Research map / Label-efficient and robust learning / Noisy labels
Normalized Loss Functions for Deep Learning with Noisy Labels
Read the paperCheck 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 |
|---|---|---|---|
| Combine normalized active and robust passive losses as NCE+RCE instead of using the strongest reported baseline | CIFAR-100 · symmetric noise η=0.8 | test accuracy: 16.18±0.46 -> 25.80±1.12 | Helped |
| Combine normalized active and robust passive losses as NFL+RCE instead of NLNL | CIFAR-10 · symmetric noise η=0.8 | test accuracy: 51.41±0.85 -> 55.06±1.08 | Helped |
| Combine NCE with MAE rather than SCE | CIFAR-100 · asymmetric noise η=0.4 | test accuracy: 42.19±0.19 -> 47.22±0.30 | Helped |
| Normalize CE to NCE (ablation) | CE on noisy-label classification · CIFAR-10 and CIFAR-100, symmetric noise η=0.6 | overfitting | Helped |
| Normalize FL to NFL (ablation) | FL on noisy-label classification · CIFAR-10 and CIFAR-100, symmetric noise η=0.6 | overfitting | Helped |
| Use normalized robust losses alone rather than an active-plus-passive combination (ablation) | NCE, NFL, MAE, and RCE · CIFAR-100, symmetric noise η=0.6 | learning sufficiency | Hurt |
| Scale normalized losses upward rather than combining active and passive losses (ablation) | NCE, NFL, MAE, and RCE · CIFAR-100, symmetric noise η=0.6 | learning sufficiency | No clear effect |
| Combine one active loss with one passive loss rather than two active or two passive losses (ablation) | APL · CIFAR-10 and CIFAR-100, symmetric noise η=0.6 | test accuracy | Helped |
| Add MAE to normalized GCE (ablation) | CIFAR-10 · symmetric noise η=0.8 | test accuracy: 29.08±0.80 -> 50.55±1.08 | Helped |
| Add RCE to normalized GCE (ablation) | CIFAR-10 · asymmetric noise η=0.4 | test accuracy: 74.98±0.32 -> 71.65±0.68 | Hurt |
| Add MAE to normalized GCE (ablation) | CIFAR-10 · symmetric noise η=0.4 | test accuracy: 83.33±0.39 -> 84.14±0.15 | Helped |
Closest papers by meaning
- Asymmetric Loss Functions for Learning with Noisy Labels (ICML 2021 · similarity 0.94)
- Active Negative Loss Functions for Learning with Noisy Labels (NeurIPS 2023 · similarity 0.94)
- Robust Loss Functions under Label Noise for Deep Neural Networks (AAAI 2017 · similarity 0.93)
- Learning from Noisy Labels with Complementary Loss Functions (AAAI 2021 · similarity 0.92)
- Enhancing Noise-Robust Losses for Large-Scale Noisy Data Learning (AAAI 2025 · similarity 0.91)
- $\epsilon$-Softmax: Approximating One-Hot Vectors for Mitigating Label Noise (NeurIPS 2024 · similarity 0.89)
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.