Research map / Label-efficient and robust learning / Noisy labels
Drainage: A Unifying Framework for Addressing Class Uncertainty
Yasser Taha, Grégoire Montavon, Nils Körber
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 |
|---|---|---|---|
| use drainage loss instead of cross-entropy | CE baseline · CIFAR-100, asymmetric noise 0.4 | accuracy: 41.63 -> 61.55 | Improved |
| use drainage loss instead of cross-entropy | CE baseline · CIFAR-10, asymmetric noise 0.45 | accuracy: 77.92 -> 82.23 | Improved |
| use drainage loss instead of GCE | GCE · CIFAR-10, instance-dependent noise 0.5 | accuracy: 59.27 -> 64.22 | Improved |
| use drainage loss instead of GCE | GCE · CIFAR-100, asymmetric noise 0.45 | accuracy: 42.70 -> 52.69 | Improved |
| use drainage loss on clean data instead of cross-entropy | CE baseline · CIFAR-10, no noise | accuracy: 92.16 -> 91.30 | Hurt |
| use drainage loss instead of CE on clean CIFAR-100 | CE baseline · CIFAR-100, no noise | accuracy: 70.05 -> 73.31 | Improved |
| use drainage loss instead of GCE | GCE · CIFAR-10N (worst labels) | accuracy: 81.70 -> 79.85 | Hurt |
| use drainage loss instead of ANL CE | ANL CE · CIFAR-100N | accuracy: 56.62 -> 57.77 | Improved |
| use drainage loss instead of ANL CE | ANL CE · Clothing1M | accuracy: 68.44 -> 70.42 | Improved |
| use drainage loss instead of ANL CE | ANL CE · ILSVRC12 (mini, 50 classes) | accuracy: 65.77 -> 64.90 | Hurt |
| use L2 instead of L1 regularization for CE (ablation) | CE · CIFAR-10, asymmetric noise 0.4 | accuracy: 82.50 -> 73.38 | Hurt |
| use drainage node probability p_d as OSR score instead of CE max softmax probability | Drainage OSR model · Open set recognition, CIFAR-10 | ROC AUC: 67.3 -> 72.4 | Improved |
Closest papers by meaning
- Confidence-based Reliable Learning under Dual Noises (NeurIPS 2022 · similarity 0.87)
- SURE: SUrvey REcipes for building reliable and robust deep networks (CVPR 2024 · similarity 0.87)
- USDNL: Uncertainty-Based Single Dropout in Noisy Label Learning (AAAI 2023 · similarity 0.87)
- Uncertainty-Aware Learning against Label Noise on Imbalanced Datasets (AAAI 2022 · similarity 0.87)
- Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 2025 · similarity 0.86)
- Enhancing Noise-Robust Losses for Large-Scale Noisy Data Learning (AAAI 2025 · 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.
