Research map / Label-efficient and robust learning / Recognition and long-tail
Binary Classification with Confidence Difference
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 |
|---|---|---|---|
| Use confidence-difference supervision with absolute-value risk correction rather than Pcomp-Teacher | ConfDiff classification · MNIST, Kuzushiji, Fashion, and CIFAR-10 (π+=0.5) | accuracy: 0.893$\pm$0.014 -> 0.964$\pm$0.001; 0.782$\pm$0.046 -> 0.867$\pm$0.006; 0.903$\pm$0.016 -> 0.967$\pm$0.001; 0.779$\pm$0. | Helped |
| Use confidence-difference supervision with absolute-value risk correction rather than Pcomp-Teacher | ConfDiff classification · Optdigits, USPS, Pendigits, and Letter (π+=0.5) | accuracy: 0.890$\pm$0.009 -> 0.962$\pm$0.006; 0.860$\pm$0.012 -> 0.959$\pm$0.004; 0.883$\pm$0.018 -> 0.988$\pm$0.003; 0.864$\pm$0. | Helped |
| Use confidence-difference supervision with absolute-value risk correction instead of soft pointwise confidence | KuaiRec recommender · KuaiRec | HR: 0.534 -> 0.570 | Helped |
| Use confidence-difference supervision with absolute-value risk correction instead of soft pointwise confidence | KuaiRec recommender · KuaiRec | NDCG: 0.380 -> 0.372 | Hurt |
| Add absolute-value risk correction rather than use the unbiased risk estimator (ablation) | ConfDiff classification · MNIST (π+=0.5) | accuracy: 0.911$\pm$0.046 -> 0.964$\pm$0.001 | Helped |
| Add absolute-value risk correction rather than use the unbiased risk estimator (ablation) | ConfDiff classification · Pendigits (π+=0.5) | accuracy: 0.945$\pm$0.052 -> 0.988$\pm$0.003 | Helped |
| Use ReLU instead of absolute-value risk correction (ablation) | ConfDiff classification · MNIST (π+=0.5) | accuracy: 0.964$\pm$0.001 -> 0.944$\pm$0.011 | Hurt |
| Use ReLU instead of absolute-value risk correction (ablation) | ConfDiff classification · Pendigits (π+=0.5) | accuracy: 0.988$\pm$0.003 -> 0.981$\pm$0.004 | Hurt |
| Use only 10% of training pairs rather than the full ConfDiff training set (ablation) | ConfDiff-ReLU and ConfDiff-ABS · Experiments with π+=0.5 | accuracy | Mixed |
Closest papers by meaning
- Confidence Difference Reflects Various Supervised Signals in Confidence-Difference Classification (ICML 2025 · similarity 0.92)
- Don’t Label Twice: Quantity Beats Quality when Comparing Binary Classifiers on a Budget (ICML 2024 · similarity 0.85)
- Learning with Complementary Labels Revisited: The Selected-Completely-at-Random Setting Is More Practical (ICML 2024 · similarity 0.84)
- Forming Auxiliary High-confident Instance-level Loss to Promote Learning from Label Proportions (CVPR 2025 · similarity 0.83)
- Retraining with Predicted Hard Labels Provably Increases Model Accuracy (ICML 2025 · similarity 0.83)
- Leveraging Uncertainty Estimates To Improve Classifier Performance (ICLR 2024 · similarity 0.83)
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.