Research And Design Of Convolutional Neural Network Based On Hybrid CMOS/Memristor Circuits

The neural network has been developing quickly for recently decades,and its unique advantages in intelligent medical,signal processing,automatic control and many other research fields have been discovered.The most prominent progress is in the field of deep learning.As the most successful network model,convolutional neural network(CNN)could extract features,higher order,abstract and closer to nature,from simple preprocessing data.But the most of the intelligent device can't provide the amount of calculation of more and more complex network model.Because memristor has excellent performance in the field of brain-like computing,memristor caused the wide attention of researchers.When a memristor is connected to a current source,the state of the memristor will change according to the amount of charge.As a result,memristor has the ability of learning and memory.In addition it has small size,low power,analog storage,non-volatile and other characteristics,so memristor is very suitable for hardware implementation of neural network system.In this thesis,we design an improved memristor crossbar array(MCA)to realize a CNN using memristors and CMOS devices.The MCA can store weights and bias accurately.A dot product between two vectors can be calculated after introducing an appropriate encoding scheme.The improved MCA is employed for convolution operations,and a classifier in a CNN.Then 'we also design a memristive CNN architecture using the improved MCA and based on the high fault-tolerance of CNNs to perform a basic CNN algorithm.In the designed architecture,the analog results of convolution operations are sampled and held before a pooling operation rather than using analog digital converters and digital analog converters between convolution and pooling operations in a previous architecture.Experimental results show the designed circuit with the area of 0.8525 cm2 can achieve a speedup of 1770× compared to a GPU platform.Compared with previous memristor-based architecture with a similar area,our design is 7.7× faster.There are many intermediate results will be produced in the computational process of CNN and need to be stored.In this thesis,we also design an improved MCA,called VMCA.VMCA is used to store feature map and implement vectorization of local receptive field.Based on VMCA,CNN circuit saves a lot of digital to analog converters and analog to digital converters,and uses very little RAM.Experimental results show that VMCA structure can be effectively applied to the CNN circuits.And the CNN circuit based on VMCA is 7x faster than previous design,but with few loss of precision.