Research map / Label-efficient and robust learning / Noisy labels
Imprecise Label Learning: A Unified Framework for Learning with Various Imprecise Label Configurations
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 Imprecise Label Learning (ILL), an EM framework that treats precise labels as latent variables to unify partial label, semi-supervised, and noisy label learning, including their mixtures, with closed-form objectives.
Problem
Learning with reduced labeling standards (partial, noisy, or missing labels), where existing methods are specialized per imprecise-label configuration and fail when multiple configurations coexist.
Approach
Model the imprecise label information as a restriction on latent true labels and maximize its likelihood with EM, so the E-step posterior over all labelings entailed by the information gives soft targets; the imprecise-label type only changes how the posterior is computed, yielding closed-form consistency-style losses for PLL, SSL, NLL and mixtures.
Key results
- On partial label learning, the method reaches 96.37% on CIFAR-10 (q=0.1), exceeding PiCO by about 2% on average across CIFAR-10 partial ratios.
- On mixed partial-noisy labels with CIFAR-10 (50k labels, q=0.1, eta=0.1), the method reaches 96.47% vs 95.83% for the best baseline DALI Mixup.
- On WebVision, the method outperforms the previous best SOP-based result by 2.05%, reaching 79.37%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ILL instead of PiCO for partial-label learning | Partial-label learning · CIFAR-10 and CIFAR-100 | accuracy: 2.13% on CIFAR-10 and 2.72% on CIFAR-100 | Helped |
| Use ILL instead of the strongest reported baseline | Semi-supervised learning · STL-10 with 40 labels and Amazon Review with 250 labels | error rate: 0.68% and 1.33% | Helped |
| Use ILL instead of prior noisy-label methods | Noisy-label learning · CIFAR-10 synthetic symmetric and asymmetric noise | accuracy | Helped |
| Use ILL instead of SOP | Noisy-label learning · CIFAR-100 asymmetric noise | accuracy | Hurt |
| Use ILL instead of the previous best method | Noisy-label learning · WebVision | accuracy: 2.05% | Helped |
| Use ILL instead of DivideMix | Noisy-label learning · Clothing1M | accuracy | No clear effect |
| Use ILL instead of partial-noisy-label baselines | Mixed imprecise label learning · CIFAR-10 and CIFAR-100 with partial, noisy, and unlabeled data | accuracy | Helped |
Surprising findings
- On CIFAR-10 and CIFAR-100 partial label learning, the method outperforms the fully-supervised reference.
- The method does not use thresholding, re-weighting, or contrastive techniques for SSL yet remains competitive.
Limitations
- Evaluated only on relatively small-scale datasets; scalability to larger datasets is unverified.
- Only balanced datasets considered; imbalanced and open-set settings are unexplored.
- The simplified instance-independent noise transition model is oversimplified, causing weaker results on CIFAR-100 asymmetric noise.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 CUB-200 STL-10 IMDB Amazon Review Clothing1M WebVision CIFAR-10N
Benchmarks: CIFAR-10 CIFAR-100 CUB-200 STL-10 IMDB Amazon Review Clothing1M WebVision
Metrics: accuracy error rate
Baselines: LWS PRODEN CC PiCO AdaMatch FixMatch FlexMatch CoMatch SimMatch FreeMatch SoftMatch DivideMix
Builds on: PiCO FixMatch SOP USB EM algorithm
Closest papers by meaning
- Adaptive Integration of Partial Label Learning and Negative Learning for Enhanced Noisy Label Learning (AAAI 2024 · similarity 0.89)
- Learning with Partial-Label and Unlabeled Data: A Uniform Treatment for Supervision Redundancy and Insufficiency (ICML 2024 · similarity 0.88)
- Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning (AAAI 2024 · similarity 0.87)
- Label Noise: Ignorance Is Bliss (NeurIPS 2024 · similarity 0.87)
- Learning with Complementary Labels Revisited: The Selected-Completely-at-Random Setting Is More Practical (ICML 2024 · similarity 0.87)
- ULAREF: A Unified Label Refinement Framework for Learning with Inaccurate Supervision (ICML 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.