Research map / Label-efficient and robust learning / Knowledge distillation
Knowledge Distillation: A Good Teacher Is Patient and Consistent
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Give teacher and student identical image views and use mixup function matching instead of fixed teacher targets (ablation) | ResNet-50 distillation from BiT-M-R152x2 · ImageNet | top-1 accuracy: 76.95 -> 82.31 | Helped |
| Train for longer, from 1200 to 9600 epochs (ablation) | FunMatch ResNet-50 · ImageNet | top-1 accuracy: 81.54 -> 82.31 | Helped |
| Replace Adam with Shampoo (ablation) | ResNet-50 distillation · ImageNet | top-1 accuracy: 81.54 -> 81.67 | Helped |
| Initialize the student with pretrained BiT-M-ResNet50 weights (ablation) | FunMatch ResNet-50 · ImageNet | top-1 accuracy: 81.11; 81.45 | Mixed |
| Use a 384px teacher and 224px student rather than 224px for both (ablation) | FunMatch ResNet-50 · ImageNet | top-1 accuracy: 82.31 -> 82.64 | Helped |
| Use a 384px and 224px ensemble teacher rather than the default teacher (ablation) | FunMatch ResNet-50 · ImageNet | top-1 accuracy: 82.64 -> 82.82 | Helped |
| Reduce student input resolution from 224px to 160px (ablation) | FunMatch ResNet-50 · ImageNet | top-1 accuracy: 82.31 -> 80.49 | Hurt |
| Replace ResNet-50 student with MobileNet v3 Large using GroupNorm (ablation) | Function matching distillation · ImageNet | top-1 accuracy: 82.31 -> 76.97 | Hurt |
| Replace MobileNet v3 GroupNorm with BatchNorm (ablation) | FunMatch MobileNet v3 · ImageNet | top-1 accuracy: 76.31 -> 73.51 | Hurt |
| Use the Small rather than Large MobileNet v3 variant (ablation) | FunMatch MobileNet v3 with GroupNorm · ImageNet | top-1 accuracy: 76.31 -> 67.57 | Hurt |
| Use two teachers for MobileNet v3 with GroupNorm (ablation) | FunMatch MobileNet v3 · ImageNet | top-1 accuracy: 74.60 -> 74.85 | Helped |
| Train with hard labels and omit the distillation loss (ablation) | ResNet-50 with the distillation training setup · ImageNet | top-1 accuracy: 78.15 -> 73.48 | Hurt |
Closest papers by meaning
- CustomKD: Customizing Large Vision Foundation for Edge Model Improvement via Knowledge Distillation (CVPR 2025 · similarity 0.89)
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 2024 · similarity 0.88)
- DFD: Distilling the Feature Disparity Differently for Detectors (ICML 2024 · similarity 0.88)
- Understanding the Role of the Projector in Knowledge Distillation (AAAI 2024 · similarity 0.88)
- VkD: Improving Knowledge Distillation using Orthogonal Projections (CVPR 2024 · similarity 0.88)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.87)
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.