Research map / Label-efficient and robust learning / Recognition and long-tail
Stochastic smoothing of the top-K calibrated hinge loss for deep imbalanced classification
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 |
|---|---|---|---|
| Smooth the calibrated top-K hinge operator with Gaussian perturbations | Balanced top-K classification on CIFAR-100 · CIFAR-100, label noise p=0.4 | top-5 accuracy: 83.6 ± 0.2 -> 87.8 ± 0.6 | Helped |
| Smooth the calibrated top-K hinge operator with Gaussian perturbations | Balanced top-K classification on CIFAR-100 · CIFAR-100, label noise p=0.1 | top-5 accuracy: 92.2 ± 0.3 -> 91.9 ± 0.1 | Hurt |
| Use uneven class-dependent margins in the noised top-K hinge loss | Noised top-K hinge on Pl@ntNet-300K · Pl@ntNet-300K test, K=5 | macro-average top-5 accuracy: 71.9 ± 0.3 -> 73.2 ± 0.5 | Helped |
| Use uneven class-dependent margins in the noised top-K hinge loss | Noised top-K hinge on ImageNet-LT · ImageNet-LT test, K=3 | top-3 accuracy: 56.0 ± 0.2 -> 56.5 ± 0.1 | Helped |
| Set smoothing parameter epsilon to 0.1 rather than 0.0 (ablation) | Noised balanced hinge on CIFAR-100 · CIFAR-100 validation | top-5 accuracy: 19.38 -> 94.46 | Helped |
| Set smoothing parameter epsilon to 1e-4 (ablation) | Noised balanced hinge on CIFAR-100 · CIFAR-100 validation | top-5 accuracy: 94.46 -> 14.84 | Hurt |
| Set smoothing parameter epsilon to 1e-3 (ablation) | Noised balanced hinge on CIFAR-100 · CIFAR-100 validation | top-5 accuracy: 94.46 -> 11.4 | Hurt |
| Increase smoothing from epsilon=1e-3 to epsilon=1e-2 (ablation) | Noised balanced hinge on CIFAR-100 · CIFAR-100 validation | top-5 accuracy: 11.4 -> 93.36 | Helped |
| Increase Monte Carlo samples from B=1 to B=100 (ablation) | Noised balanced hinge on CIFAR-100 · CIFAR-100 validation | top-5 accuracy: 94.28 -> 94.52 | No clear effect |
Closest papers by meaning
- ST$_k$: A Scalable Module for Solving Top-k Problems (NeurIPS 2024 · similarity 0.86)
- Improved Balanced Classification with Theoretically Grounded Loss Functions (NeurIPS 2025 · similarity 0.86)
- EnsLoss: Stochastic Calibrated Loss Ensembles for Preventing Overfitting in Classification (ICML 2025 · similarity 0.84)
- Balancing the Scales: A Theoretical and Algorithmic Framework for Learning from Imbalanced Data (ICML 2025 · similarity 0.84)
- Uncertainty Weighted Gradients for Model Calibration (CVPR 2025 · similarity 0.84)
- Threshold-Consistent Margin Loss for Open-World Deep Metric Learning (ICLR 2024 · 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.