Image feature selection method and module based on memristor, and neural network model

Abstract

The invention discloses an image feature selection method and module based on a memristor and a neural network model, and the method comprises the steps: obtaining a data set containing a plurality of images, each image containing M features, and the feature value of each feature being 0 or 1; counting the number Ni of the images with the eigenvalue of the i-th feature of the images in the data set being 1, and setting an image number threshold value Nth; a memristor array containing one row * M columns is constructed, each memristor has the same pulse number threshold Pth, the initial state of the memristor is volatile, and the memristor is converted into non-volatile after receiving pulses larger than or equal to the pulse number threshold Pth; and Pi pulses are applied to the memristors Ri in the ith column to modulate the conductivity states of the corresponding memristors Ri, and Pi = (Pth / Nth) * Ni. According to the invention, the conversion from the volatility of the memristor to the non-volatility of the memristor is fully utilized, the common features of the data set can be screened out, and the network learning complexity is reduced, so that high-density integration of feature processing and network learning hardware circuits is realized.

Description

Technical field [0001] The present invention belongs to the technical field of data processing, and more particularly to an image feature selection method and module, a neural network model based on an empty image feature selection method. Background technique [0002] In the era of artificial intelligence, the growth of data explosion makes high-dimensional data more and more common. For algorithms, there are some independent characteristics and noise characteristics in a large number of data. In the neural network, the character takes the complexity of network learning, and the noise can change the original characteristic value of the characteristics, resulting in different levels of different networks. Feature Selection As the pretreatment process before the training of neural network training, it is possible to effectively select important features in the case of ensuring neural network training accuracy, and filter out unrelated redundant features, reduce the complexity of n...

AI Technical Summary

Problems solved by technology

Some feature selection algorithms have been reported for data dimensionality reduction, but most of these methods are implemented by software, or use memory, processors, etc. to achieve feature selection. The circuit is relatively complex, and the energy consumption and area consumption are relatively large.

Images