Decoupling evidence-based deep learning distribution outside detection method, device and medium

By decoupling evidence-based deep learning and utilizing gradient truncation and Dirichlet distribution, the feature confusion problem in out-of-distribution sample identification of deep learning models is solved, achieving accurate differentiation and reliable detection of in-distribution and out-of-distribution samples.

CN122156829AActive Publication Date: 2026-06-05NANJING UNIV OF POSTS & TELECOMM
View PDF 4 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
NANJING UNIV OF POSTS & TELECOMM
Filing Date
2026-05-08
Publication Date
2026-06-05

AI Technical Summary

Technical Problem

Existing deep learning models struggle to accurately identify and provide reliable uncertainty estimates when faced with out-of-distribution samples, potentially leading to serious consequences in safety-critical areas. Furthermore, the coupling of classification and evidence results in feature confusion, making it difficult to distinguish between in-distribution and out-of-distribution samples.

Method used

We employ a deep learning approach based on decoupled evidence, using a gradient truncation strategy to decouple the training of the classification branch and the evidence branch. By combining a normalized cosine classification head and explicit constraints from out-of-distribution samples, we construct a decoupled evidence network model and use the concentration parameter of the Dirichlet distribution for uncertainty assessment.

Benefits of technology

It effectively solves the feature confusion problem caused by the coupling of classification and evidence, improves the model's ability to distinguish between difficult samples within the distribution and samples outside the distribution, and provides reliable uncertainty estimation and rejection decision basis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122156829A_ABST
    Figure CN122156829A_ABST
Patent Text Reader

Abstract

The application discloses a distribution-out detection method and device based on decoupling evidence deep learning and a medium, comprising: constructing a classification network based on evidence deep learning, modeling the output of the neural network as a concentration parameter of Dirichlet distribution; designing a decoupling training objective function, independently constraining the total amount of evidence and the category distribution form of the in-distribution samples and the out-of-distribution samples respectively, optimizing the model parameters by minimizing the difference between the predicted distribution and the target Dirichlet distribution; in the inference stage, calculating the differential entropy of the predicted Dirichlet distribution based on the concentration parameter of the network output, and taking it as an uncertainty measurement index for distribution-out detection. The application solves the problem that the weak feature in-distribution samples and the strong feature out-of-distribution samples are difficult to distinguish in the feature space in the traditional method by decoupling the correlation between the evidence amplitude and the distribution uncertainty, and effectively improves the distribution-out detection performance in the general image classification scene.
Need to check novelty before this filing date? Find Prior Art