Research map / Label-efficient and robust learning / Recognition and long-tail
Confidence-aware Contrastive Learning for Selective Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CCL-SC, a confidence-aware supervised contrastive method that optimizes the feature layer of selective classifiers, motivated by a PAC-Bayes generalization bound tied to intra-class variance.
Problem
Selective classification: learning classifiers that abstain on uncertain inputs while minimizing selective risk at a target coverage, where prior work only modified the classification/selection head.
Approach
Derives a generalization bound showing selective risk depends on intra-class feature variance, then adds a confidence-weighted supervised contrastive loss that defines positives/negatives by prediction correctness (MoCo-style queues) and weights it by max softmax (SR).
Key results
- On CelebA, CCL-SC achieves the lowest selective risk at all 13 coverage levels (e.g., 0.25% vs 0.41% for next best at 10% coverage... ranked first with avg rank 1.00).
- On ImageNet, CCL-SC reduces selective risk versus SAT+EM+SR across coverages (e.g., 15.76% vs 16.83% at 80% coverage), except at 10% coverage.
- On CIFAR-10, CCL-SC is best at coverage 85% and above; at 80% coverage the best is SAT+EM+SR (0.67 vs 0.69).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add confidence-aware contrastive learning (CCL-SC) rather than using SAT+EM+SR | VGG16 selective classifier · CIFAR-10, 100% coverage | selective risk (%): 6.14±0.07 -> 5.97±0.11 | Helped |
| Add CCL-SC rather than using SAT+EM+SR | VGG16 selective classifier · CIFAR-10, 80% coverage | selective risk (%): 0.67±0.07 -> 0.69±0.08 | Mixed |
| Add CCL-SC rather than using SAT+EM+SR | ResNet18 selective classifier · CelebA | selective risk (%): 19.04±0.30 -> 18.71±0.16 | Helped |
| Add CCL-SC rather than using SAT+EM+SR | VGG16 selective classifier · CIFAR-100, 100% coverage | selective risk (%): 26.96±0.14 -> 26.55±0.26 | Helped |
| Add CCL-SC rather than using SAT+EM+SR | VGG16 selective classifier · CIFAR-100, 30% coverage | selective risk (%): 0.96±0.10 -> 1.26±0.17 | Hurt |
| Add CCL-SC rather than using SAT+EM+SR | ResNet34 selective classifier · ImageNet, 50% coverage | selective risk (%): 5.57 ± 0.17 -> 4.79±0.04 | Helped |
| Add CCL-SC rather than using SAT+EM+SR | ResNet34 selective classifier · ImageNet, 10% coverage | selective risk (%): 0.55 ± 0.05 -> 0.72±0.05 | Hurt |
| Remove SR weighting from the CSC loss (ablation) | CCL-SC on CIFAR-100 · CIFAR-100 | selective risk (%) | Hurt |
| Add negative samples from other classes in addition to samples misclassified as the anchor class (ablation) | CCL-SC on CIFAR-100 (CCL-SC2) · CIFAR-100 | selective risk (%) | No clear effect |
| Replace defined misclassified-as-anchor negatives with randomly sampled negatives from other categories (ablation) | CCL-SC on CIFAR-100 (CCL-SC3) · CIFAR-100 | selective risk (%) | Hurt |
| Replace the CCL-SC positive/negative definition and loss with supervised contrastive learning (ablation) | CCL-SC on CIFAR-100 (CCL-SC+SupCon) · CIFAR-100 | selective classification performance | Hurt |
| Combine CCL-SC with SAT and entropy-minimization regularization | CCL-SC · Selective classification benchmarks | selective risk (%) | Helped |
Surprising findings
- On CIFAR-10 the discrimination among methods is low because few misclassified samples exist, so the authors focus on other datasets.
- Using SR as the confidence function substantially improves SAT and DG variants, which the authors attribute to difficulty of learning extra selection logits as class count grows.
Limitations
- Gains on CIFAR-10 are small because the dataset is nearly saturated.
- On CIFAR-100 CCL-SC is worse than SAT-related methods at 20-40% coverage.
- Requires momentum encoder and two queues, adding training overhead and hyperparameters (queue size, weight, initial epochs).
Experimental setup
Datasets: CIFAR-10 CIFAR-100 CelebA ImageNet
Benchmarks: CIFAR-10 CIFAR-100 CelebA ImageNet
Metrics: selective risk coverage Wilcoxon rank-sum test average rank
Baselines: SR SAT SAT+EM DG SelectiveNet Focal loss Adaptive Focal loss Soft AvUC loss Soft ECE loss MMCE loss AUCOC loss
Builds on: SupCon MoCo InfoNCE Self-Adaptive Training Deep Gamblers
Closest papers by meaning
- Hierarchical Selective Classification (NeurIPS 2024 · similarity 0.87)
- Selective-Supervised Contrastive Learning With Noisy Labels (CVPR 2022 · similarity 0.86)
- Investigating Why Contrastive Learning Benefits Robustness against Label Noise (ICML 2022 · similarity 0.86)
- Ensemble Selective Classification (NeurIPS 2026 · similarity 0.86)
- SAND: One-Shot Feature Selection with Additive Noise Distortion (ICML 2025 · similarity 0.83)
- Generalization Analysis for Supervised Contrastive Representation Learning under Non-IID Settings (ICML 2025 · 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.