arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2305.04474cs.CVcs.AI
收藏

TRIPS:基于文本相关图像块选择的高效视觉-语言预训练

TRIPS: Efficient Vision-and-Language Pre-training with Text-Relevant Image Patch Selection

National Engineering Research Center for Software Engineering, Peking University(软件工程国家工程研究中心,北京大学);DAMO Academy, Alibaba Group(阿里巴巴集团达摩院)

Chaoya Jiang, Haiyang Xu, Chenliang Li, Miang Yan, Wei Ye, Shikun Zhang, Bin Bi, Songfang Huang

AI总结:

本文提出TRIPS视觉-语言预训练模型,通过文本引导的图像块选择动态保留关键视觉token并融合非关键token,在不增加ViT参数的情况下实现40%加速并保持下游性能。

摘要:

Vision Transformers(ViT)已被广泛应用于大规模视觉与语言预训练(VLP)模型。尽管以往的VLP工作已经证明了ViT的有效性,但它们仍受到长视觉序列所带来的计算效率问题困扰。为解决这一问题,本文提出一种高效的视觉-语言预训练模型,即采用文本相关图像块选择的TRIPS。该模型通过视觉骨干网络中的文本引导块选择层逐步缩短视觉序列,以实现高效训练与推理。块选择层能够动态计算依赖于文本的视觉注意力,在文本引导下识别受到关注的图像token,并以端到端方式融合未受关注的token。同时,TRIPS不会为ViT引入额外参数。在多种常用基准数据集上的实验结果表明,TRIPS相比以往类似VLP模型获得40%的加速,同时在下游任务上保持具有竞争力或更好的性能。

英文摘要:

Vision Transformers (ViTs) have been widely used in large-scale Vision and Language Pre-training (VLP) models. Though previous VLP works have proved the effectiveness of ViTs, they still suffer from computational efficiency brought by the long visual sequence. To tackle this problem, in this paper, we propose an efficient vision-and-language pre-training model with \textbf{T}ext-\textbf{R}elevant \textbf{I}mage \textbf{P}atch \textbf{S}election, namely TRIPS, which reduces the visual sequence progressively with a text-guided patch-selection layer in the visual backbone for efficient training and inference. The patch-selection layer can dynamically compute text-dependent visual attention to identify the attentive image tokens with text guidance and fuse inattentive ones in an end-to-end manner. Meanwhile, TRIPS does not introduce extra parameters to ViTs. Experimental results on a variety of popular benchmark datasets demonstrate that TRIPS gain a speedup of 40\% over previous similar VLP models, yet with competitive or better downstream task performance.

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。

↑