Research map / Label-efficient and robust learning / Knowledge distillation
What to Distill? Fast Knowledge Distillation with Adaptive Sampling
Byungchul Chae, Seonyeong Heo
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 only on high teacher-student KL samples (quantity of knowledge) (ablation) | KD · CIFAR-100 VGG13->VGG8 | top-1 accuracy: 73.14 -> 73.76 (40% sampling vs 100%) | Improved |
| select samples by teacher-student KL rather than KD loss (ablation) | KD · CIFAR-100 VGG13->VGG8 | top-1 accuracy: 73.28 -> 73.67 | Improved |
| add quantity-based subsampling (ablation) | KD VGG13->VGG8 · CIFAR-100 | top-1 accuracy: 73.14 -> 73.82 | Improved |
| add curriculum sampling schedule (ablation) | KDAS · CIFAR-100 | top-1 accuracy: 73.82 -> 74.08 | Improved |
| add quality-based penalization (ablation) | KDAS · CIFAR-100 | top-1 accuracy: 74.08 -> 74.12 | Improved |
| add optimization techniques for efficiency (ablation) | KDAS · CIFAR-100 | top-1 accuracy: 74.12 -> 73.91; training time -30.22 % | Mixed |
| apply KDAS to vanilla KD on ImageNet ResNet34->ResNet18 | KD · ImageNet val | top-1 accuracy: 71.03 -> 71.35; training time -36.23% | Improved |
| apply KDAS to LogitSTD on ImageNet | LogitSTD · ImageNet val | top-1 accuracy: +0.06 and +0.03 | No clear effect |
| apply KDAS to CNN teacher with ViT students | KD · CIFAR-100 ResNet56 teacher | top-1 accuracy: 73.25 -> 77.81 (DeiT-Ti) | Improved |
Closest papers by meaning
- What Makes a Good Dataset for Knowledge Distillation? (CVPR 2025 · similarity 0.90)
- Can Students Beyond the Teacher? Distilling Knowledge from Teacher’s Bias (AAAI 2025 · similarity 0.89)
- Knowledge Distillation with Auxiliary Variable (ICML 2024 · similarity 0.89)
- Small Scale Data-Free Knowledge Distillation (CVPR 2024 · similarity 0.89)
- Why Knowledge Distillation Works in Generative Models: A Minimal Working Explanation (NeurIPS 2025 · similarity 0.88)
- Debiased Distillation for Consistency Regularization (AAAI 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.
