A Multi-conductance States Memristor-based CNN Circuit Using Quantization Method for Digital Recognition

Neural network based on memristor is one of the research hotspots in neuromorphic computation. Mostly, the weights of neural network are mapped to the conductance of memristor. During mapping stage, there are some problems, such as the weights in network simulation are infinite. While these are finite conductance states of memristor. In this paper, we propose a weight quantization method of convolutional neural network (CNN) based on memristor, in which the weight of each layer in the convolutional neural network is uniformly quantized to 32 conductance states. This study constitute array by constructing a 32 conductance states memristor model. At the same time, the conductance state of the model corresponds to the weights of the neural network. Furthermore, the array of 32-conductance state memristors are used to build the convolutional neural network circuit. The experimental results show that the CNN performs MNIST image recognition reaches an accuracy of 95.43% with the help of quantization method.