Research map / Label-efficient and robust learning / Noisy labels
Searching for Robustness: Loss Learning for Noisy Classification Tasks
Boyan Gao, Henry Gouk, Timothy M. Hospedales
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 |
|---|---|---|---|
| replace cross-entropy with the learned Taylor loss (DR) | general-purpose loss · Average rank, 80% symmetric noise (10 arch/dataset pairs, general loss) | average rank: 6.4 -> 2.0 | Improved |
| replace cross-entropy with the learned Taylor loss (AR) | general-purpose loss · Average rank, 40% asymmetric noise (10 arch/dataset pairs, general loss) | average rank: 5.0 -> 1.6 | Improved |
| dataset randomization instead of architecture randomization during meta-learning (ablation) | Taylor loss · Average rank, 80% symmetric noise (10 arch/dataset pairs, general loss) | average rank: 2.4 -> 2.0 | Mixed |
| dataset randomization instead of architecture randomization during meta-learning (ablation) | Taylor loss · Average rank, 40% asymmetric noise (10 arch/dataset pairs, general loss) | average rank: 1.6 -> 3.0 | Mixed |
| apply learned Taylor loss instead of CE on real noisy labels | ResNet-18 · Clothing1M | test accuracy: 66.88 -> 70.09 | Improved |
| learned loss trained under 40% asymmetric vs 80% symmetric noise, transferred to Clothing1M | Taylor loss · Clothing1M | test accuracy: 68.85 -> 70.09 | Mixed |
| plug-in Taylor loss vs multi-network JoCoR | ResNet-18 · Clothing1M | test accuracy: 69.79 -> 70.09 | Improved |
| train the loss specifically for the target dataset using a clean subset | Taylor loss · MNIST Sym-80% (JoCoR net) | accuracy: 84.89 -> 89.88 | Improved |
| train the loss specifically for the target dataset using a clean subset | Taylor loss · CIFAR-100 Sym-80% (JoCoR net) | accuracy: 15.49 -> 20.20 | Improved |
| use a different Taylor polynomial order (2, 3, 4 or 5) (ablation) | Taylor loss · MNIST/KMNIST/CIFAR-10 linear classifier | accuracy | No clear effect |
| apply loss trained at 80% symmetric noise to other noise levels | Taylor loss · 2MLP-MNIST, 4CNN-MNIST, VGG11-CIFAR10 | accuracy | Improved |
Closest papers by meaning
- Robust Loss Functions for Training Decision Trees with Noisy Labels (AAAI 2024 · similarity 0.89)
- Peer Loss Functions: Learning from Noisy Labels without Knowing Noise Rates (ICML 2020 · similarity 0.89)
- Robust Loss Functions under Label Noise for Deep Neural Networks (AAAI 2017 · similarity 0.89)
- Variation-Bounded Loss for Noise-Tolerant Learning (AAAI 2026 · similarity 0.88)
- Enhancing Noise-Robust Losses for Large-Scale Noisy Data Learning (AAAI 2025 · similarity 0.88)
- Label Noise: Ignorance Is Bliss (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.
