Reproduce “On the Concept Trustworthiness in Concept Bottleneck Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qihan Huang, Jie Song, Jingwen Hu, Haofei Zhang, Yong Wang, Mingli Song
A new framework, ProtoCBM, addresses concept untrustworthiness in Concept Bottleneck Models by redesigning their architecture to ensure concepts are predicted from relevant image regions, achieving an average 40.2 percentage point improvement in trustworthiness scores while maintaining classification accuracy.
Key insights
- Concept untrustworthiness is a quantifiable problem in CBMs, where predicted concepts often do not align with the actual relevant image regions.
- Interpretability can be fundamentally built into a model's architecture through design choices like localized concept prediction and part-prototypes, rather than relying solely on post-hoc explanations.
- Achieving higher concept trustworthiness in CBMs does not necessitate a sacrifice in classification performance, demonstrating that both can be pursued simultaneously.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.