Research map / Architectures and efficiency / Attention and transformers
Dynamic Token Normalization improves Vision Transformers
Read the paperCheck 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 |
|---|---|---|---|
| Replace LN with DTN | ViT-S · ImageNet | Top-1 accuracy (%): 79.9 -> 80.6 | Helped |
| Replace LN with DTN | PVT-Tiny · ImageNet | Top-1 accuracy (%): 75.1 -> 76.3 | Helped |
| Replace LN with DTN | Swin-T · ImageNet | Top-1 accuracy (%): 81.2 -> 81.9 | Helped |
| Replace LN with DTN | ViT-S · ImageNet-C | mCE: 42.8 -> 40.6 | Helped |
| Replace LN with DTN | ViT-S* · ImageNet-C | mCE: 41.9 -> 38.0 | Helped |
| Replace LN with DTN | Transformer · Long ListOps (LRA) | accuracy: 36.4 -> 37.0 | Helped |
| Replace LN with DTN | ViT-S* with DINO · ImageNet linear evaluation (100 epochs) | Top-1 accuracy (%): 74.7 -> 75.2 | Helped |
| Replace LN with DTN | PVT-Tiny with RetinaNet · COCO val2017 | AP: 36.7 -> 38.0 | Helped |
| Combine intra-token statistics with inter-token statistics (ablation) | IN variant of DTN · ImageNet | Top-1 accuracy (%): 77.6 -> 81.2 | Helped |
| Use a learnable mixture ratio λ^h instead of fixing λ^h=0.5 (ablation) | DTN with learnable positional matrix · ImageNet | Top-1 accuracy (%): 81.3 -> 81.7 | Helped |
| Initialize a^h with truncated normal rather than local-neighborhood initialization (ablation) | DTN on ViT-T* · ImageNet | Top-1 accuracy (%): 73.2 -> 72.7 | Hurt |
| Use DTN in more transformer layers (ablation) | ViT-S* · ImageNet | Top-1 accuracy (%): 81.2 -> 81.4 -> 81.5 -> 81.7 | Helped |
Closest papers by meaning
- SLAB: Efficient Transformers with Simplified Linear Attention and Progressive Re-parameterized Batch Normalization (ICML 2024 · similarity 0.87)
- Morphing Tokens Draw Strong Masked Image Models (ICLR 2025 · similarity 0.87)
- ImagePiece: Content-aware Re-tokenization for Efficient Image Recognition (AAAI 2025 · similarity 0.86)
- Transformers without Normalization (CVPR 2025 · similarity 0.86)
- Differentiable Hierarchical Visual Tokenization (NeurIPS 2025 · similarity 0.86)
- Learning to Merge Tokens via Decoupled Embedding for Efficient Vision Transformers (NeurIPS 2024 · 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.