This Looks Distinctly Like That: Grounding Interpretable Recognition in Stiefel Geometry against Neural Collapse

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Jia, Junhao, Wang, Jiaqi, Liu, Yunyou, Jing, Haodong, Wu, Yueyi, Wu, Xian, Zheng, Yefeng
Formato: Preprint
Publicado: 2026
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866912955550924800
author Jia, Junhao
Wang, Jiaqi
Liu, Yunyou
Jing, Haodong
Wu, Yueyi
Wu, Xian
Zheng, Yefeng
author_facet Jia, Junhao
Wang, Jiaqi
Liu, Yunyou
Jing, Haodong
Wu, Yueyi
Wu, Xian
Zheng, Yefeng
contents Prototype networks provide an intrinsic case based explanation mechanism, but their interpretability is often undermined by prototype collapse, where multiple prototypes degenerate to highly redundant evidence. We attribute this failure mode to the terminal dynamics of Neural Collapse, where cross entropy optimization suppresses intra class variance and drives class conditional features toward a low dimensional limit. To mitigate this, we propose Adaptive Manifold Prototypes (AMP), a framework that leverages Riemannian optimization on the Stiefel manifold to represent class prototypes as orthonormal bases and make rank one prototype collapse infeasible by construction. AMP further learns class specific effective rank via a proximal gradient update on a nonnegative capacity vector, and introduces spatial regularizers that reduce rotational ambiguity and encourage localized, non overlapping part evidence. Extensive experiments on fine-grained benchmarks demonstrate that AMP achieves state-of-the-art classification accuracy while significantly improving causal faithfulness over prior interpretable models.
format Preprint
id arxiv_https___arxiv_org_abs_2603_08374
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle This Looks Distinctly Like That: Grounding Interpretable Recognition in Stiefel Geometry against Neural Collapse
Jia, Junhao
Wang, Jiaqi
Liu, Yunyou
Jing, Haodong
Wu, Yueyi
Wu, Xian
Zheng, Yefeng
Computer Vision and Pattern Recognition
Prototype networks provide an intrinsic case based explanation mechanism, but their interpretability is often undermined by prototype collapse, where multiple prototypes degenerate to highly redundant evidence. We attribute this failure mode to the terminal dynamics of Neural Collapse, where cross entropy optimization suppresses intra class variance and drives class conditional features toward a low dimensional limit. To mitigate this, we propose Adaptive Manifold Prototypes (AMP), a framework that leverages Riemannian optimization on the Stiefel manifold to represent class prototypes as orthonormal bases and make rank one prototype collapse infeasible by construction. AMP further learns class specific effective rank via a proximal gradient update on a nonnegative capacity vector, and introduces spatial regularizers that reduce rotational ambiguity and encourage localized, non overlapping part evidence. Extensive experiments on fine-grained benchmarks demonstrate that AMP achieves state-of-the-art classification accuracy while significantly improving causal faithfulness over prior interpretable models.
title This Looks Distinctly Like That: Grounding Interpretable Recognition in Stiefel Geometry against Neural Collapse
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2603.08374