Research map / Architectures and efficiency / Attention and transformers
Neighborhood Attention Transformer
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| use NAT-T instead of Swin-T | ImageNet-1K classification · ImageNet-1K | Top-1 accuracy (%): 81.3 -> 83.2 | Helped |
| use NAT-T instead of ConvNeXt-T | MS-COCO Mask R-CNN · MS-COCO, Mask R-CNN 3x | Box AP: 46.2 -> 47.7 | Helped |
| use NAT-T instead of Swin-T | MS-COCO Cascade Mask R-CNN · MS-COCO, Cascade Mask R-CNN 3x | Box AP: 50.4 -> 51.4 | Helped |
| use NAT-T instead of ConvNeXt-T | ADE20K UPerNet · ADE20K | mIoU (multi-scale): 46.7 -> 48.4 | Helped |
| replace Swin WSA/SWSA attention with NA (ablation) | Swin-T architecture · ImageNet-1K | Top-1 accuracy (%): 81.3% -> 81.8% | Helped |
| replace Swin WSA/SWSA attention with SASA (ablation) | Swin-T architecture · ImageNet-1K | Top-1 accuracy (%): 81.3% -> 81.6% | Helped |
| use overlapping convolution downsampling instead of patch embedding and patch merge (ablation) | Swin-T · ImageNet-1K | Top-1 accuracy (%): 81.29 -> 81.78 | Helped |
| use a deeper-thinner configuration (ablation) | Swin-T with convolutional downsampling · ImageNet-1K | Top-1 accuracy (%): 81.78 -> 82.72 | Helped |
| replace SWSA with SASA (ablation) | Deeper-thinner model with convolutional downsampling · ImageNet-1K | Top-1 accuracy (%): 82.72 -> 82.54 | Hurt |
| replace SWSA with NA (ablation) | Deeper-thinner model with convolutional downsampling · ImageNet-1K | Top-1 accuracy (%): 82.72 -> 83.20 | Helped |
| increase NA kernel size from 3×3 to 7×7 (ablation) | NAT-T · ImageNet-1K | Top-1 accuracy (%): 81.4 -> 83.2 | Helped |
| increase NA kernel size from 7×7 to 9×9 (ablation) | NAT-T · ImageNet-1K | Top-1 accuracy (%): 83.2 -> 83.1 | No clear effect |
Closest papers by meaning
- Faster Neighborhood Attention: Reducing the O(n^2) Cost of Self Attention at the Threadblock Level (NeurIPS 2024 · similarity 0.91)
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.88)
- LF-ViT: Reducing Spatial Redundancy in Vision Transformer for Efficient Image Recognition (AAAI 2024 · similarity 0.87)
- Breaking the Low-Rank Dilemma of Linear Attention (CVPR 2025 · similarity 0.87)
- FasterViT: Fast Vision Transformers with Hierarchical Attention (ICLR 2024 · similarity 0.87)
- Skip-Attention: Improving Vision Transformers by Paying Less Attention (ICLR 2024 · similarity 0.87)
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.