Align before Fuse: Vision and Language Representation Learning with Momentum Distillation

  • 类型：arxiv
  • 标识：2107.07651
  • 链接：https://arxiv.org/abs/2107.07651
  • 主题：multimodal
  • 主分类：multimodal
  • 形态：method
  • 被引：2913
  • 被引来源：Semantic Scholar
  • S2被引：2913
  • OpenAlex被引：823
  • 影响力被引：343
  • TLDR：A contrastive loss to ALign the image and text representations BEfore Fusing (ALBEF) them through cross-modal attention, which enables more grounded vision and language representation learning and proposes momentum distillation, a self-training method which learns from pseudo-targets produced by a momentum model.
  • OpenAlex ID：W3184735396
  • OpenAlex DOI：10.48550/arxiv.2107.07651
  • DOI：10.48550/arxiv.2107.07651
  • DOI来源：OpenAlex
  • 开放获取：green
  • 开放获取链接：https://arxiv.org/pdf/2107.07651
  • OpenAlex更新：2026-08-23
  • 待LLM分类：否
  • 标题中文：Align before Fuse: Vision and Language Representation Learning with Momentum Distillation
  • TLDR中文：提出在通过跨模态注意力融合之前对齐图像与文本表示的对比损失（ALBEF），可实现更扎实的视觉-语言表征学习；并提出动量蒸馏，一种利用动量模型生成伪目标进行自训练的方法。
  • 来源文件：
  • [OpenAlex discover]
  • [S2 enrich]