arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2603.00518cs.CV
收藏

Vision-TTT: 高效的视觉表征学习与测试时间训练

Vision-TTT: Efficient and Expressive Visual Representation Learning with Test-Time Training

Zhejiang University(浙江大学);Amazon Web Services(亚马逊网络服务)

Quan Kong, Yanru Xiao, Yuhao Shen, Cong Wang

AI总结:

本文提出Vision-TTT,通过测试时间训练方法提升视觉表征效率与表达能力,在ImageNet分类中达到77.7%的Top-1准确率,同时在计算效率和内存消耗上显著优于现有方法。

摘要:

学习高效且表达能力强的视觉表征一直是计算机视觉研究的目标。尽管视觉Transformer(ViTs)逐渐取代传统卷积神经网络(CNNs)成为更可扩展的视觉学习者,但自注意力机制的二次复杂度限制了其应用。为解决这一挑战,我们引入新的线性时间序列建模方法测试时间训练(TTT)到视觉领域,提出Vision-TTT,将视觉序列视为数据集,并以新颖的自监督学习方式压缩视觉标记序列。通过双数据集策略和基于Conv2d的数据集预处理,Vision-TTT有效扩展了标准TTT,以建模具有全局感受野的2D视觉相关性。广泛实验表明,\texttt{Vittt-T/S/B}在ImageNet分类中分别达到77.7\%,81.8\%,82.7\%的Top-1准确率,并在下游任务中显著优于其对应方法。在1280×1280分辨率下,\texttt{Vittt-T}比DeiT-T减少79.4\%的FLOPs,运行速度快4.72倍,内存消耗少88.9\%。这些结果证明了Vision-TTT作为下一代通用视觉主干的表达能力和效率。

英文摘要:

Learning efficient and expressive visual representation has long been the pursuit of computer vision research. While Vision Transformers (ViTs) gradually replace traditional Convolutional Neural Networks (CNNs) as more scalable vision learners, their applications are plagued by the quadratic complexity of the self-attention mechanism. To address the challenge, we introduce a new linear-time sequence modeling method Test-Time Training (TTT) into vision and propose Vision-TTT, which treats visual sequences as datasets and compresses the visual token sequences in a novel self-supervised learning manner. By incorporating the dual-dataset strategy and Conv2d-based dataset preprocessing, Vision-TTT effectively extends vanilla TTT to model 2D visual correlations with global receptive fields. Extensive experiments show that \texttt{Vittt-T/S/B} achieve $77.7\%,81.8\%,82.7\%$ Top-1 accuracy on ImageNet classification and also greatly outperform their counterparts on downstream tasks. At $1280\times1280$ resolution, \texttt{Vittt-T} reduces FLOPs by $79.4\%$ and runs $4.72\times$ faster with $88.9\%$ less memory than DeiT-T. These results demonstrate the expressiveness and efficiency of Vision-TTT as a strong candidate for the next-generation generic visual backbone.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑