Research map / Label-efficient and robust learning / Recognition and long-tail
Can Class-Priors Help Single-Positive Multi-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 CRISP, a single-positive multi-label learning framework that estimates per-class priors via a threshold-based PU-style estimator and uses them in an unbiased, absolute-loss risk estimator with a logit bias for negatives.
Problem
Single-positive multi-label learning (SPMLL) assumes identical class priors when assigning pseudo-labels to unannotated labels, which is unrealistic and biases the model when class priors differ across categories.
Approach
Estimate each class's prior with a threshold-based estimator (top-bin style, from PU learning) that is guaranteed to converge, then plug the estimated priors into an absolute-loss unbiased risk estimator with a class-specific logit bias (b^j = 1 - pi^j) to shift the decision boundary.
Key results
- On VOC, COCO, NUS and CUB mAP, the method reaches 89.931, 74.913, 50.045 and 22.150, the best among compared methods (next best SMIL 89.199, 72.920, 48.743, 21.890).
- On MLL datasets the method achieves best ranking loss on 6 of 6 datasets, e.g., 0.164 on Image versus 0.251 for the next best SMIL.
- Estimating class priors from the validation set (CRISP-val) yields lower mAP than the estimator (e.g., CUB 19.600 vs 22.150).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CRISP with estimated class-priors rather than the strongest listed baseline | SPMLL · VOC | mAP: 89.224$\pm$0.017 -> 89.931$\pm$0.014 | Helped |
| Use CRISP with estimated class-priors rather than the strongest listed baseline | SPMLL · COCO | mAP: 73.272$\pm$0.006 -> 74.913$\pm$0.235 | Helped |
| Use CRISP with estimated class-priors rather than the strongest listed baseline | SPMLL · NUS | mAP: 49.590$\pm$0.021 -> 50.045$\pm$0.112 | Helped |
| Use CRISP with estimated class-priors rather than the strongest listed baseline | SPMLL · CUB | mAP: 21.890$\pm$0.347 -> 22.150$\pm$0.028 | Helped |
| Update class-priors every three epochs instead of every epoch (ablation) | CRISP · VOC | mAP: 89.931$\pm$0.014 -> 89.077$\pm$0.251 | Hurt |
| Update class-priors every three epochs instead of every epoch (ablation) | CRISP · COCO | mAP: 74.913$\pm$0.235 -> 73.930$\pm$0.399 | Hurt |
| Update class-priors every three epochs instead of every epoch (ablation) | CRISP · NUS | mAP: 50.045$\pm$0.112 -> 49.463$\pm$0.216 | Hurt |
| Update class-priors every three epochs instead of every epoch (ablation) | CRISP · CUB | mAP: 22.150$\pm$0.028 -> 19.450$\pm$0.389 | Hurt |
| Estimate class-priors from validation labels instead of using the CRISP estimator (ablation) | CRISP · VOC | mAP: 89.931$\pm$0.014 -> 89.585$\pm$0.318 | Hurt |
| Estimate class-priors from validation labels instead of using the CRISP estimator (ablation) | CRISP · COCO | mAP: 74.913$\pm$0.235 -> 74.435$\pm$0.148 | Hurt |
| Estimate class-priors from validation labels instead of using the CRISP estimator (ablation) | CRISP · NUS | mAP: 50.045$\pm$0.112 -> 49.230$\pm$0.113 | Hurt |
| Estimate class-priors from validation labels instead of using the CRISP estimator (ablation) | CRISP · CUB | mAP: 22.150$\pm$0.028 -> 19.600$\pm$1.400 | Hurt |
Surprising findings
- A fixed uniform class prior, even when tuned across values, underperforms the estimated class priors, supporting the claim that identical-prior assumptions hurt SPMLL.
Limitations
- Theoretical guarantees depend on the classifier's ability to separate positives from negatives; the error term grows with q^n/q^p.
- Exhaustive threshold search over outputs adds per-epoch overhead, though small in the reported timings.
Experimental setup
Datasets: PASCAL VOC 2021 MS-COCO 2014 NUS-WIDE CUB-200 2011 Image Scene Yeast Corel5k Mirflickr Delicious
Benchmarks: PASCAL VOC 2021 MS-COCO 2014 NUS-WIDE CUB-200 2011 Image Scene Yeast Corel5k Mirflickr Delicious
Metrics: mAP ranking loss hamming loss one-error coverage average precision
Baselines: AN AN-LS WAN EPR ROLE EM EMAPL LAGC SMILE SMIL LLN BoostLU
Builds on: PU learning top bin estimator asymmetric loss (AN) EPR ROLE EMAPL
Closest papers by meaning
- One Positive Label is Sufficient: Single-Positive Multi-Label Learning with Label Enhancement (NeurIPS 2022 · similarity 0.92)
- Revisiting Pseudo-Label for Single-Positive Multi-Label Learning (ICML 2023 · similarity 0.91)
- Correlation-Induced Label Prior for Semi-Supervised Multi-Label Learning (ICML 2024 · similarity 0.87)
- Addressing Multi-Label Learning with Partial Labels: From Sample Selection to Label Selection (AAAI 2025 · similarity 0.86)
- Learning with Complementary Labels Revisited: The Selected-Completely-at-Random Setting Is More Practical (ICML 2024 · similarity 0.86)
- Forming Auxiliary High-confident Instance-level Loss to Promote Learning from Label Proportions (CVPR 2025 · similarity 0.85)
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.