Research map / Learning theory and algorithms / Prediction and decision losses
Training Over-parameterized Models with Non-decomposable Objectives
Harikrishna Narasimhan, Aditya Krishna Menon
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 loss re-weighting with logit-adjusted loss for the diagonal-gain cost-sensitive step | ResNet cost-sensitive learning · CIFAR-10-LT | Min Rec: 0.592 -> 0.721 | Improved |
| Replace loss re-weighting with logit-adjusted loss for the diagonal-gain cost-sensitive step | ResNet cost-sensitive learning · CIFAR-100-LT | Min HT Rec: 0.083 -> 0.418 | Improved |
| Replace loss re-weighting with logit-adjusted loss for the diagonal-gain cost-sensitive step | ResNet cost-sensitive learning · TinyImgNet-LT | Min HT Rec: 0.098 -> 0.310 | Improved |
| Use hybrid cost-sensitive loss instead of re-weighted loss | ResNet cost-sensitive learning · CIFAR-10-LT | Avg Rec: 0.739 -> 0.759 | Improved |
| Use hybrid cost-sensitive loss instead of re-weighted loss | ResNet cost-sensitive learning · CIFAR-100-LT | Avg Rec: 0.408 -> 0.474 | Improved |
| Use hybrid cost-sensitive loss instead of re-weighted loss | ResNet cost-sensitive learning · TinyImgNet-LT | Avg Rec: 0.289 -> 0.350 | Improved |
| Use hybrid loss with D = diag(1/π) rather than D = diag(Gii) (ablation) | CSL hybrid loss · CIFAR-10-LT | Avg Rec: 0.759 vs 0.760 | No clear effect |
| Use hybrid loss with self-distillation rather than ordinary training (ablation) | CSL logit-adjusted method · CIFAR-10-LT | Min Rec: 0.721 -> 0.734 | Improved |
| Use hybrid loss with self-distillation rather than ordinary training (ablation) | CSL hybrid method · CIFAR-100-LT | Min HT Rec: 0.418 -> 0.467 | Improved |
| Train with proposed cost-sensitive loss instead of ERM, then apply post-shifting | CIFAR-100-LT worst-case recall over all classes · CIFAR-100-LT | Min Rec: 0.090 -> 0.120 | Mixed |
Closest papers by meaning
- Thumb on the Scale: Optimal Loss Weighting in Last Layer Retraining (NeurIPS 2025 · similarity 0.88)
- Active Learning for Cost-Sensitive Classification (ICML 2017 · similarity 0.86)
- Machine Unlearning under Overparameterization (NeurIPS 2025 · similarity 0.85)
- Learning from Noisy Labels via Self-Taught On-the-Fly Meta Loss Rescaling (AAAI 2025 · similarity 0.84)
- Enhancing Noise-Robust Losses for Large-Scale Noisy Data Learning (AAAI 2025 · similarity 0.84)
- Retraining with Predicted Hard Labels Provably Increases Model Accuracy (ICML 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.
