Research map / Architectures and efficiency / Attention and transformers
Scaling Local Self-Attention for Parameter Efficient Visual Backbones
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 unmasked blocked local attention instead of pixel-centered masked attention (ablation) | HaloNet-50 · ImageNet classification | accuracy | Helped |
| Add label smoothing and RandAugment (ablation) | HaloNet-50 · ImageNet validation | top-1 accuracy (%): 78.6 -> 79.9; +1.3 | Helped |
| Add squeeze-and-excitation and SiLU/Swish-1 (ablation) | HaloNet-50 · ImageNet validation | top-1 accuracy (%): 78.6 -> 79.0; 0.4 | Helped |
| Add squeeze-and-excitation alone (ablation) | HaloNet-50 · ImageNet validation | top-1 accuracy (%): 78.6 -> 78.6; 0.0 | No clear effect |
| Increase image size (ablation) | HaloNet · ImageNet classification | accuracy | Helped |
| Increase attention window size from 6x6 to 12x12 (ablation) | HaloNet-50 · ImageNet classification | accuracy: 1.3% | Helped |
| Use a larger query block for a fixed window size (ablation) | HaloNet-50 · ImageNet classification | accuracy | No clear effect |
| Remove haloing and use non-overlapping attention blocks (ablation) | HaloNet-50 · ImageNet classification | accuracy | Hurt |
| Replace the attention downsampling plus post-attention average pooling with strided attention (ablation) | HaloNet · Downsampling layers | accuracy | No clear effect |
| Replace attention with convolution in stages 1 and 2 (ablation) | HaloNet H7 · ImageNet classification | top-1 accuracy (%): 84.9 -> 84.7; normalized train time 1.9 -> 1.0 | Mixed |
| Add local attention in the final backbone layers (ablation) | ResNet-50 + SE · COCO detection and instance segmentation | AP^bb / AP^mk: +0.7 / +0.7 | Helped |
| Add local attention in the final backbone layers (ablation) | ResNet-101 + SE · COCO detection and instance segmentation | AP^bb / AP^mk: +0.9 / +0.6 | Helped |
Closest papers by meaning
- Stand-Alone Self-Attention in Vision Models (NeurIPS 2019 · similarity 0.93)
- Exploring Self-Attention for Image Recognition (CVPR 2020 · similarity 0.91)
- Self-Attention-Based Contextual Modulation Improves Neural System Identification (ICLR 2025 · similarity 0.90)
- 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)
- Probing the Mid-level Vision Capabilities of Self-Supervised Learning (CVPR 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.