Research map / Label-efficient and robust learning / Noisy labels
CARE: Class-Adaptive Expert Consensus for Reliable Learning with Long-Tailed Noisy Labels
Mengke Li, Haiquan Ling, Lihao Chen, Yang Lu, Yiqun Zhang, Hui Huang
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 CARE class-adaptive expert consensus to CLIP+LA | CLIP+LA · CIFAR-100-LTN joint noise IF=100 NR=0.5 | top-1 accuracy (%): 75.3 -> 76.7 | Improved |
| add CARE to CLIP+CE | CLIP+CE · CIFAR-100-LTN symmetric NR=0.6 IF=10 | top-1 accuracy (%): 74.5 -> 77.2 | Improved |
| use all three experts (BE, TE, IE) instead of any single auxiliary expert (ablation) | CARE · CIFAR-100-LTN IF=10 joint NR=0.5 | accuracy (%): 78.7 -> 80.3 | Improved |
| use class-wise Kc instead of global K (ablation) | CARE · ImageNet-LTNr IF=100 NR=40 | accuracy (%): 80.7 -> 81.9 | Improved |
| use class-wise Kc instead of global K on CIFAR-100-LTN (ablation) | CARE · CIFAR-100-LTN IF=10 NR=20 asymmetric | accuracy (%): 80.7 -> 80.8 | No clear effect |
| vary form of K (step, exponential, logarithmic, linear) (ablation) | CARE · CIFAR-100-LTN IF=100 20% symmetric | accuracy (%): 76.0 -> 77.5 | Improved |
| reduce BE weight to 0.5 so a single auxiliary expert can correct labels (ablation) | CARE · CIFAR100-IF10-NR50 symmetric | noise rate after correction (%): BE+TE 31.8, BE+IE 29.3 (from 50.0) | Improved |
| CARE versus observed-label expert only: tail-class accuracy (ablation) | CARE · CIFAR-100-LTN IF=10 joint NR=0.5 | tail accuracy (%): 72.3 -> 77.3 | Improved |
| CARE versus observed-label expert only: head-class accuracy (ablation) | CARE · CIFAR-100-LTN IF=10 joint NR=0.5 | head accuracy (%): 83.4 -> 81.8 | Hurt |
| class-agnostic label correction with lower noise rate but uncorrected tail labels | CLIP fine-tuning · CIFAR-100 IF=10 50% joint noise | accuracy (%): 79.5 -> 76.8 | Hurt |
| class-aware rectification (RLD 3) instead of class-agnostic | CLIP fine-tuning · CIFAR-100 IF=10 50% joint noise | accuracy (%): 79.5 -> 80.9 | Improved |
| add CARE on Clothing1M | baseline · Clothing1M | accuracy (%): 71.2 -> 71.4 | No clear effect |
Closest papers by meaning
- Mitigating Endogenous Confirmation Bias in Noisy Label Learning for Vision-Language Models (AAAI 2026 · similarity 0.88)
- From Head to Tail: Towards Balanced Representation in Large Vision-Language Models through Adaptive Data Calibration (CVPR 2025 · similarity 0.86)
- Robust Visual Recognition with Class-Imbalanced Open-World Noisy Data (AAAI 2024 · similarity 0.86)
- Three Heads Are Better than One: Complementary Experts for Long-Tailed Semi-supervised Learning (AAAI 2024 · similarity 0.86)
- Retraining with Predicted Hard Labels Provably Increases Model Accuracy (ICML 2025 · similarity 0.86)
- Vision-Language Models are Strong Noisy Label Detectors (NeurIPS 2024 · similarity 0.86)
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.
