Research map / Label-efficient and robust learning / Knowledge distillation
SiNGER: A Clearer Voice Distills Vision Transformers Further
Geunhyeok Yu, Sunjae Jeong, Yoonyoung Choi, Jaeseung Kim, Hyoseok Hwang
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 outlier suppression loss (ablation) | SiNGER ViT-L->ViT-T with nullspace init · NYUd-v2 | RMSE: 1.1943 -> 1.1870 | Improved |
| add information preservation loss to outlier loss (ablation) | SiNGER ViT-L->ViT-T · NYUd-v2 | RMSE: 1.1870 -> 1.1690 | Improved |
| add information preservation loss to outlier loss (ablation) | SiNGER ViT-L->ViT-T · ImageNet-1K val | accuracy (%): 11.04 -> 11.57 | Improved |
| add information preservation loss (ablation) | SiNGER · teacher vs refined teacher feature distance | Gram/feature distance: 14.22 -> 7.25 | Improved |
| add information preservation loss (ablation) | SiNGER · teacher vs student feature distance | distance: 72.36 -> 41.71 | Improved |
| use nullspace initialization instead of random (ablation) | SiNGER adapter · NYUd-v2 (ablation setup) | RMSE: 1.3048 -> 1.1943 | Improved |
| use nullspace initialization instead of random (ablation) | SiNGER adapter · alignment to null basis | E_safe: 0.2532 -> 0.8337 | Improved |
| increase LoRA rank r from 16 to 64 (ablation) | SiNGER · NYUd-v2 | RMSE: 1.4395 -> 1.6485 | Hurt |
| lower LoRA rank r from 16 to 8 (ablation) | SiNGER · NYUd-v2 | RMSE: 1.4395 -> 1.4545 | Hurt |
| change quantile threshold alpha from 0.95 to 0.90 (ablation) | SiNGER · NYUd-v2 | RMSE: 1.4395 -> 1.5989 | Hurt |
| change quantile threshold alpha from 0.95 to 0.99 (ablation) | SiNGER · NYUd-v2 | RMSE: 1.4395 -> 1.5321 | Hurt |
| add intermediate layer 11 to distillation layers 17 and 23 (ablation) | SiNGER ViT-L->ViT-T · NYUd-v2 | RMSE: 0.9406 -> 0.9624 | Hurt |
Closest papers by meaning
- Masking Teacher and Reinforcing Student for Distilling Vision-Language Models (CVPR 2026 · similarity 0.87)
- No Head Left Behind – Multi-Head Alignment Distillation for Transformers (AAAI 2024 · similarity 0.87)
- Vision Transformers with Self-Distilled Registers (NeurIPS 2025 · similarity 0.86)
- Attention Transfer Is Not Universally Effective for Vision Transformers (NeurIPS 2026 · similarity 0.85)
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 2024 · similarity 0.85)
- Vision Transformers Don't Need Trained Registers (NeurIPS 2025 · 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.
