Research map / Architectures and efficiency / Attention and transformers
KVT: k-NN Attention for Boosting Vision Transformers
Pichao Wang, Xue Wang, Fan Wang, Ming Lin, Shuning Chang, Hao Li, Rong Jin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace dense attention with k-NN attention | DeiT-Tiny · ImageNet-1K val | top-1: 72.2% -> 73.0% | Improved |
| replace dense attention with k-NN attention | PiT-Base · ImageNet-1K val | top-1: 82.0% -> 82.6% | Improved |
| replace dense attention with k-NN attention | Swin-Tiny · ImageNet-1K val | top-1: 81.2% -> 81.3% | No clear effect |
| replace dense attention with k-NN attention | Swin-T · COCO val | mAP(box): 46.0 -> 46.2 | Improved |
| replace dense attention with k-NN attention | Twins-SVT-Base · ADE20K val | mIoU: 47.4 -> 47.9 | Improved |
| use k-NN attention with k about half of the tokens instead of dense attention (ablation) | DeiT-Tiny · ImageNet-1K val | top-1: 72.2% -> 73.0% | Improved |
| add k-NN attention for faster early convergence at 10 epochs (ablation) | DeiT-S · ImageNet-1K val | top-1: 29.1% -> 31.3% | Improved |
| add k-NN attention at 10 epochs (ablation) | CvT-13 · ImageNet-1K val | top-1: 51.4% -> 54.2% | Improved |
| add k-NN attention at 300 epochs (ablation) | DeiT-S · ImageNet-1K val | top-1: 79.8% -> 80.0% | Improved |
| use k-NN attention instead of softmax temperature (ablation) | DeiT-Tiny · ImageNet-1K val | top-1: 72.5 -> 73.0 | Improved |
| use small softmax temperature t=0.05 or 0.1 (ablation) | DeiT-Tiny softmax · ImageNet-1K val | top-1: 72.2 -> crash | Hurt |
| use larger softmax temperature t=0.75 (ablation) | DeiT-Tiny softmax · ImageNet-1K val | top-1: 72.2 -> 72.5 | No clear effect |
Closest papers by meaning
- Learned Queries for Efficient Local Attention (CVPR 2022 · similarity 0.89)
- kNN Attention Demystified: A Theoretical Exploration for Scalable Transformers (ICLR 2025 · similarity 0.89)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · similarity 0.89)
- A Survey on Vision Transformer (TPAMI 2022 · similarity 0.89)
- Skip-Attention: Improving Vision Transformers by Paying Less Attention (ICLR 2024 · similarity 0.88)
- Vision Transformer With Deformable Attention (CVPR 2022 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
