Research map / Architectures and efficiency / Attention and transformers
Locality Guidance for Improving Vision Transformers on Tiny Datasets
Kehan Li, Runyi Yu, Zhennan Wang, Li Yuan, Guoli Song, Jie Chen
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 locality guidance from a lightweight CNN | DeiT-Ti · CIFAR-100 | top-1 accuracy: 65.08 -> 78.15 | Improved |
| add locality guidance from a lightweight CNN | DeiT-Ti · Oxford Flowers | top-1 accuracy: 50.06 -> 68.50 | Improved |
| add locality guidance from a lightweight CNN | PVTv2-B0 · CIFAR-100 | top-1 accuracy: 77.44 -> 79.30 | Improved |
| add locality guidance from a lightweight CNN | PVT-Ti · Chaoyang | top-1 accuracy: 73.68 -> 85.65 | Improved |
| add locality guidance (ours) instead of L_drloc | T2T-ViT-14 · CIFAR-100 (100 epochs) | top-1 accuracy: 68.03 -> 77.84 | Improved |
| use feature guidance instead of logit distillation (DeiT-Hard) | DeiT-Tiny · CIFAR-100 | top-1 accuracy: 73.25 -> 78.15 | Improved |
| raise guidance ratio R from 0.25 to 1.0 (use all CNN features) (ablation) | DeiT-Tiny + guidance · CIFAR-100 (100 epochs) | top-1 accuracy: 70.56 -> 77.29 | Improved |
| add guidance loss weight beta (0.0 -> 2.5) (ablation) | DeiT-Tiny + guidance · CIFAR-100 (100 epochs) | top-1 accuracy: 65.08 -> 77.29 | Improved |
| use learnable linear projection instead of SP (ablation) | DeiT-Tiny + guidance · CIFAR-100 (100 epochs) | top-1 accuracy: 67.36 -> 77.29 | Improved |
| use learnable linear projection instead of AT (ablation) | DeiT-Tiny + guidance · CIFAR-100 (100 epochs) | top-1 accuracy: 73.51 -> 77.29 | Improved |
| use a stronger guidance CNN, ResNet-110 instead of ResNet-56 (ablation) | DeiT-Tiny + guidance · CIFAR-100 (100 epochs) | top-1 accuracy: 77.29 -> 76.62 | No clear effect |
| use a weaker guidance CNN, ResNet-20 instead of ResNet-56 (ablation) | DeiT-Tiny + guidance · CIFAR-100 (100 epochs) | top-1 accuracy: 77.29 -> 72.91 | Hurt |
Closest papers by meaning
- TinyViT: Fast Pretraining Distillation for Small Vision Transformers (ECCV 2022 · similarity 0.89)
- Efficient Training of Visual Transformers with Small Datasets (NeurIPS 2021 · similarity 0.88)
- Bridging the Gap Between Vision Transformers and Convolutional Neural Networks on Small Datasets (NeurIPS 2022 · similarity 0.88)
- ConViT: Improving Vision Transformers with Soft Convolutional Inductive Biases (ICML 2021 · similarity 0.88)
- Boosting Vanilla Lightweight Vision Transformers via Re-parameterization (ICLR 2024 · similarity 0.88)
- Closest Neighbors are Harmful for Lightweight Masked Auto-encoders (CVPR 2025 · 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.
