New research rethinks concept bottleneck models for better interpretability

Two new research papers explore the interpretability of Concept Bottleneck Models (CBMs), which aim to make deep learning models more transparent by factoring predictions through human-understandable concepts. The first paper introduces 'Clarity,' a diagnostic measure to assess the trade-off between downstream performance and the semantic alignment of concept activations, finding that models can optimize task performance by deviating from semantic alignment. The second paper proposes 'Representation Integrity' as a crucial property for CBMs, introducing metrics like group coherence and concept coverage to evaluate how well concept-supporting features are organized, suggesting that concept integrity is a vital criterion beyond simple accuracy. AI

IMPACT These papers introduce new frameworks for evaluating and improving the interpretability of concept bottleneck models, potentially leading to more trustworthy AI systems.

RANK_REASON Two academic papers published on arXiv introducing new methods and metrics for evaluating concept bottleneck models.

Read on arXiv cs.LG →

• arXiv
• CatalyzeX
• Concept Bottleneck Models
• DagsHub
• Gaoxiang Huang
• Gotit.pub
• Hugging Face
• Konstantinos P. Panousis
• Representation Integrity
• ScienceCast

• paper
• safety

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →