Concept Bottleneck Models: Interpretability vs. Robustness Trade-off Explored

A new research paper explores the robustness of Concept Bottleneck Models (CBMs), which are designed for interpretability. The study argues that previous findings on CBM robustness have been contradictory due to a conflation of different robustness notions and perturbation types. Researchers introduced a framework to compare CBMs with standard classifiers under geometric and semantic perturbations, finding that interpretability does not inherently guarantee robustness. Instead, concept bottlenecks redistribute sensitivity, revealing a trade-off between interpretability and robustness that is dependent on the task structure and perturbation regime. AI

IMPACT Clarifies the relationship between model interpretability and robustness, suggesting that interpretability does not automatically lead to better robustness.

RANK_REASON Research paper published on arXiv detailing a new framework for evaluating the robustness of Concept Bottleneck Models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• Concept Bottleneck Models
• Connected Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• Hugging Face
• IArxiv Recommender
• Influence Flower
• Litmaps
• machine learning
• ScienceCast
• scite Smart Citations

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →