Research map / Label-efficient and robust learning / Knowledge distillation
Respecting Transfer Gap in Knowledge Distillation
Read the paperCheck 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 |
|---|---|---|---|
| apply IPWD reweighting to KD (two-stage) | KD · CIFAR-100, resnet32x4->resnet8x4 | top-1 accuracy (%): 73.33 -> 76.03 | Helped |
| apply IPWD reweighting to KD (two-stage) | KD · CIFAR-100, ResNet50->MobileNetV2 | top-1 accuracy (%): 67.35 -> 70.25 | Helped |
| apply IPWD to KD | KD · ImageNet, ResNet-50->MobileNet-v1 | top-1 accuracy (%): 70.49 -> 72.65 | Helped |
| apply IPWD to KD | KD · ImageNet, ResNet-34->ResNet-18 | top-1 accuracy (%): 70.67 -> 71.88 | Helped |
| add IPWD as plug-in to SSKD | SSKD · CIFAR-100, WRN-40-2->ShuffleNetV1 | top-1 accuracy (%): 78.49 -> 79.17 | Helped |
| add IPWD to PS-KD (late stage) | PS-KD · CIFAR-100, ResNet-101 self-distillation | top-1 accuracy (%): 80.57 -> 81.39 | Helped |
| add IPWD to PS-KD (late stage) | PS-KD · CIFAR-100, ResNeXt-29 | ECE (%): 9.15 -> 4.93 | Helped |
| remove CLS head and logit normalization from weight estimation (ablation) | IPWD · CIFAR-100, resnet32x4->ShuffleNetV1 | top-1 accuracy (%): 76.03 -> 52.81 | Hurt |
| remove CLS head, keep logit normalization (ablation) | IPWD · CIFAR-100, resnet32x4->ShuffleNetV1 | top-1 accuracy (%): 76.03 -> 75.49 | Hurt |
| remove logit normalization, keep CLS head (ablation) | IPWD · CIFAR-100, WRN-40-2->ShuffleNetV1 | top-1 accuracy (%): 76.61 -> 75.96 | Hurt |
| start IPWD from the beginning of training instead of last 75 epochs (ablation) | PS-KD + IPWD · CIFAR-100, ResNeXt-29 | top-1 accuracy (%): 83.30 -> 82.86 | Hurt |
| train teacher with label smoothing | IPWD vs KD · CIFAR-100, ResNet50->MobileNetV2 | top-1 accuracy (%): 66.79 vs 66.23 | Mixed |
Closest papers by meaning
- Knowledge Distillation with Auxiliary Variable (ICML 2024 · similarity 0.91)
- Can Students Beyond the Teacher? Distilling Knowledge from Teacher’s Bias (AAAI 2025 · similarity 0.91)
- Speculative Knowledge Distillation: Bridging the Teacher-Student Gap Through Interleaved Sampling (ICLR 2025 · similarity 0.90)
- Cooperative Knowledge Distillation: A Learner Agnostic Approach (AAAI 2024 · similarity 0.89)
- ABKD: Pursuing a Proper Allocation of the Probability Mass in Knowledge Distillation via $\alpha$-$\beta$-Divergence (ICML 2025 · similarity 0.89)
- Revisit the Essence of Distilling Knowledge through Calibration (ICML 2024 · similarity 0.89)
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.