Research map / Architectures and efficiency / Attention and transformers
Shunted Self-Attention via Multi-Scale Token Aggregation
Check 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 multi-scale shunted self-attention | Shunted Transformer · ImageNet-1K | Top-1 accuracy: 82.2 -> 82.9 | Helped |
| use multi-scale shunted self-attention | Shunted Transformer · COCO, Mask R-CNN 3× schedule + MS | bbox AP: 47.2 -> 49.1 | Helped |
| use multi-scale shunted self-attention | Shunted Transformer · ADE20K, Upernet 160K | mIoU: 45.8 -> 48.9 | Helped |
| replace non-overlapping patch embedding with the proposed patch embedding (ablation) | Shunted-S · ImageNet-1K | Top-1 accuracy: 81.5 -> 82.9 | Helped |
| replace overlapping patch embedding with the proposed patch embedding (ablation) | Shunted-S · ImageNet-1K | Top-1 accuracy: 82.6 -> 82.9 | Helped |
| replace linear aggregation with the proposed token aggregation function (ablation) | Shunted-S · ImageNet-1K | Top-1 accuracy: 82.1 -> 82.9 | Helped |
| replace convolutional spatial reduction with the proposed token aggregation function (ablation) | Shunted-S · ImageNet-1K | Top-1 accuracy: 82.6 -> 82.9 | Helped |
| add a detail-specific depth-wise convolution to the feed-forward block (ablation) | ViT · ImageNet-1K | Top-1 accuracy: 79.8 -> 80.7 | Helped |
| add a detail-specific depth-wise convolution to the feed-forward block (ablation) | Shunted Transformer · ImageNet-1K | Top-1 accuracy: 82.6 -> 82.9 | Helped |
| replace the traditional feed-forward layer with Conv-Feedforward (ablation) | Shunted Transformer · ImageNet-1K | Top-1 accuracy: 82.6 -> 82.7 | Helped |
| replace traditional feed-forward with Conv-Feedforward (ablation) | ViT · ImageNet-1K | Top-1 accuracy: 79.8 -> 80.5 | Helped |
Closest papers by meaning
- Focal Attention for Long-Range Interactions in Vision Transformers (NeurIPS 2021 · similarity 0.90)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · similarity 0.89)
- Skip-Attention: Improving Vision Transformers by Paying Less Attention (ICLR 2024 · similarity 0.88)
- Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers (CVPR 2024 · similarity 0.87)
- Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials (NeurIPS 2025 · similarity 0.87)
- One Meta-tuned Transformer is What You Need for Few-shot Learning (ICML 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.