Research map / Label-efficient and robust learning / Knowledge distillation
Knowledge Distillation As Efficient Pre-Training: Faster Convergence, Higher Data-Efficiency, and Better Transferability
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 |
|---|---|---|---|
| Use KDEP (SVD+PTS) instead of supervised pre-training with the same 10% data and 90 epochs | R18 pre-training · Image classification average | top-1 accuracy: 71.05 -> 75.74 | Helped |
| Use KDEP (SVD+PTS) instead of supervised pre-training with the same 100% data and 90 epochs | R18 pre-training · Image classification average | top-1 accuracy: 77.13 -> 78.40 | Helped |
| Use KDEP (SVD+PTS) instead of supervised pre-training with the same 10% data and 90 epochs | R18 pre-training · Semantic segmentation average | mIoU: 57.41 -> 59.61 | Helped |
| Use KDEP (SVD+PTS) instead of supervised pre-training with the same 10% data and 90 epochs | R18 pre-training · COCO detection | AP: 27.6 -> 28.6 | Helped |
| Add Power Temperature Scaling after SVD (ablation) | KDEP feature alignment, R50 → R18 · Image classification average, 10% ImageNet-1K, 90 epochs | top-1 accuracy: 75.09 -> 75.74 | Helped |
| Add Scale Normalization after SVD (ablation) | KDEP feature alignment, R50 → R18 · Image classification average, 10% ImageNet-1K, 90 epochs | top-1 accuracy: 75.09 -> 75.34 | Helped |
| Add Std Matching after SVD (ablation) | KDEP feature alignment, R50 → R18 · Image classification average, 10% ImageNet-1K, 90 epochs | top-1 accuracy: 75.09 -> 75.87 | Helped |
| Use PTS with T=0.1 instead of no PTS (ablation) | KDEP, R50 → R18 · Image classification average, 10% ImageNet-1K, 90 epochs | top-1 accuracy: 75.09 -> 75.74 | Helped |
| Add multiple-layer feature distillation (ablation) | KDEP, MS R50 → R18 · Image classification average, 10% ImageNet-1K, 90 epochs | top-1 accuracy: 75.74 -> 76.11 | Helped |
| Add multiple-layer feature distillation (ablation) | KDEP, MS R50 → MNV2 · Image classification average, 10% ImageNet-1K, 90 epochs | top-1 accuracy: 75.37 -> 75.20 | Hurt |
| Replace SVD dimension alignment with 1×1 convolution alignment (ablation) | KDEP, MS R50 → R18 · Image classification average, 10% ImageNet-1K, 90 epochs | top-1 accuracy: 75.74 -> 73.10 | Hurt |
| Use logits KD instead of feature-based KDEP (ablation) | R50 → R18 · Image classification average, 10% ImageNet-1K, 90 epochs | top-1 accuracy: 71.95 vs 74.79 | Hurt |
Closest papers by meaning
- VkD: Improving Knowledge Distillation using Orthogonal Projections (CVPR 2024 · similarity 0.89)
- Dataset Distillation via Knowledge Distillation: Towards Efficient Self-Supervised Pre-training of Deep Networks (ICLR 2025 · similarity 0.89)
- Small Scale Data-Free Knowledge Distillation (CVPR 2024 · similarity 0.89)
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 2024 · similarity 0.89)
- All You Need in Knowledge Distillation Is a Tailored Coordinate System (AAAI 2025 · similarity 0.88)
- Generative Model-Based Feature Knowledge Distillation for Action Recognition (AAAI 2024 · 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.