Research map / Label-efficient and robust learning / Recognition and long-tail
Long-Tailed Partial Label Learning by Head Classifier and Tail Classifier Cooperation
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 a head-classifier/tail-classifier cooperation method with a learned classifier weight estimation module for long-tailed partial label learning, trained from scratch on ResNet-18.
Problem
Partial label learning where each instance has candidate labels and the class distribution is long-tailed, so existing PLL methods are biased toward head classes and sacrifice tail performance (or vice versa).
Approach
Train two classifiers on a shared backbone: a head classifier (imbalanced sampling with calibrated logits) and a tail classifier (uniform sampling with logit adjustment), derive soft pseudo-labels from the moving average of the tail classifier's output and class distribution from them, and fuse the two via a learned classifier weight estimation module that infers per-sample head/tail belongingness
Key results
- On CIFAR10 with η=0.5 and γ=200, the method improves over the best compared method by 13.46%.
- On PASCAL VOC, overall accuracy 67.45% vs. 61.83% for the best compared method (SoLar), a 5.62% gain.
- Ablation: using only the tail classifier gives 69.31% and only the head classifier 68.12% on CIFAR10 (η=0.3, γ=200), versus 81.97% for the full model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use cooperating head and tail classifiers with CWE fusion instead of only the tail classifier (ablation) | Long-tailed partial label learning · CIFAR10 (η = 0.3, γ = 200) | overall accuracy: 69.31 -> 81.97 | Helped |
| use cooperating head and tail classifiers with CWE fusion instead of only the head classifier (ablation) | Long-tailed partial label learning · CIFAR10 (η = 0.3, γ = 200) | overall accuracy: 68.12 -> 81.97 | Helped |
| add soft-pseudo-label losses (ablation) | Long-tailed partial label learning · CIFAR10 (η = 0.3, γ = 200) | overall accuracy: 63.49 -> 81.97 | Helped |
| add mixup (ablation) | Long-tailed partial label learning · CIFAR10 (η = 0.3, γ = 200) | overall accuracy: 72.32 -> 81.97 | Helped |
| add logit adjustment (ablation) | Long-tailed partial label learning · CIFAR10 (η = 0.3, γ = 200) | overall accuracy: 73.79 -> 81.97 | Helped |
| add reliable-sample selection (ablation) | Long-tailed partial label learning · CIFAR10 (η = 0.3, γ = 200) | overall accuracy: 74.25 -> 81.97 | Helped |
| use head-tail classifier cooperation with CWE | Long-tailed partial label learning · CIFAR10 (η = 0.5, γ = 200) | accuracy: 64.25 ± 2.02 -> 77.71 ± 1.12 | Helped |
| use head-tail classifier cooperation with CWE | Long-tailed partial label learning · CIFAR100 (η = 0.1, γ = 50) | accuracy: 45.47 ± 0.74 -> 51.26 ± 1.31 | Helped |
| use head-tail classifier cooperation with CWE | Long-tailed partial label learning · PASCAL VOC | overall accuracy: 61.83 -> 67.45 | Helped |
Limitations
- Relies on a fixed set of hyperparameters (mixup ζ, µ, α and β ramp schedules, τ) chosen per imbalance regime.
- Evaluation limited to CIFAR10/100 with synthetic candidate-label generation and one real-world PLL dataset.
Experimental setup
Datasets: CIFAR10 CIFAR100 PASCAL VOC 2007
Benchmarks: CIFAR10-LT CIFAR100-LT PASCAL VOC (LT-PLL)
Metrics: classification accuracy many/medium/few shot accuracy
Baselines: LW CC PRODEN CAVL PiCO RECORDS SoLar
Builds on: PRODEN PiCO SoLar RECORDS logit adjustment mixup
Closest papers by meaning
- Three Heads Are Better than One: Complementary Experts for Long-Tailed Semi-supervised Learning (AAAI 2024 · similarity 0.87)
- Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic Long-Tailed Semi-Supervised Learning (NeurIPS 2025 · similarity 0.87)
- Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition (NeurIPS 2024 · similarity 0.87)
- Feature Fusion from Head to Tail for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.87)
- What Makes Partial-Label Learning Algorithms Effective? (NeurIPS 2024 · similarity 0.87)
- Long-Tailed Learning as Multi-Objective Optimization (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.