Research map / Label-efficient and robust learning / Recognition and long-tail
Learning to Re-weight Examples with Optimal Transport for Imbalanced Classification
Dan dan Guo, Zhuo Li, meixi zheng, He Zhao, Mingyuan Zhou, Hongyuan Zha
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 |
|---|---|---|---|
| use OT-based example re-weighting | ResNet-32 · CIFAR-LT-100 IF 200 | top-1 error: 65.30 -> 54.97 | Improved |
| use OT-based example re-weighting | ResNet-152 · Places-LT | top-1 error: 60.7 -> 60.32±0.02 | Improved |
| use OT-based example re-weighting | ResNet-50 · ImageNet-LT | top-1 error: 52.61 -> 52.36±0.01 | Improved |
| use OT-based example re-weighting | BERT base uncased · SST-2 1000:10 | accuracy: 86.50±0.38 -> 87.10±0.05 | Improved |
| combine label and feature costs rather than using either alone (ablation) | OT re-weighting · CIFAR-LT-100 IF 200 | top-1 error: 55.06 -> 54.97; 55.04 -> 54.97 | Improved |
| use all meta examples rather than prototypes (ablation) | OT re-weighting · CIFAR-LT-100 IF 200 | top-1 error: 54.97 -> 54.98 | No clear effect |
| use random meta samples rather than prototypes (ablation) | OT re-weighting · CIFAR-LT-100 IF 200 | top-1 error: 54.97 -> 55.03 | No clear effect |
| maintain and iteratively update weights rather than re-estimating from scratch (ablation) | OT re-weighting · CIFAR-LT-100 IF 200 | top-1 error: 55.07 -> 54.97 | Improved |
| use a prototype meta distribution instead of all meta examples (ablation) | OT weight net · SST-2 | accuracy: 86.71±0.24 -> 87.12±0.03; 86.49±0.60 -> 87.11±0.01; 86.47±0.51 -> 87.09±0.07; 86.24±0.60 -> 87.12±0.01 | No clear effect |
| use randomly sampled meta examples instead of prototypes (ablation) | OT weight net · SST-2 | accuracy: 87.12±0.03 -> 86.59±0.50; 87.11±0.01 -> 86.14±0.52; 87.09±0.07 -> 86.08±0.77; 87.12±0.01 -> 85.85±0.52 | Hurt |
| build the meta set dynamically from imbalanced training data instead of using an additional meta set (ablation) | OT weight net · SST-2 | accuracy: 87.08±0.09 -> 87.12±0.03; 87.13±0.04 -> 87.11±0.01; 87.14±0.08 -> 87.09±0.07; 87.10±0.05 -> 87.12±0.01 | No clear effect |
| use a weight net instead of a weight vector (ablation) | OT re-weighting · ModelNet10 10:100 | instance accuracy: 90.40±0.17 -> 89.84±0.73 | Hurt |
Closest papers by meaning
- Improving the Accuracy of Learning Example Weights for Imbalance Classification (ICLR 2022 · similarity 0.87)
- Revive Re-weighting in Imbalanced Learning by Density Ratio Estimation (NeurIPS 2024 · similarity 0.87)
- Optimal Transport Reweighting for Robust Learning under Spurious Correlations and Label Noise (NeurIPS 2026 · similarity 0.85)
- Inverse Weight-Balancing for Deep Long-Tailed Learning (AAAI 2024 · similarity 0.85)
- Gradient Reweighting: Towards Imbalanced Class-Incremental Learning (CVPR 2024 · similarity 0.85)
- Balancing the Scales: A Theoretical and Algorithmic Framework for Learning from Imbalanced Data (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.
