Research map / Label-efficient and robust learning / Recognition and long-tail
What Makes Partial-Label Learning Algorithms Effective?
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.
The paper analyzes what drives success in partial-label learning, argues that mini-batch pseudo-label purification is the key principle, and proposes SASM as a minimal algorithm plus the StreamPurify instance-dependent warm-up.
Problem
Modern partial-label learning (PLL) methods are complicated and differ widely, making it unclear which components actually drive effectiveness on benchmark-simulated partial-label datasets.
Approach
Dissects SOTA PLL methods into components and finds that mini-batch pseudo-label purification (progressive transition from uniform to one-hot targets) is the essential ingredient, while data augmentation and other borrowed techniques mainly help representation learning; distills a minimal single-augmentation single-model algorithm (SASM) and proposes StreamPurify, an instance-dependent warm-up.
Key results
- SASM, a single-augmentation single-model method with no extra components, performs comparably to or better than SOTA PLL methods such as PiCO, PaPi and CroSel in most settings (e.g., CIFAR-100 flipping 0.05: about 80% for DADM/SADM vs. 80.40 for CroSel).
- Hard pseudo-labels (DASM-H) drop CIFAR-100 accuracy at flip rate 0.7 from ~79 to 33.22, showing that progressive soft purification is essential.
- StreamPurify improves SASM by up to 1.82 points on mini-ImageNet (instance-dependent setting) and improves most other methods across datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace mini-batch pseudo-label updates with hard pseudo-label targets (DASM-H) (ablation) | DASM · CIFAR-100, η=0.1 | accuracy (%): 79.62 -> 33.22 | Hurt |
| Use one-step EM with uniform targets for 50 epochs, then fixed hard labels (DASM-S) (ablation) | DASM · mini-ImageNet, instance-dependent | accuracy (%): 54.71 -> 36.71 | Hurt |
| Update pseudo-labels epoch-wise rather than mini-batch-wise (DASM-E) (ablation) | DASM · mini-ImageNet, instance-dependent | accuracy (%): 54.71 -> 53.44 | Hurt |
| Remove data augmentation (ablation) | DASM · CIFAR-100, η=0.05 | accuracy (%): 79.70 -> 60.18 | Hurt |
| Replace DPLL's two losses with alternative DA Match/loss formulations (ablation) | DPLL · CIFAR-100, η=0.05 | accuracy (%): 76.81; 79.35; 79.75; 79.53 | No clear effect |
| Remove DA Match purification and retain only the ABS loss (Eq. 6) (ablation) | DPLL · mini-ImageNet, instance-dependent | accuracy (%): 52.22 -> 44.69 | Hurt |
| Use hard pseudo-labels (DADM-H) instead of mini-batch purification (ablation) | DADM · CIFAR-100, η=0.1 | accuracy (%): 79.13 -> 35.72 | Hurt |
| Use one-step EM updates (DADM-S) (ablation) | DADM · mini-ImageNet, instance-dependent | accuracy (%): 53.69 -> 36.42 | Hurt |
| Update pseudo-labels epoch-wise (DADM-E) (ablation) | DADM · mini-ImageNet, instance-dependent | accuracy (%): 53.69 -> 52.18 | Hurt |
| Remove the second augmentation view (ablation) | DADM · mini-ImageNet, instance-dependent | accuracy (%): 53.69 -> 54.64 | Helped |
| Remove the prototypical classifier and streamline PaPi as PRODEN+ (ablation) | PaPi · mini-ImageNet, instance-dependent | accuracy (%): 57.10 -> 52.39 | Hurt |
| Remove data augmentation (ablation) | PRODEN · CIFAR-100, η=0.05 | accuracy (%): 79.85 -> 62.47 | Hurt |
Surprising findings
- Data augmentation mainly improves representation learning rather than true-label disambiguation; using augmented views to supervise original ones degrades accuracy.
- Ablating the ABS-style loss does not necessarily hurt: the simple maximum-likelihood loss under SGD implicitly yields progressive purification, matching SASM (Lemma 3.1).
Limitations
- Experiments are limited to simulated PLL on vision benchmarks; applicability to other domains is not tested.
- Results use re-implementations focusing on core components, so reported numbers may be slightly lower than originals (e.g., no mixup for PaPi).
- Broader impact not discussed and compute resources not fully specified.
Experimental setup
Datasets: FMNIST CIFAR-10 CIFAR-100 mini-ImageNet
Benchmarks: FMNIST CIFAR-10 CIFAR-100 mini-ImageNet
Metrics: classification accuracy
Baselines: PRODEN PiCO PaPi CroSel DPLL CC LWC SoLar
Builds on: PRODEN DPLL PiCO PaPi CroSel
Closest papers by meaning
- Realistic Evaluation of Deep Partial-Label Learning Algorithms (ICLR 2025 · similarity 0.91)
- Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning (AAAI 2024 · similarity 0.90)
- Partial Label Learning with a Partner (AAAI 2024 · similarity 0.89)
- Conformal Prediction for Partial Label Learning (AAAI 2025 · similarity 0.89)
- Does Label Smoothing Help Deep Partial Label Learning? (ICML 2024 · similarity 0.88)
- Reduction-based Pseudo-label Generation for Instance-dependent Partial Label Learning (NeurIPS 2025 · 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.