CBM-zero: Concept Bottleneck Model With Zero Performance Loss
Keywords: interpretability, explainability, concept bottleneck model
Abstract: Interpreting machine learning models with high-level, human-understandable \emph{concepts} has gained increasing interest. The concept bottleneck model (CBM) is a popular approach to providing interpretable models, relying on first predicting the presence of concepts in a given input, and then using these concept scores to predict a label of interest. Yet, CBMs suffer from lower accuracy compared with standard black-box models, as they use a surrogate (and thus, interpretable) predictor in lieu of the original model. In this work, we propose an approach that allows us to find a CBM in any standard black-box model via an invertible mapping from its latent space to an interpretable concept space. This method preserves the original black-box model's prediction and thus has zero performance drop while providing human-understandable explanations. We evaluate the accuracy and interpretability of our method across various benchmarks, demonstrating state-of-the-art explainability metrics while enjoying superior accuracy.
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.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 11475
Loading