Research map / Label-efficient and robust learning / Recognition and long-tail
Identifying Hard Noise in Long-Tailed Sample Distribution
Xuanyu Yi, Kaihua Tang, Xian-Sheng Hua, Joo-Hwee Lim, Hanwang Zhang
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 environment-invariant hard noise identification (H2E) vs MentorMix-RS | H2E · ImageNet-NLT red 10% noise | Top-1 accuracy (%): 62.20 -> 64.86 | Improved |
| use H2E instead of MentorMix-RS | H2E · ImageNet-NLT blue 30% noise | Top-1 accuracy (%): 48.27 -> 51.52 | Improved |
| iterate hard-to-easy transformation (H2E-iter) | H2E · ImageNet-NLT red 30% noise | Top-1 accuracy (%): 58.38 -> 60.66 | Improved |
| add augmentation to construct environments (3 environments) (ablation) | H2E · ImageNet-NLT red 10% noise | Top-1 accuracy (%): 62.79 -> 64.86 | Improved |
| increase number of environments from 2 to 3 (ablation) | H2E · ImageNet-NLT red 10% noise (with augmentation) | Top-1 accuracy (%): 62.15 -> 64.86 | Improved |
| increase number of environments from 3 to 4 (ablation) | H2E · ImageNet-NLT red 10% noise (with augmentation) | Top-1 accuracy (%): 64.86 -> 65.38 | No clear effect |
| use H2E as stage 2 instead of CE+RW after CF stage 1 (ablation) | H2E · ImageNet-NLT red 10% noise | Top-1 accuracy (%): 56.54 -> 63.90 | Improved |
| use H2E stage 1 instead of LDAM before H2E stage 2 (ablation) | H2E · ImageNet-NLT red 10% noise | Top-1 accuracy (%): 61.94 -> 64.86 | Improved |
| use H2E stage 2 instead of ERM+RW after H2E stage 1 (ablation) | H2E · ImageNet-NLT red 10% noise | Top-1 accuracy (%): 60.20 -> 64.86 | Improved |
| use H2E instead of LDAM on Animal10-NLT | H2E · Animal10-NLT (eta=100) | Top-1 accuracy (%): 68.21 -> 66.58 | No clear effect |
| use H2E instead of BBN on Food101-NLT | H2E · Food101-NLT (eta=100) | Top-1 accuracy (%): 53.16 -> 58.66 | Improved |
| use red (realistic) instead of blue (synthetic) noise | CE · ImageNet-NLT 10% noise | Top-1 accuracy (%): 45.80 -> 54.36 | Improved |
Closest papers by meaning
- Robust Logit Adjustment for Learning with Long-Tailed Noisy Data (AAAI 2025 · similarity 0.87)
- EAT: Towards Long-Tailed Out-of-Distribution Detection (AAAI 2024 · similarity 0.86)
- Rethinking Benign Overfitting in Two-Layer Neural Networks (ICML 2025 · similarity 0.86)
- Out-of-Distribution Detection in Long-Tailed Recognition with Calibrated Outlier Class Learning (AAAI 2024 · similarity 0.86)
- Long-Tailed Out-of-Distribution Detection: Prioritizing Attention to Tail (AAAI 2025 · similarity 0.86)
- TailedCore: Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection (CVPR 2025 · similarity 0.85)
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.
