ConceptExplainer: Interactive Explanation for Deep Neural Networks from a Concept Perspective.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Human-Computer Interaction

Background:

• Traditional deep learning interpretability methods lack global and fine-grained explanations.
• Concept-based explanations offer intuitive, flexible insights into model behavior.
• Automated concept discovery presents challenges in navigating the concept space.

Purpose of the Study:

• To design, develop, and validate ConceptExplainer, a visual analytics system.
• To enable interactive exploration of the concept space for deep learning model interpretation.
• To address challenges faced by users in understanding deep learning model behavior.

Main Methods:

• Iterative prototyping to develop the ConceptExplainer system.
• User study to validate the system's effectiveness in addressing interpretation challenges.
• Usage scenarios to demonstrate interactive analysis of model behavior.

Main Results:

• ConceptExplainer facilitates structured navigation and exploration of the concept space.
• The system provides concept-based insights at instance, class, and global levels.
• Validated through user studies and diverse usage scenarios.

Conclusions:

• ConceptExplainer effectively supports users in interpreting deep learning models.
• The system aids in identifying important concepts for classification and detecting data bias.
• Enables understanding of concept sharing across different classes.