Research map / Label-efficient and robust learning / Knowledge distillation
DistPro: Searching a Fast Knowledge Distillation Process via Meta Optimization
Xueqing Deng, Dawei Sun, Shawn Newsam, Peng Wang
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 |
|---|---|---|---|
| learn searched distillation process alpha instead of equal pathway weights (ablation) | DistPro · CIFAR100 (ResNet32x4 -> ShuffleNet-v1) | top-1 accuracy (%): 73.54 -> 77.18 | Improved |
| learn searched distillation process alpha instead of equal pathway weights (ablation) | DistPro · CIFAR100 (WRN-40-2 -> ShuffleNet-v1) | top-1 accuracy (%): 74.28 -> 77.24 | Improved |
| use time-varying process instead of final alpha_T only (ablation) | DistPro · CIFAR100 (ResNet56 -> ResNet20) | top-1 accuracy (%): 71.24 -> 72.03 | Improved |
| use time-varying process instead of final alpha_T only (ablation) | DistPro · CIFAR100 (ResNet32x4 -> ShuffleNet-v2) | top-1 accuracy (%): 76.64 -> 77.54 | Improved |
| use time-varying process instead of final alpha_T only (ablation) | DistPro · CIFAR100 (WRN-40-2 -> ShuffleNet-v1) | top-1 accuracy (%): 77.19 -> 77.24 | No clear effect |
| use DistPro process instead of ReviewKD fixed pathways | ReviewKD · CIFAR100 (ResNet56 -> ResNet20) | top-1 accuracy (%): 71.89 -> 72.03 | Improved |
| use searched process instead of equal weights alongside IFVD (ablation) | DistPro · CityScapes | mIoU (%): 68.49 -> 69.12 | Improved |
| use searched process instead of equal weights (ablation) | DistPro · NYUv2 depth | RMSE: 0.2030 -> 0.1972 | Improved |
| use biased softmax normalization instead of sigmoid (ablation) | DistPro · CIFAR100 (WRN-28-4 -> WRN-16-4) | accuracy (%): 78.80 -> 79.78 | Improved |
| use biased softmax normalization instead of plain softmax (ablation) | DistPro · CIFAR100 (WRN-28-4 -> WRN-16-4) | accuracy (%): 79.41 -> 79.78 | Improved |
| use DistPro instead of ReviewKD at short training of 10 epochs | MobileNet student · ImageNet1K (ResNet-50 -> MobileNet, 10 epochs) | top-1 accuracy (%): 62.03 -> 65.25 | Improved |
| use DistPro instead of ReviewKD at 100 epochs | MobileNet student · ImageNet1K (ResNet-50 -> MobileNet, 100 epochs) | top-1 accuracy (%): decreased to 0.14% | Improved |
Closest papers by meaning
- DisWOT: Student Architecture Search for Distillation WithOut Training (CVPR 2023 · similarity 0.92)
- Direction Sensitivity–Based Knowledge Distillation: Optimization-Aware Low-Rank Knowledge Transfer (AAAI 2026 · similarity 0.91)
- KD-Zero: Evolving Knowledge Distiller for Any Teacher-Student Pairs (NeurIPS 2023 · similarity 0.91)
- Adaptive Dual Guidance Knowledge Distillation (AAAI 2025 · similarity 0.90)
- Progressively Knowledge Distillation via Re-parameterizing Diffusion Reverse Process (AAAI 2024 · similarity 0.89)
- All You Need in Knowledge Distillation Is a Tailored Coordinate System (AAAI 2025 · 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.
