Research map / Label-efficient and robust learning / Noisy labels
SuperLoss: A Generic Loss for Robust Curriculum Learning
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 |
|---|---|---|---|
| append SuperLoss on top of task loss | Faster R-CNN box classification · Pascal VOC 60% label noise | AP50: 69.5 -> 74.9 | Helped |
| append SuperLoss on top of focal loss | RetinaNet · Pascal VOC 60% label noise | AP50: 52.0 -> 59.7 | Helped |
| append SuperLoss on clean data | Faster R-CNN · Pascal VOC 0% noise | AP50: 81.4 -> 81.4 | No clear effect |
| append SuperLoss on clean CIFAR-10 | WideResNet-28-10 · CIFAR-10 clean | accuracy (%): 95.8 -> 96.0 | No clear effect |
| append SuperLoss under 40% symmetric label noise | WideResNet-28-10 · CIFAR-100 40% noise | accuracy (%): 71.05 vs 70.93 (Data Parameters) | Helped |
| append SuperLoss to ResNet-18 on WebVision | ResNet-18 · WebVision | accuracy (%): 65.5 -> 66.7 | Helped |
| use SuperLoss with smooth-l1 under 80% noise | ResNet-18 age regression · UTKFace 80% noise | MAE: 17.56 -> 13.09 | Helped |
| append SuperLoss, noisy Landmarks-full training | ResNet-50+GeM contrastive · ROxf Medium | mAP: 54.4 -> 62.7 | Helped |
| append SuperLoss on clean Landmarks | ResNet-50+GeM contrastive · Landmarks-clean average mAP | mAP: 57.2 -> 57.2 | No clear effect |
| fix tau to log(C) vs running average (ablation) | SuperLoss · CIFAR-10/100 | accuracy | No clear effect |
| fix tau to log(C) vs exponential average (ablation) | SuperLoss with Faster R-CNN · Pascal VOC 60% label noise | AP50: 73.8 -> 74.9 | Helped |
| use closed-form optimal confidence instead of learned per-sample confidence (ablation) | SuperLoss vs Data Parameters · CIFAR-10 40% noise | accuracy (%): 91.10 -> 91.55 | Helped |
Closest papers by meaning
- Robust Curriculum Learning: from clean label detection to noisy label self-correction (ICLR 2021 · similarity 0.88)
- Dual-Level Curriculum Meta-Learning for Noisy Few-Shot Learning Tasks (AAAI 2024 · similarity 0.86)
- CurBench: Curriculum Learning Benchmark (ICML 2024 · similarity 0.85)
- Learning to Solve Complex Problems via Dataset Decomposition (NeurIPS 2025 · similarity 0.85)
- Bandit Guided Submodular Curriculum for Adaptive Subset Selection (NeurIPS 2025 · similarity 0.84)
- On the Mechanisms of Weak-to-Strong Generalization: A Theoretical Perspective (NeurIPS 2025 · similarity 0.84)
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.