Interpretable local migration mutual learning method based on attention map

Abstract

The invention discloses a local mutual learning method based on attention map migration, and aims to improve the annotation precision of a model attention map so as to improve the interpretability of a model. The method comprises the steps: adopting two lightweight models; in the migration process, measuring the distance between attention maps between the models through a Pearson algorithm, and learning the attention maps of the opposite sides mutually along with the training process. In order to avoid the influence of negative migration, region division is carried out on the attention map, a threshold value is set, and the attention map part with high similarity is selected for migration. Compared with the most advanced method, the algorithm provided by the invention has the advantages that the average decline rate of 28.2% and the average growth rate of 29.5% of visual interpretability confidence coefficient are respectively realized, the algorithm is superior to other methods based on input disturbance and class activation mapping to a great extent, and meanwhile, the most responsive region in the sample picture can be labeled, and is not limited to a visual visualization area.

Description

technical field [0001] The invention belongs to the technical field of image recognition, and can be applied to deep learning model interpretation in industrial, financial, medical and other scenarios. Background technique [0002] In convolutional neural models, the interpretability of feature visualization plays an important role in intuitively understanding model decisions. Simonyan et al. proposed a gradient-based approach to this class, which visualizes the features of the class based on the output score of the maximal class in a deep convolutional model. Zeiler et al. propose a deconvolution model that shares weights with the original model and is used to project certain features onto the image plane for visualization of pictures. The method of activation maximization is further optimized by introducing regularization to make the visualization clearer and more interpretable. Shi et al. proposed an upconvolution model to invert CNN feature maps into images, viewing up...

AI Technical Summary

Problems solved by technology

[0006] Purpose of the invention: To solve the problem of poor interpretability of the model due to the deviation of the attention map labeling area constructed by the current model

Images