Research map / Label-efficient and robust learning / Knowledge distillation
How to Train the Teacher Model for Effective Knowledge Distillation
Shayan Mohajer Hamidi, Xizhen Deng, Renhao Tan, Linfeng Ye, Ahmed Hussein Salamah
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 |
|---|---|---|---|
| train the teacher with MSE loss instead of cross-entropy (ablation) | KD methods (AT, PKT, SP, CC, RKD, VID, CRD, REVIEW, DKD, HSAKD) · CIFAR-100 | student top-1 accuracy: up to +2.67 | Improved |
| train the teacher with MSE loss instead of cross-entropy (ablation) | KD, REVIEW+KD · ImageNet (ResNet34-ResNet18) | student top-1 accuracy: 71.03 -> 71.58 (KD) | Improved |
| train the teacher with MSE loss instead of cross-entropy (ablation) | teacher network · CIFAR-100 and ImageNet | teacher accuracy: 79.34 -> 74.54 (ResNet-50 CIFAR-100) | Hurt |
| train the teacher with MSE loss with feature-based methods combined with KD (ablation) | AT+KD, PKT+KD, etc. · CIFAR-100 | student top-1 accuracy: +2.24 (AT+KD, ResNet-32x4 -> ShuffleNetV2) | Improved |
| train the teacher with MSE loss for CC+KD (ablation) | CC+KD · CIFAR-100 (ResNet-50 -> VGG-8) | student top-1 accuracy: 73.48 -> 70.79 | Hurt |
| use MSE teacher in different-architecture teacher-student pairs (ablation) | KD methods · CIFAR-100 | student top-1 accuracy | Improved |
| train the teacher with MSE loss in binary classification KD (ablation) | KD, AT, CC, VID · binary CIFAR variants | student accuracy: 71.40 -> 71.23 (KD, CIFAR-50-74) | Mixed |
| use MSE teacher in semi-supervised distillation (ablation) | ResNet18 student · CIFAR-10 with 1-8% labels | student accuracy | Improved |
Closest papers by meaning
- Can Students Beyond the Teacher? Distilling Knowledge from Teacher’s Bias (AAAI 2025 · similarity 0.90)
- Bayesian Knowledge Distillation: A Bayesian Perspective of Distillation with Uncertainty Quantification (ICML 2024 · similarity 0.90)
- Bayes Conditional Distribution Estimation for Knowledge Distillation Based on Conditional Mutual Information (ICLR 2024 · similarity 0.89)
- Knowledge Distillation with Auxiliary Variable (ICML 2024 · similarity 0.89)
- Toward Student-oriented Teacher Network Training for Knowledge Distillation (ICLR 2024 · similarity 0.89)
- Speculative Knowledge Distillation: Bridging the Teacher-Student Gap Through Interleaved Sampling (ICLR 2025 · similarity 0.88)
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.
