Research map / Architectures and efficiency / Attention and transformers
An Empirical Study of Spatial Attention Mechanisms in Deep Networks
Xizhou Zhu, Dazhi Cheng, Zheng Zhang, Stephen Lin, Jifeng Dai
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 Transformer attention (full 1111) over no attention (ablation) | ResNet-50 detector · COCO object detection | mAP: 36.4 -> 38.8 | Improved |
| add Transformer attention (full 1111) over no attention (ablation) | Transformer base · WMT14 En-De | BLEU: 20.9 -> 27.7 | Improved |
| add query and key content term E1 to self-attention (ablation) | Transformer attention · Cityscapes | mIoU: 76.6 -> 76.7 | No clear effect |
| add query and key content term E1 to encoder-decoder attention (ablation) | Transformer attention · WMT14 En-De | BLEU | Improved |
| add key content only term E3 to E2 (ablation) | Transformer attention · COCO object detection | mAP: 38.6 -> 38.8 | Improved |
| replace E2 with deformable convolution (ablation) | Transformer attention · COCO object detection | mAP: 38.6 -> 39.9 | Improved |
| replace E2 with deformable convolution (ablation) | Transformer attention · Cityscapes | mIoU: 76.3 -> 77.2 | Improved |
| replace E2 with deformable convolution (ablation) | Transformer attention · WMT14 En-De | BLEU: 27.4 -> 27.3 | No clear effect |
| add key content only term E3 to deformable convolution (ablation) | deformable convolution · COCO object detection | mAP: 39.9 -> 40.8 | Improved |
| replace E2 with dynamic convolution (kernel 31) (ablation) | Transformer attention · COCO object detection | mAP: 38.6 -> 37.9 | Hurt |
| replace E2 with dynamic convolution (kernel 31) (ablation) | Transformer attention · Cityscapes | mIoU: 75.5 -> 74.2 | Hurt |
| replace E2 with dynamic convolution (kernel 31) (ablation) | Transformer attention · WMT14 En-De | BLEU: 27.4 -> 27.6 | No clear effect |
Closest papers by meaning
- Rethinking Attention: Exploring Shallow Feed-Forward Neural Networks as an Alternative to Attention Layers in Transformers (Student Abstract) (AAAI 2024 · similarity 0.88)
- Towards Interpretable and Efficient Attention: Compressing All by Contracting a Few (NeurIPS 2025 · similarity 0.87)
- On the Role of Attention Masks and LayerNorm in Transformers (NeurIPS 2024 · similarity 0.87)
- Hierarchical Self-Attention: Generalizing Neural Attention Mechanics to Multi-Scale Problems (NeurIPS 2025 · similarity 0.87)
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.87)
- Graph Convolutions Enrich the Self-Attention in Transformers! (NeurIPS 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.
