Research map / Architectures and efficiency / Attention and transformers
Token-Label Alignment for Vision Transformers
Han Xiao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, Jiwen Lu
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 token-label alignment to CutMix targets | DeiT-S · ImageNet-1k val | top-1 accuracy (%): 79.8 -> 80.6 | Improved |
| align labels layer by layer instead of only last-layer attention (ablation) | DeiT-S · ImageNet-1k val | top-1 accuracy (%): 80.1 -> 80.6 | Improved |
| align only at some middle layers (ablation) | DeiT-S · ImageNet-1k val | top-1 accuracy (%): 79.8 -> 80.2 | Improved |
| keep normalization in alignment (ablation) | TL-Align · ImageNet-1k val | top-1 accuracy (%): 80.3 -> 80.6 | Improved |
| apply TL-Align to MixUp+CutMix (ablation) | DeiT-S · ImageNet-1k val | top-1 accuracy (%): 79.8 -> 80.2 | Improved |
| apply TL-Align to block-wise mixing (ablation) | DeiT-S · ImageNet-1k val | top-1 accuracy (%): 80.0 -> 80.3 | Improved |
| use TL-Align instead of TransMix on same recipe | DeiT-S · ImageNet-1k val | top-1 accuracy (%): 80.1 -> 80.6 | Improved |
| use saliency or attention-based mixing (SaliencyMix, Attentive-CutMix) instead of CutMix | DeiT-S · ImageNet-1k val | top-1 accuracy (%): 79.8 -> 79.2 (SaliencyMix), 77.5 (Attentive-CutMix) | Hurt |
| add TL-Align on Swin-T for COCO detection | Swin-T Cascade Mask R-CNN · COCO val | AP box: 50.4 -> 50.5 | No clear effect |
| add TL-Align to DeiT-T | DeiT-T · ImageNet-A | top-1 accuracy (%): 7.7 -> 6.1 | Hurt |
| add TL-Align to DeiT-B | DeiT-B · ImageNet-R | top-1 accuracy (%): 45.3 -> 44.4 | Hurt |
| add TL-Align training cost | DeiT-S · training speed | image/s: 322 -> 311 | Hurt |
Closest papers by meaning
- TokenMix: Rethinking Image Mixing for Data Augmentation in Vision Transformers (ECCV 2022 · similarity 0.90)
- Learning to Merge Tokens via Decoupled Embedding for Efficient Vision Transformers (NeurIPS 2024 · similarity 0.87)
- ImagePiece: Content-aware Re-tokenization for Efficient Image Recognition (AAAI 2025 · similarity 0.86)
- Token-to-Token Alignment of Text Embeddings for Semantic Blending (NeurIPS 2026 · similarity 0.85)
- MADTP: Multimodal Alignment-Guided Dynamic Token Pruning for Accelerating Vision-Language Transformer (CVPR 2024 · similarity 0.85)
- Video Token Merging for Long Video Understanding (NeurIPS 2024 · similarity 0.85)
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.
