Research map / Label-efficient and robust learning / Recognition and long-tail
Controller-Guided Partial Label Consistency Regularization with Unlabeled Data
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 ConCont, a partial label learning method that uses unlabeled data with controller-guided (p-score) label-level and representation-level consistency regularization plus adaptive class-balanced confidence thresholds.
Problem
Partial label learning with insufficient partially annotated examples, where consistency-regularization PLL methods degrade sharply due to weak initial supervision, overfitting and class imbalance.
Approach
A controller computes a partial confidence score (p-score) from label-set size, candidate margin, and non-candidate probability to gate label-level consistency and select positives/negatives for MoCo-style representation-level contrastive consistency; per-class thresholds are adapted by minimizing cross-entropy to a uniform distribution of filtered pseudo-labels.
Key results
- On CIFAR-10 with 10% partial examples, ConCont improves over the best baseline by 5.99%, 10.32% and 16.37% at q=0.1, 0.3 and 0.5.
- ConCont reaches 90.46% accuracy on CIFAR-10 (10%, q=0.5) and 62.85% on CIFAR-100 (20%, q=0.1), versus 74.09% for DPLL at CIFAR-10 q=0.5 and much lower for other baselines.
- Adaptive thresholds outperform DA and CPL class-imbalance methods by 1.38% on CIFAR-10 (10%, q=0.5) and 1.81% on CIFAR-100 (20%, q=0.1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add unlabeled examples to ConCont (ablation) | ConCont · CIFAR-10 (10%, q=0.5) | accuracy: 76.69 -> 90.46 | Helped |
| add label-level consistency regularization (ablation) | ConCont · CIFAR-10 (10%, q=0.5) | accuracy: 80.62 -> 90.46 | Helped |
| add representation-level consistency regularization (ablation) | ConCont · CIFAR-10 (10%, q=0.5) | accuracy: 86.76 -> 90.46 | Helped |
| use partial cross-entropy instead of traditional cross-entropy (ablation) | ConCont · CIFAR-10 (10%, q=0.5) | accuracy: 44.32 -> 90.46 | Helped |
| use the three-branch augmentation design rather than two-branch (ablation) | ConCont · CIFAR-10 (10%, q=0.5) | accuracy: 90.23 -> 90.46 | Helped |
| use adaptive confidence thresholds rather than vanilla thresholds (ablation) | ConCont · CIFAR-10 (10%, q=0.5) | accuracy: 86.50 -> 90.46 | Helped |
| use adaptive confidence thresholds rather than vanilla thresholds (ablation) | ConCont · CIFAR-100 (20%, q=0.1) | accuracy: 59.50 -> 62.85 | Helped |
| replace adaptive thresholds with Distribution Alignment (ablation) | ConCont · CIFAR-10 (10%, q=0.5) | accuracy: 90.46 -> 89.08 | Hurt |
| replace adaptive thresholds with Curriculum Pseudo-Labeling (ablation) | ConCont · CIFAR-100 (20%, q=0.1) | accuracy: 62.85 -> 46.32 | Hurt |
| use the proposed positive/negative sample selection rather than HCP (ablation) | ConCont representation-level regularization · CIFAR-10 (10%, q=0.5) | accuracy: 89.76 -> 90.46 | Helped |
| use the proposed positive/negative sample selection rather than HCPN (ablation) | ConCont representation-level regularization · CIFAR-100 (20%, q=0.1) | accuracy: 62.50 -> 62.85 | Helped |
Surprising findings
- Two-branch variant (same augmentation pair for label and representation consistency) is competitive with the full three-branch design on several settings.
- Using true labels for supervised contrastive learning (Y_true) only beats ConCont by 0.2% on CIFAR-10, and is 1.64% higher on CIFAR-100, suggesting limited headroom from pseudo-label quality.
Limitations
- Experiments limited to image classification benchmarks (CIFAR-10/100, SVHN) with small ResNet-18 backbone
- PARM could not be run on CIFAR-100 due to computational cost, so it is missing from those comparisons
- Several hyperparameters (thresholds, loss weights, learning rate for thresholds) are set per dataset
Experimental setup
Datasets: CIFAR-10 CIFAR-100 SVHN
Benchmarks: CIFAR-10 CIFAR-100 SVHN
Metrics: accuracy
Baselines: DPLL PiCO SSPL PARM DPLL* PiCO*
Builds on: DPLL PiCO MoCo FlexMatch ReMixMatch
Closest papers by meaning
- Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning (AAAI 2024 · similarity 0.89)
- Generative Modeling Reinvents Supervised Learning: Label Repurposing with Predictive Consistency Learning (ICML 2025 · similarity 0.88)
- Learning with Partial-Label and Unlabeled Data: A Uniform Treatment for Supervision Redundancy and Insufficiency (ICML 2024 · similarity 0.88)
- Conformal Prediction for Partial Label Learning (AAAI 2025 · similarity 0.88)
- CroSel: Cross Selection of Confident Pseudo Labels for Partial-Label Learning (CVPR 2024 · similarity 0.88)
- Partial Label Learning with a Partner (AAAI 2024 · similarity 0.88)
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.