Research map / Label-efficient and robust learning / Noisy labels
Making Deep Neural Networks Robust to Label Noise: A Loss Correction Approach
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| apply forward loss correction using the true noise matrix instead of cross-entropy | Fully connected network · MNIST, asymmetric N=0.6 | accuracy: 53.0 ± 0.6 -> 97.3 ± 0.0 | Helped |
| use forward rather than backward correction with the true noise matrix | Fully connected network · MNIST, asymmetric N=0.6 | accuracy: 93.0 ± 0.9 -> 97.3 ± 0.0 | Helped |
| estimate the noise matrix rather than use the true matrix for forward correction (ablation) | Fully connected network · MNIST, asymmetric N=0.6 | accuracy: 97.3 ± 0.0 -> 64.9 ± 4.4 | Hurt |
| apply forward loss correction using the true noise matrix instead of cross-entropy | 14-layer ResNet · CIFAR-10, asymmetric N=0.6 | accuracy: 57.6 -> 84.5 | Helped |
| estimate the noise matrix rather than use the true matrix for forward correction (ablation) | 14-layer ResNet · CIFAR-10, asymmetric N=0.6 | accuracy: 84.5 -> 74.8 | Hurt |
| apply backward loss correction using the true noise matrix instead of cross-entropy | Word embedding · IMDB, asymmetric N=0.4 | accuracy: 58.1 ± 0.5 -> 82.1 ± 0.1 | Helped |
| estimate the noise matrix rather than use the true matrix for backward correction (ablation) | Word embedding · IMDB, asymmetric N=0.4 | accuracy: 82.1 ± 0.1 -> 77.0 ± 1.4 | Hurt |
| apply forward loss correction using the true noise matrix instead of cross-entropy | Word embedding + LSTM · IMDB, asymmetric N=0.4 | accuracy: 71.4 ± 1.3 -> 84.7 ± 0.6 | Helped |
| apply forward loss correction using the true noise matrix instead of cross-entropy | 44-layer ResNet · CIFAR-100, asymmetric N=0.6 | accuracy: 17.1 -> 68.4 | Helped |
| estimate the noise matrix rather than use the true matrix for forward correction (ablation) | 44-layer ResNet · CIFAR-100, asymmetric N=0.6 | accuracy: 68.4 -> 15.9 | Hurt |
| use forward rather than backward correction with the true noise matrix | 44-layer ResNet · CIFAR-100, asymmetric N=0.6 | accuracy: 36.8 -> 68.4 | Helped |
| fine-tune the forward-corrected model on clean 50k data (ablation) | Forward-corrected 50-ResNet · Clothing1M | accuracy: 69.84 -> 80.38 | Helped |
Closest papers by meaning
- Noise-Robust Learning from Multiple Unsupervised Sources of Inferred Labels (AAAI 2022 · similarity 0.91)
- Training Noise-Robust Deep Neural Networks via Meta-Learning (CVPR 2020 · similarity 0.90)
- Augmentation Strategies for Learning With Noisy Labels (CVPR 2021 · similarity 0.90)
- Robust Classification via Regression for Learning with Noisy Labels (ICLR 2024 · similarity 0.90)
- $\epsilon$-Softmax: Approximating One-Hot Vectors for Mitigating Label Noise (NeurIPS 2024 · similarity 0.89)
- Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 2025 · similarity 0.88)
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.