Research map / Label-efficient and robust learning / Recognition and long-tail
Addressing Multi-Label Learning with Partial Labels: From Sample Selection to Label Selection
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 Co-Label Selection (CLS), a two-network co-training method for multi-label learning with partial labels that removes likely false-negative and non-informative negative labels instead of dropping samples.
Problem
Multi-label learning with partial labels (ML-PL), where each sample has only a subset of its positive labels annotated; single-network training suffers confirmation bias and sample-selection co-training wastes data.
Approach
Shift co-training from sample selection to label selection: two networks each drop high-loss negative labels for the peer (likely false negatives) and drop low-loss negative labels for itself (to counter positive-negative imbalance), so all samples remain in training.
Key results
- On MS-COCO with single positive label, CLS reaches 67.4 mAP vs 66.1 for JoCoR (+1.3).
- On PASCAL VOC 2007 with single label, CLS reaches 88.6 mAP, 0.9 below ROLE's... (CLS 88.6 vs ASL 85.8, ML-GCN 83.5), outperforming ASL and ML-GCN by 2.8% and 5.1%.
- On NUS-WIDE with 75% labels left, CLS obtains 58.9 mAP, 0.2 above ASL (58.7).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CLS instead of the strongest mAP baseline | Multi-label learning with partial labels · MS-COCO, 25% labels left | mAP: 1.1% | Helped |
| Use CLS instead of the strongest mAP baseline | Multi-label learning with partial labels · MS-COCO, single label | mAP: 1.3% | Helped |
| Use CLS instead of ROLE | Multi-label learning with partial labels · PASCAL VOC 2007, 75% labels left | mAP: 0.3% | Hurt |
| Use CLS instead of ROLE | Multi-label learning with partial labels · PASCAL VOC 2007, 50% labels left | mAP: 0.1% | Helped |
| Use CLS instead of ROLE | Multi-label learning with partial labels · PASCAL VOC 2007, 25% labels left | mAP: 0.7% | Helped |
| Use CLS instead of ROLE | Multi-label learning with partial labels · PASCAL VOC 2007, single label | mAP: 0.9% | Helped |
| Add removal of high-loss negative labels (ablation) | CLS · MS-COCO, 25% labels left | mAP | Helped |
| Add removal of high-loss negative labels (ablation) | CLS · MS-COCO, single label | mAP | Helped |
| Add removal of low-loss negative labels (ablation) | CLS · MS-COCO, 25% labels left | mAP | Helped |
| Add removal of low-loss negative labels (ablation) | CLS · MS-COCO, single label | mAP | Helped |
| Replace label selection with label correction (ablation) | CLS on MS-COCO · MS-COCO, 25% labels left and single label | mAP | Hurt |
Surprising findings
- Correcting false-negative labels (Co-Label Correction) performs worse than simply removing them, degrading after the first few epochs.
Limitations
- Relies on a hyperparameter schedule (reserving rates tau_h, tau_l) chosen by cross-validation per dataset and configuration.
- Performance on PASCAL VOC 2007 with 75% labels is slightly below ROLE.
Experimental setup
Datasets: MS-COCO PASCAL VOC 2007 NUS-WIDE
Benchmarks: MS-COCO PASCAL VOC 2007 NUS-WIDE
Metrics: mAP CF1 OF1
Baselines: ML-GCN CTRAN ASL Co-teaching Co-teaching+ JoCoR WAN ROLE
Builds on: Co-teaching Co-teaching+ JoCoR Binary Cross Entropy
Closest papers by meaning
- Adaptive Integration of Partial Label Learning and Negative Learning for Enhanced Noisy Label Learning (AAAI 2024 · similarity 0.88)
- Multi-Label Supervised Contrastive Learning (AAAI 2024 · similarity 0.88)
- Learning with Complementary Labels Revisited: The Selected-Completely-at-Random Setting Is More Practical (ICML 2024 · similarity 0.88)
- Double-Layer Hybrid-Label Identification Feature Selection for Multi-View Multi-Label Learning (AAAI 2024 · similarity 0.87)
- Learning with Partial-Label and Unlabeled Data: A Uniform Treatment for Supervision Redundancy and Insufficiency (ICML 2024 · similarity 0.87)
- 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.