Research map / Label-efficient and robust learning / Semi-supervised learning
Learning with Partial-Label and Unlabeled Data: A Uniform Treatment for Supervision Redundancy and Insufficiency
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 SPMI, a mutual-information-based label channel that dynamically expands and condenses pseudo candidate label sets so partial-label and unlabeled data are handled uniformly in semi-supervised partial label learning.
Problem
Semi-supervised partial label learning (SSPLL), where redundant partial labels and insufficient unlabeled data coexist; existing approaches compose separate PLL and SSL components, which the authors argue is incremental.
Approach
A label channel that treats all data uniformly as pseudo candidate label sets: label expansion adds likely true labels (via a mutual-information-derived criterion q(k|z) > p(k)) and label condensation removes the least relevant label (via a KL-based information score), with an initialization step that seeds candidate sets for unlabeled data from complementary-style loss comparisons.
Key results
- SPMI reaches 92.81% on CIFAR-10 (l=4000, p=0.3) versus 92.19% for the best PRODEN-based SSL baseline in the PLL comparison setting (FixMatch-based PLL baselines reported in Table 1), and 66.43% vs 52.06% for PRODEN+FixMatch on CIFAR-10 (l=10000 setting listed in Table 1 columns).
- On partial-complementary label learning, SPMI reaches 94.55% on CIFAR-10 versus 93.55% for the best CLL baseline (NLL) and 78.67% on CIFAR-100 versus 75.35%.
- Ablation shows removing label generation drops CIFAR-100 accuracy from 73.63% to 40.12%, and removing label condensation drops CIFAR-10 accuracy from 92.81% to 85.14%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SPMI instead of the strongest listed PLL+FixMatch baseline | semi-supervised partial-label learning · F-MNIST, CIFAR-10, CIFAR-100, and SVHN (Table 1 settings) | test accuracy | Helped |
| Use SPMI instead of the strongest listed SSL+PRODEN baseline | semi-supervised partial-label learning · F-MNIST and SVHN (Table 2 settings) | test accuracy | Helped |
| Use SPMI instead of the strongest listed SSL+PRODEN baseline | semi-supervised partial-label learning · CIFAR-10 (l=4000, p=0.3) and CIFAR-100 (l=10000, p=0.05) | test accuracy | Hurt |
| Add initialization of pseudo candidate label sets (ablation) | SPMI on unlabeled data · F-MNIST (l=4000, p=0.3) | test accuracy: 89.38±0.15 -> 89.82±0.04 | Helped |
| Add label generation (ablation) | SPMI on unlabeled data · F-MNIST (l=4000, p=0.3) | test accuracy: 87.51±0.36 -> 89.82±0.04 | Helped |
| Add label condensation (ablation) | SPMI on unlabeled data · CIFAR-10 (l=4000, p=0.3) | test accuracy: 85.14±0.55 -> 92.81±0.06 | Helped |
Surprising findings
- SPMI is slightly weaker than FlexMatch and FreeMatch when labeled data is extremely scarce (e.g., CIFAR-10 with l=1000, p=0.3 and CIFAR-100 with l=5000, p=0.05).
- Directly combining PLL methods with FixMatch can fail badly under limited labels, e.g., PLCR and POP reach only 71.19% and 68.63% on CIFAR-10 with l=4000, p=0.3 vs 88.85% for PRODEN-FixMatch.
Limitations
- Experiments are limited to image benchmarks with simulated uniform partial label generation.
- Slightly weaker than FlexMatch/FreeMatch in extremely low-label regimes.
- Theoretical guarantee for initialization relies on a mild assumption that the true label output exceeds random output (1/c).
Experimental setup
Datasets: Fashion-MNIST CIFAR-10 CIFAR-100 SVHN
Benchmarks: Fashion-MNIST CIFAR-10 CIFAR-100 SVHN
Metrics: test accuracy (mean±std over three seeds)
Baselines: PRODEN PLCR POP MixMatch FixMatch FlexMatch FreeMatch Forward L-W NLL
Builds on: PRODEN FixMatch information bottleneck FlexMatch
Closest papers by meaning
- Rethinking Consistent Multi-Label Classification under Inexact Supervision (ICLR 2026 · similarity 0.90)
- All Labels Are Not Created Equal: Enhancing Semi-Supervision via Label Grouping and Co-Training (CVPR 2021 · similarity 0.90)
- Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning (AAAI 2024 · similarity 0.89)
- Reduction-based Pseudo-label Generation for Instance-dependent Partial Label Learning (NeurIPS 2025 · similarity 0.89)
- ULAREF: A Unified Label Refinement Framework for Learning with Inaccurate Supervision (ICML 2024 · similarity 0.89)
- Imprecise Label Learning: A Unified Framework for Learning with Various Imprecise Label Configurations (NeurIPS 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.