Research map / Label-efficient and robust learning / Recognition and long-tail
Targeted Supervised Contrastive Learning for Long-Tailed Recognition
Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio S. Feris, Piotr Indyk, Dina Katabi
Check 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 |
|---|---|---|---|
| Add uniformly distributed targets and online matching to supervised contrastive learning | KCL · CIFAR-10-LT, rho=100 | top-1 accuracy: 77.6 -> 79.7 | Improved |
| Add uniformly distributed targets and online matching to supervised contrastive learning | KCL · CIFAR-10-LT, rho=50 | top-1 accuracy: 81.7 -> 82.9 | Improved |
| Add uniformly distributed targets and online matching to supervised contrastive learning | KCL · CIFAR-100-LT, rho=100 | top-1 accuracy: 42.8 -> 43.8 | Improved |
| Add uniformly distributed targets and online matching to supervised contrastive learning | KCL · ImageNet-LT, all | top-1 accuracy: 51.5 -> 52.4 | Improved |
| Add uniformly distributed targets and online matching to supervised contrastive learning | KCL · iNaturalist 2018, all | top-1 accuracy: 68.6 -> 69.7 | Improved |
| Use class-balanced sampling with KCL instead of instance-balanced sampling (ablation) | KCL · CIFAR-10-LT, rho=100 | top-1 accuracy: 77.6 -> 75.5 | Hurt |
| Use class-balanced sampling with KCL instead of instance-balanced sampling (ablation) | KCL · CIFAR-100-LT, rho=100 | top-1 accuracy: 42.8 -> 41.5 | Hurt |
| Replace online matching with fixed random class-target assignment (ablation) | TSC · ImageNet-LT | top-1 accuracy: 52.4 -> 50.8 | Hurt |
| Replace online matching with fixed random class-target assignment (ablation) | TSC · ImageNet-LT | reasonability (R): 7.14 -> 7.81 | Hurt |
| Use KCL's balanced positive sampling rather than FCL's positive sampling (ablation) | TSC · ImageNet-LT, all | top-1 accuracy: 51.8 -> 52.4 | Improved |
| Replace RIDE stage-1 training with TSC | RIDE (4 experts) · ImageNet-LT, all | top-1 accuracy: 55.4 -> 56.9 | Improved |
Closest papers by meaning
- Long-Tailed Recognition by Mutual Information Maximization between Latent Features and Ground-Truth Labels (ICML 2023 · similarity 0.92)
- Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition (NeurIPS 2024 · similarity 0.91)
- Decoupled Contrastive Learning for Long-Tailed Recognition (AAAI 2024 · similarity 0.90)
- Long-tailed Object Detection Pretraining: Dynamic Rebalancing Contrastive Learning with Dual Reconstruction (NeurIPS 2024 · similarity 0.89)
- Long-Tailed Recognition via Information-Preservable Two-Stage Learning (NeurIPS 2025 · similarity 0.88)
- A Tale of Two Classes: Adapting Supervised Contrastive Learning to Binary Imbalanced Datasets (CVPR 2025 · similarity 0.87)
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.
