Research map / Label-efficient and robust learning / Noisy labels
Detecting Corrupted Labels Without Training a Model to Predict
Zhaowei Zhu, Zihao Dong, Yang Liu
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 |
|---|---|---|---|
| detect corrupted labels using feature-neighborhood ranking instead of training on noisy labels | SimiFeat-R vs. training-based methods · CIFAR-10 and CIFAR-100, synthetic and human noise | F1-score (%) | Improved |
| use local neighborhood voting rather than training-based detection | SimiFeat-V vs. training-based methods · CIFAR-10 and CIFAR-100, synthetic and human noise | F1-score (%) | Improved |
| directly use extracted features instead of fine-tuned model logits | SimiFeat vs. methods using the same CLIP extractor · CIFAR-10 and CIFAR-100 | F1-score (%) | Mixed |
| add one-epoch domain-specific warmup to the feature extractor (ablation) | SimiFeat on Clothing1M · Clothing1M | best-epoch test accuracy: 72.37 -> 73.64 | Improved |
| add balanced sampling during training on selected data (ablation) | SimiFeat-selected Clothing1M training set · Clothing1M | best-epoch test accuracy: 73.64 -> 73.97 | Improved |
| filter noisy training instances before training (ablation) | Clothing1M standard training · Clothing1M | best-epoch test accuracy: 70.32 -> 72.37 | Improved |
| use CLIP features rather than ImageNet ResNet50 features for selection (ablation) | SimiFeat on Clothing1M · Clothing1M | best-epoch test accuracy: 72.37 -> 72.54 | Improved |
| use feature representations directly instead of a noisy-label-trained logit layer (ablation) | SimiFeat-V vs. deep k-NN · CIFAR-10 and CIFAR-100 | F1-score (%) | Mixed |
Closest papers by meaning
- Using Trusted Data to Train Deep Networks on Labels Corrupted by Severe Noise (NeurIPS 2018 · similarity 0.91)
- Distilling Effective Supervision From Severe Label Noise (CVPR 2020 · similarity 0.90)
- Understanding and Utilizing Deep Neural Networks Trained with Noisy Labels (ICML 2019 · similarity 0.89)
- Delving into Noisy Label Detection with Clean Data (ICML 2023 · similarity 0.89)
- Learning Discriminative Dynamics with Label Corruption for Noisy Label Detection (CVPR 2024 · similarity 0.87)
- Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 2025 · similarity 0.87)
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.
