Deep Neural Network System Based On Memristor Crossbars And Its Application

Compared with the traditional neural network method,the deep learning neural network has a significant feature of large amount of computation,so it is urgent to have a higher performance and lower power consumption of the computing system to run the deep neural network.The goal of this thesis is to design a deep neural network circuit based on memristor crossbar,which can greatly improve the speed and energy efficiency.Semantic segmentation and sentiment analysis are realized in the proposed circuit system.In this thesis,a memristor based deep neural network memristor is designed by combining memristor crossbars with full convolutional neural network and long short-term memory neural network,which are widely used in deep learning.Memristor is used as the synapse of neural network to reduce the energy consumption and delay of the whole circuit system.The main memristor based circuit structures for full convolutional neural network and long short-term memory neural network are quite different.In the design process of full convolutional neural network,this thesis utilizes memristor crossbars to realize three important neural network layers including convolution layer,maximum pooling layer and deconvolution layer,and then integrates these network layers into a complete memristor-based neural network system through auxiliary circuit connections.The integrated neural network circuit is applied to image semantic segmentation task.In the design of long short-term memory neural network,the feature extraction layer and external classification layer circuit are realized by memristor crossbars in this thesis.Then,the circuits are connected to form a complete memristor-based long short-term memory neural network circuit with auxiliary circuits.The text is converted to the corresponding word vector and then converted to the corresponding input voltages.The system then outputs the level of sentiment analysis of the text through the memristor-based long short-term memory network circuit to realize the end-to-end sentiment analysis.The research results of this thesis are to realize the memristor-based deep neural network,mainly focusing on the design of memristor-based full convolutional neural network and memristor-based long short-term memory neural network,which can realize the forward propagation in circuit level,as well as end-to-end image segmentation and text sentiment analysis.In this thesis,the memristors store the value of neurons and also participate in the operation processing,which can achieve a significant speed improvement.And the design scheme of the main layers can be extended to many other deep neural network circuit design,which will be helpful to the applications of memristor in artificial intelligence chip in the future.