VTranM: Vision Transformer Explainability with Vector Transformations Measurement
Primary Area: visualization or interpretation of learned representations
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.
Keywords: Post-hoc Explainability, Vision Transformer, Explainable AI
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: We propose a Vision Transformer explanation method that evaluates the effects of transformed vectors, which faithfully captures comprehensive vector contributions over the entire model.
Abstract: While Vision Transformers, characterized by their growing complexity, excel in various computer vision tasks, the intricacies of their internal dynamics remain largely unexplored. To embed visual information, Vision Transformers draw representations from image patches as transformed vectors and subsequently integrate them using attention weights. However, current explanation methods only focus on attention weights without considering essential information from the corresponding transformed vectors, failing to accurately illustrate the rationales behind models' predictions. To accommodate the contributions of transformed vectors, we propose VTranM, a novel explanation method leveraging our introduced vector transformation measurement. Specifically, our measurement faithfully evaluates transformation effects by considering changes in vector length and directional correlation. Furthermore, we use an aggregation framework to incorporate attention and vector transformation information across layers, thus capturing the comprehensive vector contributions over the entire model. Experiments on segmentation and perturbation tests demonstrate the superiority of VTranM compared to state-of-the-art explanation methods.
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: 2664
Loading