Research map / Architectures and efficiency / Attention and transformers
Scratching Visual Transformer's Back with Uniform Attention
Nam Hyeon-Woo, Kim Yu-Ji, Byeongho Heo, Dongyoon Han, Seong Joon Oh, Tae-Hyun Oh
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 |
|---|---|---|---|
| add Context Broadcasting (CB) to ViT-S | ViT-S · ImageNet-1K val | top-1 accuracy: 79.9 -> 80.8 | Improved |
| insert CB at end of MLP block rather than in MSA or both (ablation) | CB in ViT-S · ImageNet-1K val | top-1 accuracy: 80.1 -> 80.5 | Improved |
| place CB at Mid or End of MLP rather than Front (ablation) | CB in ViT-S · ImageNet-1K val | top-1 accuracy: 79.9 -> 80.5 | Improved |
| insert CB only in upper layers instead of all layers (ablation) | CB · ImageNet-1K val, ViT-Ti/S/B | top-1 accuracy: 80.5 -> 80.8 (ViT-S) | Improved |
| compare CB against Squeeze-and-Excitation module | ViT-S · ImageNet-1K val | top-1 accuracy: 80.3 -> 80.8 | Improved |
| reduce number of attention heads with CB (ablation) | ViT-Ti with CB · ImageNet-1K val | accuracy gap | Improved |
| add CB to ViT-S for attention entropy | ViT-S · layer-wise attention entropy | entropy | Improved |
| add CB to ViT backbones for segmentation | ViT-Ti/S/B UperNet · ADE20K 40K iterations | mIoU: +1.0, +0.4, +0.8 | Improved |
| add CB_S to ViT-S for segmentation | ViT-S UperNet · ADE20K 160K iterations | mIoU: 43.3 -> 43.1 | Hurt |
| add CB to ViT-Ti for detection | ViT-Ti Mask R-CNN · COCO val | AP box: 34.8 -> 35.1 | Improved |
| add CB to ViT-S for robustness | ViT-S · FGSM attack | accuracy: 27.2 -> 32.3 | Improved |
| add CB_S to ViLT image tokens, text tokens or both | ViLT · VQAv2 | accuracy: 71.28 -> 71.46 | Improved |
Closest papers by meaning
- Mobile Attention: Mobile-Friendly Linear-Attention for Vision Transformers (ICML 2024 · similarity 0.86)
- Dissecting Query-Key Interaction in Vision Transformers (NeurIPS 2024 · similarity 0.86)
- QT-ViT: Improving Linear Attention in ViT with Quadratic Taylor Expansion (NeurIPS 2024 · similarity 0.86)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · similarity 0.86)
- On the Role of Attention Masks and LayerNorm in Transformers (NeurIPS 2024 · similarity 0.85)
- Attention Transfer Is Not Universally Effective for Vision Transformers (NeurIPS 2026 · similarity 0.85)
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.
