Research map / Label-efficient and robust learning / Knowledge distillation
From Knowledge Distillation to Self-Knowledge Distillation: A Unified Approach with Normalized Loss and Customized Soft Labels
Zhendong Yang, Ailing Zeng, Zhe Li, Tianke Zhang, Chun Yuan, Yu Li
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 normalized non-target loss (NKD) on top of target loss (ablation) | KD, ResNet34 to ResNet18 · ImageNet val | Top-1 accuracy (%): 71.06 -> 71.96 | Improved |
| use KD's unnormalized non-target loss instead of NKD's normalized one (ablation) | KD, ResNet34 to ResNet18 · ImageNet val | Top-1 accuracy (%): 71.96 -> 71.33 | Hurt |
| add target-class distillation alone (ablation) | KD, ResNet34 to ResNet18 · ImageNet val | Top-1 accuracy (%): 69.90 -> 71.06 | Improved |
| replace KD with NKD loss | ResNet18 on ImageNet · ImageNet val | Top-1 accuracy (%): 71.03 -> 71.96 | Improved |
| combine target and non-target distillation in USKD (ablation) | MobileNet · ImageNet val | Top-1 accuracy (%): 70.18 -> 70.38 | Improved |
| apply USKD to ViT-like models via an extra linear layer on a middle layer | DeiT-Tiny, DeiT-Small, Swin-Tiny, RegNetX-1.6GF · ImageNet val | Top-1 accuracy (%): DeiT-Tiny 74.42 -> 74.97 | Improved |
| use USKD-trained ResNet-50 as backbone for Mask R-CNN | Mask R-CNN · COCO val2017 | AP / AR (box): +0.3 mAP, +0.5 mAR | Improved |
| add USKD self-distillation during training | MobileNet-V2 · ImageNet training time | minutes per epoch: 10.17 -> 10.18 | No clear effect |
Closest papers by meaning
- Decoupled Knowledge Distillation (CVPR 2022 · similarity 0.92)
- Improving Knowledge Distillation via Regularizing Feature Direction and Norm (ECCV 2024 · similarity 0.92)
- Multi-label Self Knowledge Distillation (AAAI 2025 · similarity 0.90)
- Distilling the Knowledge in Data Pruning (ICML 2025 · similarity 0.89)
- GIFT: Unlocking Full Potential of Labels in Distilled Dataset at Near-zero Cost (ICLR 2025 · similarity 0.89)
- Logit Standardization in Knowledge Distillation (CVPR 2024 · similarity 0.89)
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.
