Prototype-Grounded Concept Models for Verifiable Concept Alignment
Stefano Colamonaco, David Debot, Pietro Barbiero, Giuseppe Marra
TLDR
PGCMs enhance CBM interpretability by grounding concepts in visual prototypes, allowing direct inspection and correction of concept misalignments.
Key contributions
- Introduces Prototype-Grounded Concept Models (PGCMs) for verifiable concept alignment.
- Grounds human-understandable concepts in explicit visual prototypes (image parts).
- Enables direct inspection of concept semantics and targeted human intervention.
- Matches CBM performance while boosting transparency, interpretability, and intervenability.
Why it matters
This paper addresses a key limitation in interpretable AI by making concept alignment verifiable. By grounding concepts in visual prototypes, PGCMs offer a more trustworthy and transparent approach to deep learning. This allows users to directly inspect and correct concept meanings, fostering greater confidence in AI systems.
Original Abstract
Concept Bottleneck Models (CBMs) aim to improve interpretability in Deep Learning by structuring predictions through human-understandable concepts, but they provide no way to verify whether learned concepts align with the human's intended meaning, hurting interpretability. We introduce Prototype-Grounded Concept Models (PGCMs), which ground concepts in learned visual prototypes: image parts that serve as explicit evidence for the concepts. This grounding enables direct inspection of concept semantics and supports targeted human intervention at the prototype level to correct misalignments. Empirically, PGCMs match the predictive performance of state-of-the-art CBMs while substantially improving transparency, interpretability, and intervenability.
📬 Weekly AI Paper Digest
Get the top 10 AI/ML arXiv papers from the week — summarized, scored, and delivered to your inbox every Monday.