Exploring Compositionality in Vision Transformers using Wavelet Representations
Keywords: Vision Transfomers, Explainability, Compositionality, Latent Representations
TL;DR: Understanding compositionality trends in Vision Transformer models.
Abstract: Insights into the workings of the transformer have been elicited by analyzing its representations when trained and tested on language data. In this paper, we turn an analytical lens to the representations of variants of the Vision Transformers. This work is aimed to gain insights into the geometric structure of the latent spaces of each encoding layer. We use representation-similarity measures, and representation-visualization approaches to analyse the impact of training regimes on the latent manifolds learned. We then use our approach to design a test for quantifying the extent to which these latent manifolds respect the compositional structure of the input space. We restrict our analysis to compositional structure induced by the Discrete Wavelet Transform (DWT). Interestingly, our empirical analysis reveals that ViT patch representations give notions of compositionality with respect to the DWT primitives.
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: 13839
Loading