Concept-Based Explanations in Computer Vision: Where Are We and Where Could We Go?

Concept-based explanations offer more intuitive understanding than traditional methods: Unlike saliency-based techniques that only highlight which regions of an image matter, concept-based approaches explain AI decisions using semantically meaningful parts (concepts) that humans can readily understand, making AI systems more transparent to non-experts.

The field is still developing core methodological choices: The paper identifies three critical areas still being worked out—which concepts to use for explanations, how to represent those concepts, and how to control them—suggesting that concept-based explanation methods lack standardized approaches and best practices.

Knowledge representation techniques could improve future explanations: The authors propose drawing inspiration from knowledge representation and learning fields to better control and structure concepts, indicating that integrating insights from other AI disciplines could make explanations more reliable and systematic.

This is an active research frontier with significant gaps: By calling for a "critical look at advances and trends" and proposing "underexplored areas," the paper signals that while concept-based explanations are promising, the field still has substantial room for development before these methods can be reliably deployed in high-stakes applications.