Research map / Label-efficient and robust learning / Semi-supervised learning
Twice Class Bias Correction for Imbalanced 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 Twice Class Bias Correction (TCBC), which corrects FixMatch's training objective with an estimated training-sample class prior and refines unlabeled pseudo-labels using an estimated current-parameter class bias, improving class-imbalanced semi-supervised learning on CIFAR10/100-LT and STL10-LT.
Problem
Class-imbalanced semi-supervised learning, where labeled data is long-tailed and the unlabeled class distribution is unknown and possibly different, causing model bias and biased pseudo-labels toward majority classes.
Approach
Two corrections: (1) logit adjustment of the training loss using a dynamically estimated class prior from participating training samples so the model learns a class-balanced posterior; (2) a momentum-estimated class bias term under current parameters that is added to logits to produce less biased pseudo-labels.
Key results
- On CIFAR10-LT with gamma_u=1, TCBC reaches 92.4% vs 73.0% for FixMatch (+19.4 points) and 87.1% for L2AC.
- On STL10-LT (N1=150), TCBC achieves 77.6% vs 75.1% for L2AC with gamma_l=10.
- Averaged over unlabeled imbalance settings on CIFAR10-LT, TCBC beats DASO by about 4.0 points (86.8 vs 82.8).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply TCBC instead of FixMatch | CIFAR10-LT, N1=1500, M1=3000 · γl=100, γu=100 | accuracy (%): 77.5 -> 84.0 | Helped |
| Apply TCBC instead of FixMatch | CIFAR10-LT, N1=1500, M1=3000 · γl=100, γu=1 | accuracy (%): 81.5 -> 92.8 | Helped |
| Apply TCBC instead of FixMatch | CIFAR100-LT · γl=10, γu=1 | accuracy (%): 58.6 -> 63.2 | Helped |
| Apply TCBC instead of FixMatch | STL10-LT, N1=450, M1=100k · γl=10, γu=N/A | accuracy (%): 72.4 -> 84.5 | Helped |
| Add model bias correction to FixMatch (ablation) | CIFAR10-LT · γu=100 | accuracy (%): 67.8 -> 79.2 | Helped |
| Add model bias correction to FixMatch (ablation) | CIFAR10-LT · γu=1 | accuracy (%): 73.0 -> 92.3 | Helped |
| Add model bias correction to FixMatch (ablation) | CIFAR10-LT · γu=1/100 | accuracy (%): 62.5 -> 79 | Helped |
| Add pseudo-label refinement to FixMatch (ablation) | CIFAR10-LT · γu=100 | accuracy (%): 67.8 -> 77.8 | Helped |
| Add pseudo-label refinement to FixMatch (ablation) | CIFAR10-LT · γu=1 | accuracy (%): 73.0 -> 92.1 | Helped |
| Add pseudo-label refinement to FixMatch (ablation) | CIFAR10-LT · γu=1/100 | accuracy (%): 62.5 -> 76.8 | Helped |
| Add pseudo-label refinement to model bias correction (ablation) | CIFAR10-LT · γu=100 | accuracy (%): 79.2 -> 80.3 | Helped |
| Add pseudo-label refinement to model bias correction (ablation) | CIFAR10-LT · γu=1 | accuracy (%): 92.3 -> 92.4 | Helped |
Surprising findings
- Pseudo-label refinement alone performs poorly when unlabeled data is imbalanced (gamma_u=100 or 1/100), requiring model bias correction to work well.
Limitations
- Pseudo-label refinement alone is unstable under imbalanced unlabeled data, depending on the model bias correction.
- Relies on estimating class distributions from participating training samples over a window of recent iterations (T=50*K), with no theoretical guarantee stated.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 STL-10 SUN397
Benchmarks: CIFAR10-LT CIFAR100-LT STL10-LT SUN397
Metrics: top-1 accuracy
Baselines: FixMatch DARP CReST+ ABC DASO L2AC Supervised
Builds on: FixMatch logit adjustment balanced softmax DASO DARP
Closest papers by meaning
- Sampling Control for Imbalanced Calibration in Semi-Supervised Learning (AAAI 2026 · similarity 0.90)
- CReST: A Class-Rebalancing Self-Training Framework for Imbalanced Semi-Supervised Learning (CVPR 2021 · similarity 0.90)
- CDMAD: Class-Distribution-Mismatch-Aware Debiasing for Class-Imbalanced Semi-Supervised Learning (CVPR 2024 · similarity 0.89)
- BaCon: Boosting Imbalanced Semi-supervised Learning via Balanced Feature-Level Contrastive Learning (AAAI 2024 · similarity 0.87)
- LCGC: Learning from Consistency Gradient Conflicting for Class-Imbalanced Semi-Supervised Debiasing (AAAI 2025 · similarity 0.87)
- Balanced Data, Imbalanced Spectra: Unveiling Class Disparities with Spectral Imbalance (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.