Research map / Label-efficient and robust learning / Recognition and long-tail
Label-similarity Curriculum Learning
Ürün Dogan, Aniket Anand Deshmukh, Marcin Bronislaw Machura, Christian Igel
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 one-hot labels with label-similarity curriculum (LCL) | ResNet-50 · ImageNet 5% data | top-1 accuracy: 38.21 -> 42.21 | Improved |
| replace one-hot labels with LCL | ResNet-50 · ImageNet 100% data | top-1 accuracy: 76.25 -> 76.89 | Improved |
| replace one-hot labels with LCL | ResNet-101 · AWA2 5% data | top-1 accuracy: 23.09 -> 45.78 | Improved |
| replace one-hot labels with LCL | ResNet-101 · AWA2 5% data | top-5 accuracy: 53.31 -> 76.47 | Improved |
| replace one-hot labels with LCL | ResNeXt-101 · ImageNet 100% data | top-5 accuracy: 93.95 -> 96.15 | Improved |
| replace one-hot labels with LCL | WRN · CIFAR100 100% data | top-1 accuracy: 79.52 -> 81.25 | Improved |
| replace static label smoothing with LCL | DenseNet-BC · CIFAR100 100% data | top-1 accuracy: 83.01 -> 83.2 | Improved |
| replace knowledge distillation with LCL | ResNet-50 · ImageNet 5% data | top-1 accuracy: 39.81 -> 42.21 | Improved |
| increase epsilon from 0.9 to 0.999 (slower decay to one-hot) (ablation) | LCL ResNet-50 · ImageNet 5% data | top-1 accuracy: 41.24 -> 42.21 | Improved |
| increase epsilon from 0.9 to 0.999 (ablation) | LCL ResNet-101 · AWA2 10% data | top-1 accuracy: 47.21 -> 53.39 | Improved |
| replace one-hot labels with LCL on SENet-154 top-5 at 10% data | SENet-154 · ImageNet 10% data | top-5 accuracy: 83.71 -> 83.6 | Hurt |
| use curriculum with limited data (5% vs 100%) gain size (ablation) | LCL ResNet-50 · ImageNet | top-1 accuracy: 76.25 -> 76.89 | Improved |
Closest papers by meaning
- On The Power of Curriculum Learning in Training Deep Networks (ICML 2019 · similarity 0.87)
- Curriculum Labeling: Revisiting Pseudo-Labeling for Semi-Supervised Learning (AAAI 2021 · similarity 0.87)
- Robust Curriculum Learning: from clean label detection to noisy label self-correction (ICLR 2021 · similarity 0.87)
- Interactive Multi-Label CNN Learning With Partial Labels (CVPR 2020 · similarity 0.86)
- All Labels Are Not Created Equal: Enhancing Semi-Supervision via Label Grouping and Co-Training (CVPR 2021 · similarity 0.86)
- Calibrated Disambiguation for Partial Multi-label Learning (AAAI 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.
