Research map / Label-efficient and robust learning / Semi-supervised learning
Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic Long-Tailed Semi-Supervised 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 CPG, a controllable pseudo-label generation framework for realistic long-tailed semi-supervised learning under arbitrary unlabeled distributions, combining reliable pseudo-label filtering, logit-adjusted training, class-aware augmentation and an auxiliary consistency branch.
Problem
Long-tailed semi-supervised learning where the unlabeled data distribution is unknown and may be arbitrary, causing confirmation bias from erroneous high-confidence pseudo-labels and degraded minority-class performance.
Approach
Expand the labeled set with reliable pseudo-labels selected by a dynamic filter (confidence thresholds on weak and strong views, agreement, and voting), so the updated labeled set has a known class distribution; then train a logit-adjusted (Bayes-optimal) classifier on it, which in turn yields more reliable pseudo-labels in a self-reinforcing cycle.
Key results
- On CIFAR-10-LT (N_max=400, M_max=4600, gamma=100) with arbitrary unlabeled distribution, CPG outperforms baselines by 15.97 pp, reaching 82.10% vs. lower baselines.
- Averaged over settings, CPG gains 11.14 pp on CIFAR-10-LT, 3.09 pp on CIFAR-100-LT, 4.06 pp on SVHN-LT and 2.57 pp on Food-101-LT.
- Ablation on CIFAR-10-LT: CSOC alone adds 14.41 pp under arbitrary distribution; CAA+CSOC combined gives 14.82 pp, average gain 10.65 pp.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CPG rather than the strongest listed baseline, SoftMatch | CIFAR-10-LT · CIFAR-10-LT (average across distributions and imbalance ratios) | accuracy (%): 76.41 vs 65.27 | Helped |
| Add CPG rather than the strongest listed baseline, FixMatch | CIFAR-100-LT · CIFAR-100-LT (average across distributions and imbalance ratios) | accuracy (%): 47.25 vs 44.16 | Helped |
| Add CPG rather than the strongest listed baseline, SoftMatch | Food-101-LT · Food-101-LT (average across distributions and imbalance ratios) | accuracy (%): 23.67 vs 21.10 | Helped |
| Add CPG rather than the strongest listed baseline, SimPro | SVHN-LT · SVHN-LT (average across distributions) | accuracy (%): 94.06 vs 88.38 | Helped |
| Add CPG rather than SimPro | ESC-50-LT · ESC-50-LT (long-tailed labeled, arbitrary unlabeled) | accuracy (%): 69.13 vs 68.75 | Helped |
| Add CPG rather than SimPro | ImageNet-127 · ImageNet-127 (32×32) | accuracy (%): 44.58 vs 43.31 | Helped |
| Add CPG rather than ACR | ImageNet-127 · ImageNet-127 (64×64) | accuracy (%): 50.43 vs 47.61 | Helped |
| Add the auxiliary branch (AB) to CPG without AB, CSOC, and CAA (ablation) | CPG · CIFAR-10-LT (γ=100, arbitrary unlabeled) | accuracy (%): 65.18 -> 65.98 (+0.80 pp) | Helped |
| Add class-aware adaptive augmentation (CAA) to the AB baseline (ablation) | CPG with AB · CIFAR-10-LT (γ=100, arbitrary unlabeled) | accuracy (%): 65.98 -> 68.97 (+2.99 pp) | Helped |
| Add the controllable self-reinforcing optimization cycle (CSOC) to the AB baseline (ablation) | CPG with AB · CIFAR-10-LT (γ=100, arbitrary unlabeled) | accuracy (%): 65.98 -> 80.39 (+14.41 pp) | Helped |
| Add CAA and CSOC together to the AB baseline (ablation) | CPG with AB · CIFAR-10-LT (γ=100, arbitrary unlabeled) | accuracy (%): 65.98 -> 82.33 (+14.82 pp) | Helped |
| Add CAA to the AB+CSOC configuration (ablation) | CPG · CIFAR-100-LT (γ=10, arbitrary unlabeled) | accuracy (%): 49.25 -> 51.85 | Helped |
Surprising findings
- Supervised-only LA training upper-bounds several SSL/ReaLTSSL baselines by a wide margin on CIFAR-10-LT, while CPG's gains are largest under arbitrary unlabeled distributions.
- Reported win/tie/loss counts show CPG wins in 130 of 162 comparisons with zero losses.
Limitations
- Theoretical guarantee relies on assumptions (Lipschitz, bounded loss, pseudo-label error rate) and is loose in practice.
- Pseudo-label reliability is limited in early training, so an auxiliary consistency branch is needed to stabilize learning.
Experimental setup
Datasets: CIFAR-10-LT CIFAR-100-LT Food-101-LT SVHN-LT ImageNet-127 ESC-50-LT
Benchmarks: CIFAR-10-LT CIFAR-100-LT Food-101-LT SVHN-LT ImageNet-127 ESC-50-LT
Metrics: accuracy macro-F1 win/tie/loss via pairwise t-test
Baselines: FixMatch FreeMatch SoftMatch ACR SimPro CDMAD Supervised CE Supervised LA
Builds on: FixMatch logit adjustment SimPro USB
Closest papers by meaning
- Semi-Supervised Regression with Heteroscedastic Pseudo-Labels (NeurIPS 2025 · similarity 0.91)
- Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition (NeurIPS 2024 · similarity 0.90)
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · similarity 0.90)
- Bi-Level Optimization for Semi-Supervised Learning with Pseudo-Labeling (AAAI 2025 · similarity 0.90)
- Three Heads Are Better than One: Complementary Experts for Long-Tailed Semi-supervised Learning (AAAI 2024 · similarity 0.89)
- SimPro: A Simple Probabilistic Framework Towards Realistic Long-Tailed Semi-Supervised Learning (ICML 2024 · 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.