Loading [a11y]/accessibility-menu.js
Generating Explanations for Model Incorrect Decisions via Hierarchical Optimization of Conceptual Sensitivity | IEEE Conference Publication | IEEE Xplore

Generating Explanations for Model Incorrect Decisions via Hierarchical Optimization of Conceptual Sensitivity

Publisher: IEEE

Abstract:

The capacity to analyze the causes of poor decisions made by visual recognition models is becoming increasingly crucial as the security requirements of various real-world...View more

Abstract:

The capacity to analyze the causes of poor decisions made by visual recognition models is becoming increasingly crucial as the security requirements of various real-world systems continue to escalate. However, the complex structure and blackbox nature within deep neural networks constrain the mining of their error causes. Based on this, we propose a concept-based (e.g. a group of pixel blocks that contain leaves represents the concept of leaves) automated strong localization interpretation framework, called hierarchically optimized concept-sensitive interpretation (HOCS), to provide quantitative analysis of the semantics of wrong decisions in the classification network is provided from two directions of internal and external information interference of samples. HOCS was applied to models with spurious correlation and well-distributed data in the training set. The results showed that it provided concrete explanations in a way that was understandable to humans and demonstrated the significant advantages of HOCS in terms of efficiency and accuracy.
Date of Conference: 30 June 2024 - 05 July 2024
Date Added to IEEE Xplore: 09 September 2024
ISBN Information:

ISSN Information:

Publisher: IEEE
Conference Location: Yokohama, Japan

References

References is not available for this document.