Research map / Label-efficient and robust learning / Knowledge distillation
Rethinking the Dark Knowledge and Kullback-Leibler Divergence Loss in Knowledge Distillation Under Capacity Mismatching
Yingchao Wang, Wenqi Niu, Xingshan Yao, Li You, Weilun Fei
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 |
|---|---|---|---|
| use cosine similarity instead of MSE on the relative confidence matrix (ablation) | RCKD · CIFAR-100 (ResNet-32x4 -> ResNet-8x4) | top-1 accuracy: 75.56 -> 76.89 | Improved |
| use cosine similarity instead of MSE on the relative confidence matrix (ablation) | RCKD · CIFAR-100 (WRN-40-2 -> ShuffleNet-V1) | top-1 accuracy: 76.13 -> 76.97 | Improved |
| use cosine similarity instead of MSE on the relative confidence matrix (ablation) | RCKD · CIFAR-100 (VGG-13 -> MobileNet-V2) | top-1 accuracy: 68.33 -> 69.77 | Improved |
| replace KL divergence distillation with RCKD | KD · CIFAR-100 (ResNet-32x4 -> ResNet-8x4) | top-1 accuracy: 73.33 -> 76.89 | Improved |
| replace KL divergence distillation with RCKD | KD · ImageNet (ResNet-34 -> ResNet-18) | top-1 accuracy: 71.03 -> 72.02 | Improved |
| replace KL divergence distillation with RCKD | KD · CIFAR-100 (VGG-13 -> MobileNet-V2) | top-1 accuracy: 67.37 -> 69.77 | Improved |
| use RCKD instead of the best logit-based baseline | logit-based KD · CIFAR-100 (ResNet-50 -> MobileNet-V2) | top-1 accuracy: 70.35 -> 69.67 | Hurt |
| use RCKD instead of ReviewKD feature-based distillation | feature-based KD · ImageNet (ResNet-50 -> MobileNet-V1) | top-1 accuracy: 72.56 -> 72.41 | Hurt |
| vary the RCKD weight beta (too small or too large) (ablation) | RCKD · CIFAR-100 | top-1 accuracy | Mixed |
Closest papers by meaning
- ABKD: Pursuing a Proper Allocation of the Probability Mass in Knowledge Distillation via $\alpha$-$\beta$-Divergence (ICML 2025 · similarity 0.92)
- Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models (NeurIPS 2026 · similarity 0.91)
- PLD: A Choice-Theoretic List-Wise Knowledge Distillation (NeurIPS 2025 · similarity 0.91)
- All You Need in Knowledge Distillation Is a Tailored Coordinate System (AAAI 2025 · similarity 0.91)
- Revisit the Essence of Distilling Knowledge through Calibration (ICML 2024 · similarity 0.90)
- Maintaining Fairness in Logit-based Knowledge Distillation for Class-Incremental Learning (AAAI 2025 · similarity 0.90)
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.
