Research map / Label-efficient and robust learning / Recognition and long-tail
Partial Label Learning with a Partner
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 PLCP, a mutual-supervision framework that couples any partial label learning classifier with a partner classifier trained on non-candidate labels, with blurring and a collaborative term, to identify and rectify mislabeled samples.
Problem
Partial label learning methods struggle to rectify mislabeled samples: false-positive candidate labels can keep high labeling confidence and mislead final predictions.
Approach
Introduce a partner classifier that models non-candidate labels (which are accurate) and link it to the base PLL classifier via a collaborative term, with a blurring mechanism (negative temperature) that softens confidences; both classifiers supervise each other in an iterative mutual-supervision loop.
Key results
- PLCP coupled with PL-KNN raises FG-NET accuracy from 0.036 to 0.076 and PL-SVM from 0.043 to 0.081.
- Average improvements reported range from 3.61% (PL-CL) to 53.98% (PL-KNN) across real-world datasets; B-PLCP significantly outperforms base classifier in all real-world cases by pairwise t-test.
- On CIFAR-10/100, B-PLCP consistently outperforms the base deep classifier in 83.3% of settings (e.g., PICO on CIFAR-10 q=0.1: 94.39% to 94.80%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Couple PL-AGGD with PLCP | PL-AGGD · Real-world datasets | accuracy: +5.10% | Helped |
| Couple PLL methods with PLCP | Stand-alone PLL methods · Real-world datasets | accuracy | Helped |
| Couple deep PLL methods with PLCP | PICO and PRODEN · CIFAR-10 and CIFAR-100 | accuracy: 83.3% cases | Helped |
| Remove the blurring mechanism (ablation) | PL-AGGD-PLCP · Real-world datasets | accuracy: 87.5% cases | Hurt |
| Replace the non-candidate-label partner with a candidate-label partner (ablation) | PL-AGGD-PLCP · Real-world datasets | accuracy: 87.5% cases | Hurt |
| Remove kernel extension from the partner classifier (ablation) | PL-AGGD-PLCP · Real-world datasets | accuracy | Hurt |
| Use k values that are too small or k ≥ 0 instead of the blurring setting (ablation) | PLCP · Partial label learning | accuracy | Hurt |
Surprising findings
- Removing the blurring mechanism or using a partner built on candidate labels (O variant) significantly hurts performance in most cases, showing the non-candidate partner is key.
Limitations
- Gains on deep-learning methods on CIFAR are modest (often small absolute improvements).
- Partner classifier design and hyperparameters (lambda, alpha, gamma, k) are set by hand and sensitive; blur temperature must be negative.
Experimental setup
Datasets: FG-NET Lost MSRCv2 Mirflickr Soccer Player Yahoo!News CIFAR-10 CIFAR-100
Benchmarks: FG-NET Lost MSRCv2 Mirflickr Soccer Player Yahoo!News CIFAR-10 CIFAR-100
Metrics: classification accuracy transductive accuracy MAE
Baselines: PL-CL PL-AGGD SURE LALO PL-SVM PL-KNN PICO PRODEN
Builds on: PICO PRODEN PL-AGGD SURE LALO
Closest papers by meaning
- Realistic Evaluation of Deep Partial-Label Learning Algorithms (ICLR 2025 · similarity 0.90)
- Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning (AAAI 2024 · similarity 0.89)
- What Makes Partial-Label Learning Algorithms Effective? (NeurIPS 2024 · similarity 0.89)
- Adaptive Integration of Partial Label Learning and Negative Learning for Enhanced Noisy Label Learning (AAAI 2024 · similarity 0.88)
- Learning with Partial-Label and Unlabeled Data: A Uniform Treatment for Supervision Redundancy and Insufficiency (ICML 2024 · similarity 0.88)
- Navigating Real-World Partial Label Learning: Unveiling Fine-Grained Images with Attributes (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.