise,the original image with Gaussian noise and the contaminated image.It is found that convolutional neural networks have better classification performance when the memristive stochastic degree is low.The research results of this thesis mainly include the establishment a mathematical model of the stochastic memristor,the simulation of the effect of stochastic memristors on the performance of multi-layer neural networks,the circuit implementation of memristor-based convolutional layer and pooling layer,the simulation of stochastic memristors' influence on the classification performance of noisy image.The results of this thesis are of great significance for the implementation of memristor-based artificial neural networks.