Vision Transformers 是循环注意力学习者
摘要
自注意力机制一直是视觉 Transformer 取得进展的关键因素。然而,其二次复杂度在高分辨率场景下带来沉重的计算负担,限制了实际应用。先前方法通过引入局部性或稀疏性等手工模式来缓解此问题,却不可避免地损害模型容量。本文提出一种新颖注意力范式,称为循环注意力(Circulant Attention),通过利用自注意力的内在高效模式实现。具体而言,我们首先识别到视觉 Transformer 中的自注意力矩阵常常近似于具有循环子块的块循环矩阵(Block Circulant matrix with Circulant Blocks, BCCB),这是一种其矩阵乘法可在 时间内完成的结构矩阵。基于此有趣的模式,我们显式地将注意力图建模为其最近的 BCCB 矩阵,并提出一种高效计算算法进行快速计算。 resulting approach closely mirrors vanilla self-attention, differing only in its use of BCCB matrices. Since our design is inspired by the inherent efficient paradigm, it not only delivers computation complexity, but also largely maintains the capacity of standard self-attention. Extensive experiments on diverse visual tasks demonstrate the effectiveness of our approach, establishing circulant attention as a promising alternative to self-attention for vision Transformer architectures. Code is available at https://github.com/LeapLabTHU/Circulant-Attention.
引用
@article{arxiv.2512.21542,
title = {Vision Transformers are Circulant Attention Learners},
author = {Dongchen Han and Tianyu Li and Ziyi Wang and Gao Huang},
journal= {arXiv preprint arXiv:2512.21542},
year = {2025}
}
备注
AAAI 2026