Research map / Label-efficient and robust learning / Self-supervised and contrastive
Unsqueeze [CLS] Bottleneck to Learn Rich Representations
Qing Su, Shihao Ji
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 DINO/iBOT objectives with UDI (stratified local-prediction consolidation plus multimodal prediction) | ViT-S/16 SSL · ImageNet-1K linear probing | top-1 accuracy: 77.4 -> 77.6 | Improved |
| use the extra multimodal class token for linear evaluation | UDI · ImageNet-1K linear probing | top-1 accuracy: 77.6 -> 75.1 | Hurt |
| use the extra multimodal class token for transfer learning and fine-tuning | UDI · ImageNet-1K fine-tuning | top-1 accuracy: 82.6 -> 83.0 | Improved |
| add UDI self-attention pooled patch-level semantic constraint to DINO (ablation) | DINO ViT-S/16 100 epochs · COCO detection | AP box: 40.6 -> 42.6 | Improved |
| add shared projector for image- and patch-level representations (ablation) | UDI · ImageNet-1K linear | top-1 accuracy: 75.6 -> 75.8 | Improved |
| add shared projector to EsViT constraint (ablation) | EsViT · COCO detection | AP box: 41.8 -> 41.2 | Hurt |
| add multimodal objective L_cls+ (ablation) | UDI · ImageNet-1K k-NN | k-NN top-1: 73.0 -> 74.9 | Improved |
| extra class token without multimodal objective (register token only) (ablation) | UDI · ImageNet-1K linear | top-1 accuracy | No clear effect |
| increase blending factor alpha from 0.1 to 0.5 (ablation) | UDI · ImageNet-1K linear | top-1 accuracy: 75.9 -> 76.3 | Improved |
| add UDI to DINO-style pretraining versus SelfPatch | ViT-S/16 · COCO Mask R-CNN | AP box: 42.1 -> 43.2 | Improved |
| use UDI with register token (DINO+reg.) | DINO · ImageNet-1K linear | top-1 accuracy: 76.1 -> 76.1 | No clear effect |
Closest papers by meaning
- UDON: Universal Dynamic Online distillatioN for generic image representations (NeurIPS 2024 · similarity 0.89)
- A Survey on Self-Supervised Learning: Algorithms, Applications, and Future Trends (TPAMI 2024 · similarity 0.86)
- Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation (ICLR 2025 · similarity 0.86)
- Towards Universal Dataset Distillation via Task-Driven Diffusion (CVPR 2025 · similarity 0.85)
- Minimal Semantic Sufficiency Meets Unsupervised Domain Generalization (NeurIPS 2025 · similarity 0.84)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.84)
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.
