Research map / Label-efficient and robust learning / Noisy labels
CroSel: Cross Selection of Confident Pseudo Labels for 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 CroSel, which uses two networks that select confident pseudo labels from memory banks of historical predictions for each other, plus a MixUp-based co-mix consistency term, to disambiguate partial-label learning.
Problem
Partial-label learning, where each training example has a candidate label set containing the true label; identifying the true label accurately and completely is challenging, and noisy pseudo labels hurt training.
Approach
Cross selection: two networks each keep a FIFO memory bank of softmax outputs over past epochs and select high-confidence, stable, candidate-consistent labels to train the other network; a co-mix consistency loss (weak/strong augmentation cross-generated soft targets mixed with MixUp) supervises the remaining unselected examples.
Key results
- CroSel reaches 84.07% test accuracy on CIFAR-100 (q=0.1) versus 82.39% for the next best PoP, and state-of-the-art on most settings.
- On CIFAR-10 (q=0.5) CroSel gets 97.34% accuracy with 96.25% selection ratio and 99.44% selection accuracy.
- Under fine-grained superclass candidate sets on CIFAR-100, CroSel achieves 83.34% accuracy versus 82.04% for PoP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CroSel instead of CRDPLL | WRN-34-10 on CIFAR-10 (q=0.1) · CIFAR-10 (q=0.1) | accuracy: 97.31±.04% vs 97.41±.06% | Hurt |
| Use CroSel instead of the strongest baseline | WRN-34-10 on CIFAR-10 (q=0.3) · CIFAR-10 (q=0.3) | accuracy: 97.50±.05% vs 97.38±.04% | Helped |
| Use CroSel instead of the strongest baseline | WRN-34-10 on SVHN (q=0.5) · SVHN (q=0.5) | accuracy: 97.86±.06% vs 97.70±.05% | Helped |
| Use CroSel instead of the strongest baseline | WRN-34-10 on CIFAR-100 (q=0.1) · CIFAR-100 (q=0.1) | accuracy: 84.07±.16% vs 82.39±.04% | Helped |
| Use CroSel instead of PoP | WRN-34-10 on CIFAR-100 fine-grained · CIFAR-100 fine-grained | accuracy: 83.34% vs 82.04% | Helped |
| Add selection criterion cr1 (ablation) | CroSel on CIFAR-100 (q=0.1) · CIFAR-100 (q=0.1) | accuracy: 73.12% -> 84.07% | Helped |
| Add selection criteria cr1 and cr2 (ablation) | CroSel on CIFAR-100 (q=0.1) · CIFAR-100 (q=0.1) | accuracy: 72.00% -> 84.07% | Helped |
| Add selection criteria cr1, cr2, and cr3 (ablation) | CroSel on CIFAR-100 (q=0.1) · CIFAR-100 (q=0.1) | accuracy: 70.68% -> 84.07% | Helped |
| Add co-mix regularization (ablation) | CroSel on CIFAR-100 (q=0.1) · CIFAR-100 (q=0.1) | accuracy: 77.04% -> 84.07% | Helped |
| Add selection criterion cr2 (ablation) | CroSel on CIFAR-100 (q=0.1) · CIFAR-100 (q=0.1) | accuracy: 79.90% -> 84.07% | Helped |
| Apply co-mix regularization to all data rather than only unselected data (ablation) | CroSel on CIFAR-100 (q=0.1) · CIFAR-100 (q=0.1) | accuracy: 77.61% -> 84.07% | Helped |
| Add co-mix regularization on all data (ablation) | CroSel on CIFAR-100 (q=0.1) · CIFAR-100 (q=0.1) | accuracy: 70.68% -> 84.07% | Helped |
Surprising findings
- Performance does not necessarily decline monotonically with increasing label noise; the method may perform best at moderate noise levels.
- On SVHN the selection ratio drops significantly as noise increases (to about 73-76% at q=0.3-0.5), which the authors attribute to MixUp disturbing the feature space.
Limitations
- Selection ratio drops notably on SVHN as noise grows, indicating degraded selection under MixUp.
- Stricter selection criteria can be hard to satisfy early in training, requiring a warm-up and tuned t and gamma per setting.
- Results reported only on three image benchmarks with synthetic partial labels.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 SVHN
Benchmarks: CIFAR-10 CIFAR-100 SVHN
Metrics: test accuracy selection ratio selection accuracy
Baselines: PoP CRDPLL PiCO LWS PRODEN CC
Builds on: CC Co-teaching MixUp PiCO PRODEN
Closest papers by meaning
- Realistic Evaluation of Deep Partial-Label Learning Algorithms (ICLR 2025 · similarity 0.90)
- Conformal Prediction for Partial Label Learning (AAAI 2025 · similarity 0.89)
- Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning (AAAI 2024 · similarity 0.88)
- Controller-Guided Partial Label Consistency Regularization with Unlabeled Data (AAAI 2024 · similarity 0.88)
- Adaptive Integration of Partial Label Learning and Negative Learning for Enhanced Noisy Label Learning (AAAI 2024 · 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.