Conceptualize Any Network: A Concept Extraction Framework for Holistic Interpretability of Image Classifiers

Hossein Moghaddam, Quentin Bouniot, Jayneel Parekh, Pavlo Mozharovskyi, Florence d'Alché-Buc

Keywords: Explainability, Computer Vision, CNN, ViT

TL;DR: Conceptualize Any Network

Abstract:

Attribution-based and concept-based methods dominate the area of post-hoc explainability for vision classifiers. While attribution-based methods highlight crucial regions of the input images to justify model predictions, concept-based methods provide explanations rooted in high-level properties that are generally more understandable for humans. In this work, we introduce ``Conceptualize Any Network'' (CAN), a comprehensive post-hoc explanation framework that combines the wide scope of attribution-based methods and the understandability of concept-based methods.
Designed to be model agnostic, CAN is capable of explaining any network that allows for the extraction of feature attribution maps, expanding its applicability to both CNNs and Vision Transformers (ViTs). Moreover, unlike existing concept-based methods for vision classifiers, CAN extracts a set of concepts shared across all classes, enabling a unified explanation of the model as a whole.
Extensive numerical experiments across different architectures, datasets, and feature attribution methods showcase the capabilities of CAN in Conceptualizing Any Network faithfully, concisely, and consistently.
Furthermore, we managed to scale our framework to all of ImageNet's classes which has not been achieved before.

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: 9925