A Memristor-Based Convolutional Neural Network Implementation Method

Abstract

The invention discloses a method for implementing a convolutional neural network based on a memristor, which includes preparing a convolutional neural network to be trained consisting of a convolution + pooling layer and a fully connected layer, and inputting image information of a training set to obtain a fully The output value of the connection layer; according to the deviation between the output value of the fully connected layer of the convolutional neural network to be trained and the standard information, the convolutional neural network to be trained is trained using the backpropagation function; when the number of training times is reached, it is determined whether the training accuracy is Up to the standard, if not up to the standard, retrain the trained convolutional neural network until the accuracy reaches the standard; obtain the required convolutional neural network and other steps. Its remarkable effects are: reducing the impact on precision, preparation difficulty and time on the hardware; realizing a more advanced multi-layer convolutional neural network; not easily affected by hardware noise, preparation difficulty and time, and being able to exhibit greater adaptability.

Description

technical field [0001] The invention relates to the technical fields of artificial intelligence and material science, in particular, a method for realizing a memristor-based convolutional neural network. Background technique [0002] Memristor (Memristor) is a circuit device (dρ=Mdq) that represents the relationship between magnetic flux and charge. It has the dimension of resistance, and its resistance value is determined by the charge flowing through it. By measuring the resistance value of the memristor, the amount of charge flowing through it can be known, thereby having the function of memorizing charge. The resistance-switching properties of memristors are thought to be similar to synapses in neurons in the brain. Studies have shown that the human brain is a network composed of billions of neurons, through about 10 15 communicate through synaptic connections. Human cognitive abilities are generated by computations performed in this vast network, and many fundamental...

AI Technical Summary

Problems solved by technology

[0004] However, although the above method can realize the neural network on the memristor array and apply it to simple classification, the neural network implemented is relatively basic, and the arrangement of the memristor array is unique.

In this arrangement, the scale of the memristor array increases sharply with the complexity of the application, which not only doubles the difficulty and time of preparation, but also makes the noise caused by hardware crosstalk more serious and affects the accuracy.

Images