Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Sample Difficulty from Pre-trained Models for Reliable Prediction
Peng Cui, Dan Zhang, Zhijie Deng, Yinpeng Dong, Jun Zhu
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 |
|---|---|---|---|
| replace constant-weight entropy regularization with sample difficulty-aware weighting (ablation) | ER · ImageNet1k | ECE (%): 3.720 -> 1.602 | Improved |
| replace constant-weight entropy regularization with sample difficulty-aware weighting (ablation) | ER · ImageNet1k | Top-1 accuracy (%): 73.68 -> 74.11 | Improved |
| add difficulty-aware weighting to cross entropy | CE · CIFAR-100 | ECE (%): 7.251 -> 3.410 | Improved |
| PolyLoss baseline vs proposed | PolyLoss · ImageNet1k | ECE (%): 7.882 -> 1.602 | Improved |
| use CLIP-ViT-B instead of ViT-B as difficulty scorer (ablation) | Proposed · ImageNet1k | ECE (%): 2.701 -> 1.602 | Improved |
| use CLIP-ViT-B instead of MAE-ViT-B as scorer (ablation) | Proposed · ImageNet1k | ECE (%): 1.925 -> 1.602 | Improved |
| use CLIP-ViT-B instead of DINOv2 as scorer (ablation) | Proposed · ImageNet1k | Top-1 accuracy (%): 74.01 -> 74.11 | Improved |
| use RMD instead of MD for difficulty (ablation) | Proposed · ImageNet1k | ECE (%): 2.125 -> 1.613 | Improved |
| use RMD instead of K-means distance (ablation) | Proposed · ImageNet1k | Top-1 accuracy (%): 73.78 -> 74.11 | Improved |
| use RMD instead of correctness ranking loss | CRL · ImageNet1k | ECE (%): 2.272 -> 1.602 | Improved |
| use RMD instead of CRL on CIFAR-10 | CRL · CIFAR-10 | ECE (%): 0.957 -> 1.212 | Hurt |
| increase alpha from 0.05 to 0.30 in entropy regularization (constant weight) (ablation) | ER · ImageNet1k | ECE (%): 3.308 -> 33.26 | Hurt |
Closest papers by meaning
- SURE: SUrvey REcipes for building reliable and robust deep networks (CVPR 2024 · similarity 0.86)
- MaxSup: Overcoming Representation Collapse in Label Smoothing (NeurIPS 2025 · similarity 0.85)
- Exploring Learning Complexity for Efficient Downstream Dataset Pruning (ICLR 2025 · similarity 0.85)
- Training Deep Neural Networks with Virtual Smoothing Classes (AAAI 2025 · similarity 0.85)
- Provable Benefits of Unsupervised Pre-training and Transfer Learning via Single-Index Models (ICML 2025 · similarity 0.84)
- Lightweight Dataset Pruning without Full Training via Example Difficulty and Prediction Uncertainty (ICML 2025 · similarity 0.84)
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.
