Research map / Architectures and efficiency / Attention and transformers
Beyond Self-Attention: External Attention Using Two Linear Layers for Visual Tasks
Meng-Hao Guo, Zheng-Ning Liu, Tai‐Jiang Mu, Shi‐Min Hu
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 self-attention with external attention (DoubleNorm) (ablation) | FCN ResNet-50 · PASCAL VOC val (FCN ablation, 30k iters) | mIoU: 76.6 -> 77.4 | Improved |
| add double normalization instead of softmax in external attention (ablation) | FCN + EA · PASCAL VOC val (FCN ablation, 30k iters) | mIoU: 75.3 -> 77.4 | Improved |
| use DoubleNorm instead of softmax in self-attention (ablation) | FCN + SA · PASCAL VOC val (FCN ablation, 30k iters) | mIoU: 76.2 -> 76.6 | Improved |
| increase memory units from 8 to 64 (ablation) | FCN + EA · PASCAL VOC val (FCN ablation, 30k iters) | mIoU: 77.1 -> 77.4 | Improved |
| increase memory units from 64 to 256 (ablation) | FCN + EA · PASCAL VOC val (FCN ablation, 30k iters) | mIoU: 77.4 -> 77.0 | Hurt |
| reduce output stride from 16 to 8 (ablation) | FCN + EA · PASCAL VOC val (FCN ablation, 30k iters) | mIoU: 77.4 -> 77.8 | Improved |
| use ResNet-101 instead of ResNet-50 (ablation) | FCN + EA · PASCAL VOC val (FCN ablation, 30k iters) | mIoU: 77.4 -> 78.3 | Improved |
| add external attention to FCN (ablation) | FCN ResNet-50 · PASCAL VOC val (FCN ablation, 30k iters) | mIoU: 75.7 -> 77.4 | Improved |
| use multi-head external attention instead of single-head (ablation) | T2T-ViT-7 Performer MEA · ImageNet-1K | top-1 accuracy: 63.2 -> 68.6 | Improved |
| increase input size from 224 to 384 (ablation) | T2T-ViT-7 MEA · ImageNet-1K | top-1 accuracy: 68.6 -> 70.9 | Improved |
| replace LN by BN in MLP blocks (ablation) | EAMLP-7 · ImageNet-1K | top-1 accuracy: 68.9 -> 70.0 | Improved |
| replace LN by BN in MLP blocks (ablation) | EAMLP-14/19 · ImageNet-1K | top-1 accuracy: 78.9 -> Failed | Hurt |
Closest papers by meaning
- Scaling Local Self-Attention for Parameter Efficient Visual Backbones (CVPR 2021 · similarity 0.90)
- On the Integration of Self-Attention and Convolution (CVPR 2022 · similarity 0.89)
- Exploring Self-Attention for Image Recognition (CVPR 2020 · similarity 0.89)
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.87)
- Revisiting the Integration of Convolution and Attention for Vision Backbone (NeurIPS 2024 · similarity 0.86)
- Degrees of Freedom for Linear Attention: Distilling Softmax Attention with Optimal Feature Efficiency (NeurIPS 2025 · similarity 0.86)
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.
