NeurIPS-2021 Intriguing Properties of Vision Transformers

· NeurIPS· · vision-language, transformer

Paper: Intriguing Properties of Vision Transformers

Properties of ViT — skim-level notes

  1. Because ViT adopts a transformer backbone, it can aggregate global context early; standard convolutional networks usually need many layers before they effectively capture image-wide information.
  2. Improved robustness: stronger performance under occlusion, distribution shift, adversarial attacks, permuted patch order, and in semantic segmentation settings.

<HR align=left color=#987cb9 SIZE=1>