Research map / Label-efficient and robust learning / Knowledge distillation
Knowledge Distillation with Refined Logits
Wujie Sun, Defang Chen, Siwei Lyu, Genlang Chen, Chun Chen, Can Wang
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 sample confidence distillation to CE (ablation) | RLD · CIFAR-100 ResNet32x4 -> ResNet8x4 | top-1 accuracy (%): 72.50 -> 73.55 | Improved |
| add masked correlation (M_ge) to CE (ablation) | RLD · CIFAR-100 ResNet32x4 -> ResNet8x4 | top-1 accuracy (%): 72.50 -> 75.64 | Improved |
| add masked correlation (M_g) to CE (ablation) | RLD · CIFAR-100 ResNet32x4 -> ResNet8x4 | top-1 accuracy (%): 72.50 -> 75.50 | Improved |
| add sample confidence to masked correlation with M_ge (ablation) | RLD · CIFAR-100 ResNet32x4 -> ResNet8x4 | top-1 accuracy (%): 75.64 -> 76.64 | Improved |
| add sample confidence to masked correlation with M_g (ablation) | RLD · CIFAR-100 ResNet32x4 -> ResNet8x4 | top-1 accuracy (%): 75.50 -> 75.53 | No clear effect |
| mask classes with M_ge instead of M_g in full RLD (ablation) | RLD · CIFAR-100 ResNet32x4 -> ResNet8x4 | top-1 accuracy (%): 75.53 -> 76.64 | Improved |
| use RLD instead of DKD with inferior teacher ResNet56 | RLD · CIFAR-100 ResNet56 -> WRN-40-2 | top-1 accuracy (%): 77.34 -> 78.03 | Improved |
| use RLD instead of DKD with inferior teacher VGG13 | RLD · CIFAR-100 VGG13 -> WRN-40-2 | top-1 accuracy (%): 77.45 -> 77.88 | Improved |
| use RLD instead of DKD with LSKD logit standardization | RLD · CIFAR-100 ResNet50 -> MN-V2 | top-1 accuracy (%): 70.45 -> 71.06 | Improved |
| use RLD instead of DKD | RLD · ImageNet ResNet50 -> MN-V1 | top-1 accuracy (%): 72.05 -> 72.75 | Improved |
| use RLD instead of DKD | RLD · CIFAR-100 ResNet110 -> ResNet32 | top-1 accuracy (%): 74.11 -> 74.02 | Hurt |
| use RLD instead of DKD | RLD · CIFAR-100 ResNet110 -> ResNet20 | top-1 accuracy (%): 71.06 -> 71.67 | Improved |
Closest papers by meaning
- Logit Standardization in Knowledge Distillation (CVPR 2024 · similarity 0.91)
- PLD: A Choice-Theoretic List-Wise Knowledge Distillation (NeurIPS 2025 · similarity 0.91)
- Maintaining Fairness in Logit-based Knowledge Distillation for Class-Incremental Learning (AAAI 2025 · similarity 0.91)
- Enhancing Logits Distillation with Plug&Play Kendall's $\tau$ Ranking Loss (ICML 2025 · similarity 0.90)
- Scaled Decoupled Distillation (CVPR 2024 · similarity 0.90)
- Can Students Beyond the Teacher? Distilling Knowledge from Teacher’s Bias (AAAI 2025 · 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.
