Concept Bottleneck Models under Label Noise
Keywords: Concept bottleneck models, Label noise, Sharpness-aware minimization
TL;DR: We analyze the impact of label noise on concept bottleneck models, revealing significant performance issues, and propose sharpness-aware minimization to mitigate these challenges.
Abstract: Concept bottleneck models (CBMs) are a class of interpretable neural network models that make the final predictions based on intermediate representations known as concepts. With these concepts being human-interpretable, CBMs enable one to better understand the decisions made by neural networks. Despite this advantage, we find that CBMs face a critical limitation: they require additional labeling efforts for concept annotation, which can easily increase the risk of mislabeling, i.e., CBMs need to be trained with noisy labels. In this work, we systematically investigate the impact of label noise on CBMs, demonstrating that it can significantly compromise both model performance and interpretability. Specifically, we measure the impact of varying levels of label noise across different training schemes, through diverse lenses including extensive numerical evaluations, feature visualizations, and in-depth analysis of individual concepts, identifying key factors contributing to the breakdowns and establishing a better understanding of underlying challenges. To mitigate these issues, we propose leveraging a robust optimization technique called sharpness-aware minimization (SAM). By improving the quality of intermediate concept predictions, SAM enhances both the subsequent concept-level interpretability and final target prediction performance.
Primary Area: interpretability and explainable AI
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 940
Loading