Trusted multi-view classification method based on evidence deep learning

Abstract

The invention discloses a credible multi-view classification method based on evidence deep learning, and the method comprises the following steps: S1, carrying out the sample definition, setting a data set to have N samples, and enabling each sample to have V views; s2, estimating the classification uncertainty of the single-view-angle data according to the single-view-angle evidence; s3, multi-view evidence fusion: spreading global information to each view by using a degradation layer, so that each view can learn evidence based on the global information; and S4, optimizing the target, and optimizing all parameters in the model by using a gradient descent algorithm. According to the method provided by the invention, the prediction accuracy is improved, and the complementary information between the visual angles, which is deep and easy to ignore, is mined by using the degeneration layer designed in the invention, so that the uncertainty which better conforms to human cognition is output during prediction.

Description

technical field [0001] The invention relates to the technical field of deep learning, in particular to a credible multi-view classification method based on evidence deep learning. Background technique [0002] Multi-view classification means that each sample in the data set contains features from multiple different perspectives. For example, doctors often need to comprehensively consider features from multiple perspectives such as MRI and clinical laboratory results when diagnosing cancer. The use of deep learning to achieve multi-view classification tasks can make up for the shortcomings of traditional machine learning that are difficult to mine deep information from multi-view data. Most of the past researches are devoted to how to improve the prediction accuracy, but ignore the reliability of decision-making. However, in many high-risk applications, not only the model needs to obtain the result of the decision, but also the confidence level of the decision. For example,...

AI Technical Summary

Problems solved by technology

However, in fact, previous studies can only output high uncertainty when dealing with perspective 2 because they cannot distinguish whether the lesion is on the left or the right, that is, the complete category information, so they cannot dig out the deep complementary information of the two perspectives.

Ignoring information from either perspective in medical diagnosis increases the likelihood of misdiagnosis

[0004] In multi-view classification problems, it is still a great challenge to mine the complementary information between views and obtain the uncertainty in line with human cognition without reducing the accuracy.

Images