Guardado en:
Detalles Bibliográficos
Autores principales: Jeanneret, Guillaume, Simon, Loïc, Jurie, Frédéric
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:https://arxiv.org/abs/2502.17196
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866910917687508992
author Jeanneret, Guillaume
Simon, Loïc
Jurie, Frédéric
author_facet Jeanneret, Guillaume
Simon, Loïc
Jurie, Frédéric
contents Visual transformers have achieved remarkable performance in image classification tasks, but this performance gain has come at the cost of interpretability. One of the main obstacles to the interpretation of transformers is the self-attention mechanism, which mixes visual information across the whole image in a complex way. In this paper, we propose Hindered Transformer (HiT), a novel interpretable by design architecture inspired by visual transformers. Our proposed architecture rethinks the design of transformers to better disentangle patch influences at the classification stage. Ultimately, HiT can be interpreted as a linear combination of patch-level information. We show that the advantages of our approach in terms of explicability come with a reasonable trade-off in performance, making it an attractive alternative for applications where interpretability is paramount.
format Preprint
id arxiv_https___arxiv_org_abs_2502_17196
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Disentangling Visual Transformers: Patch-level Interpretability for Image Classification
Jeanneret, Guillaume
Simon, Loïc
Jurie, Frédéric
Computer Vision and Pattern Recognition
Artificial Intelligence
Visual transformers have achieved remarkable performance in image classification tasks, but this performance gain has come at the cost of interpretability. One of the main obstacles to the interpretation of transformers is the self-attention mechanism, which mixes visual information across the whole image in a complex way. In this paper, we propose Hindered Transformer (HiT), a novel interpretable by design architecture inspired by visual transformers. Our proposed architecture rethinks the design of transformers to better disentangle patch influences at the classification stage. Ultimately, HiT can be interpreted as a linear combination of patch-level information. We show that the advantages of our approach in terms of explicability come with a reasonable trade-off in performance, making it an attractive alternative for applications where interpretability is paramount.
title Disentangling Visual Transformers: Patch-level Interpretability for Image Classification
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2502.17196