Research map / Label-efficient and robust learning / Recognition and long-tail
Reduction-based Pseudo-label Generation for Instance-dependent Partial Label Learning
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 reduction-based pseudo-labels for instance-dependent partial label learning, generated by weighted aggregation of a multi-branch auxiliary model whose branches are each trained with one label excluded, with meta-learned weights.
Problem
Instance-dependent partial label learning (ID-PLL), where candidate labels correlate with features, causes the training model to overfit incorrect candidate labels and produce poor self-generated pseudo-labels, creating a training bottleneck.
Approach
Train a multi-branch auxiliary model where branch j is trained in a label subspace excluding label j, and aggregate branch outputs with a meta-learned weight vector to form reduction-based pseudo-labels that avoid the influence of strongly associated incorrect labels; theoretically argued to be more consistent with the Bayes optimal classifier than pseudo-labels from the predictive model itself.
Key results
- On CIFAR-10, CIFAR-100 and TinyImageNet the method reaches 87.53%, 65.03% and 40.74% accuracy, the best among compared methods (e.g., PLDIRK 84.63%, 58.17%, 25.77%).
- On real-world datasets it attains 81.07% (Lost), 75.27% (BirdSong), 51.65% (MSRCv2), 56.94% (Soccer Player), 68.01% (Yahoo!News), significantly above the next best method on each.
- The meta-learned weight variant outperforms the uniform-weight variant across all eight datasets (e.g., 87.53% vs 85.43% on CIFAR-10, 46.62% vs 51.65% gap on MSRCv2 is 5.03 points).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use reduction-based pseudo-label generation | ID-PLL methods · CIFAR-10, CIFAR-100, TinyImageNet | accuracy: 87.53$\pm$0.21\%; 65.03$\pm$0.21\%; 40.74$\pm$0.64\% | Helped |
| use reduction-based pseudo-label generation | ID-PLL methods · Lost, BirdSong, MSRCv2, Soccer Player, Yahoo!News | accuracy: 81.07$\pm$0.74\%; 75.27$\pm$0.23\%; 51.65$\pm$0.65\%; 56.94$\pm$0.34\%; 68.01$\pm$0.19\% | Helped |
| replace meta-learned branch weights with uniform weights (ablation) | Reduction-based pseudo-label generation · CIFAR-10 | accuracy: 85.43$\pm$0.52\%$\bullet$ -> 87.53$\pm$0.21\% | Helped |
| replace meta-learned branch weights with uniform weights (ablation) | Reduction-based pseudo-label generation · CIFAR-100 | accuracy: 61.18$\pm$0.34\%$\bullet$ -> 65.03$\pm$0.21\% | Helped |
| replace meta-learned branch weights with uniform weights (ablation) | Reduction-based pseudo-label generation · TinyImageNet | accuracy: 35.35$\pm$1.01\%$\bullet$ -> 40.74$\pm$0.64\% | Helped |
| replace meta-learned branch weights with uniform weights (ablation) | Reduction-based pseudo-label generation · Lost | accuracy: 77.14$\pm$1.62\%$\bullet$ -> 81.07$\pm$0.74\% | Helped |
| replace meta-learned branch weights with uniform weights (ablation) | Reduction-based pseudo-label generation · BirdSong | accuracy: 73.18$\pm$0.71\%$\bullet$ -> 75.27$\pm$0.23\% | Helped |
| replace meta-learned branch weights with uniform weights (ablation) | Reduction-based pseudo-label generation · MSRCv2 | accuracy: 46.62$\pm$1.54\%$\bullet$ -> 51.65$\pm$0.65\% | Helped |
| replace meta-learned branch weights with uniform weights (ablation) | Reduction-based pseudo-label generation · Soccer Player | accuracy: 55.48$\pm$0.65\%$\bullet$ -> 56.94$\pm$0.34\% | Helped |
| replace meta-learned branch weights with uniform weights (ablation) | Reduction-based pseudo-label generation · Yahoo!News | accuracy: 66.82$\pm$0.13\%$\bullet$ -> 68.01$\pm$0.19\% | Helped |
Surprising findings
- Vanilla uniform-weight variant still beats most ID-PLL baselines, though the meta-learned weights add further gains.
- Method is noticeably more time-costly than baselines due to bi-level optimization, as acknowledged in conclusion.
Limitations
- Higher training cost from bi-level optimization.
- Theoretical guarantees rely on mild assumptions (Tsybakov condition, bound on auxiliary model approximation error) that are not empirically verified.
- Requires a clean validation set for meta-learning the weights.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 TinyImageNet Lost BirdSong MSRCv2 Soccer Player Yahoo!News
Benchmarks: CIFAR-10 CIFAR-100 TinyImageNet Lost BirdSong MSRCv2 Soccer Player Yahoo!News
Metrics: classification accuracy pairwise t-test at 0.05 significance
Baselines: PLDIRK PLSDCT POP PLIDGP PLABLE VPLL PLCR PiCO CAVL LWS RC-PLL CC-PLL
Builds on: instance-dependent PLL candidate label generation PiCO meta-weight-net style online bi-level optimization
Closest papers by meaning
- Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning (AAAI 2024 · similarity 0.91)
- Partial Label Causal Representation Learning for Instance-Dependent Supervision and Domain Generalization (AAAI 2025 · similarity 0.89)
- Learning with Partial-Label and Unlabeled Data: A Uniform Treatment for Supervision Redundancy and Insufficiency (ICML 2024 · similarity 0.89)
- What Makes Partial-Label Learning Algorithms Effective? (NeurIPS 2024 · similarity 0.88)
- Realistic Evaluation of Deep Partial-Label Learning Algorithms (ICLR 2025 · similarity 0.88)
- Partial Label Learning with a Partner (AAAI 2024 · similarity 0.87)
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.