MIMOSA: Multimodal Concept-based representations

Eleonora Poeta, Gabriele Ciravegna, Eliana Pastor, Tania Cerquitelli, Elena Baralis

Keywords: Concept-based model, Multimodal, Explainability

Abstract:

In recent years, deep learning-based architectures have significantly improved multimodal representation. However, interpretability remains challenging with traditional attention and gradient-based methods, offering limited insights into decision-making processes. Concept-based explainability provides intrinsic model interpretability by mapping raw data to higher-level abstractions, yet it has only been applied to unimodal data.
We present MIMOSA (MultIMOdal concept-based repreSentAtion), a unified multimodal model that integrates concept-based interpretability. Our research shows that exploiting a joint multimodal conceptual representation achieves comparable accuracy with multimodal black-box models, surpassing approaches based on unimodal concepts. This unified representation also prevents misclassification of concepts between modalities and improves concept interventions. Through a concept decoder, MIMOSA can extract concept visualizations for each modality.
Experimental results obtained from three distinct multimodal datasets substantiate the efficacy of our approach, showcasing enhanced interpretability in multimodal models.

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: 11143