Research map / Label-efficient and robust learning / Recognition and long-tail
Rethinking Importance Weighting for Deep Learning under Distribution Shift
Tongtong Fang, Nan Lu, Gang Niu, Masashi Sugiyama
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 |
|---|---|---|---|
| add a feature extractor to IW (SIW-L vs IW) (ablation) | importance weighting · CIFAR-10 0.4 symmetric noise | accuracy (%): 44.31 -> 76.39 | Improved |
| make weights dynamic (DIW1-L vs SIW-L) (ablation) | static IW with FE · CIFAR-10 0.4 symmetric noise | accuracy (%): 76.39 -> 75.27 | Mixed |
| make weights dynamic (DIW1-F vs SIW-F) (ablation) | static IW with FE · CIFAR-10 0.4 symmetric noise | accuracy (%): 65.58 -> 78.23 | Improved |
| update the feature extractor during training (DIW2-L vs DIW1-L) (ablation) | DIW · CIFAR-10 0.4 symmetric noise | accuracy (%): 75.27 -> 80.40 | Improved |
| pretrain FE for -F variant (DIW3-F vs DIW2-F) (ablation) | DIW-F · CIFAR-10 0.4 symmetric noise | accuracy (%): 77.48 -> 78.75 | Improved |
| pretrain FE for -L variant (DIW3-L vs DIW2-L) (ablation) | DIW-L · CIFAR-10 0.4 symmetric noise | accuracy (%): 80.40 -> 80.10 | No clear effect |
| DIW vs Reweight under label noise | DIW2-L · CIFAR-100 0.4 symmetric noise | accuracy (%): 42.15 -> 53.66 | Improved |
| DIW vs best baseline under class-prior shift rho=100 | DIW · Fashion-MNIST | accuracy (%): 83.48 -> 84.02 | Improved |
| DIW vs IW weight estimation under class-prior shift rho=100 | DIW · Fashion-MNIST | MAE to true weights: 1.10 -> 0.45 | Improved |
| DIW vs Reweight under pair-flip noise 0.3 | DIW · Fashion-MNIST | accuracy (%): 88.74 -> 88.19 | Hurt |
Closest papers by meaning
- Is Importance Weighting Incompatible with Interpolating Classifiers? (ICLR 2022 · similarity 0.88)
- Optimizing importance weighting in the presence of sub-population shifts (ICLR 2025 · similarity 0.88)
- Understanding the role of importance weighting for deep learning (ICLR 2021 · similarity 0.87)
- Revive Re-weighting in Imbalanced Learning by Density Ratio Estimation (NeurIPS 2024 · similarity 0.84)
- Inverse Weight-Balancing for Deep Long-Tailed Learning (AAAI 2024 · similarity 0.82)
- Gradient-Weight Alignment as a Train-Time Proxy for Generalization in Classification Tasks (NeurIPS 2025 · similarity 0.82)
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.
