Research map / Label-efficient and robust learning / Semi-supervised learning
BaCon: Boosting Imbalanced Semi-supervised Learning via Balanced Feature-Level Contrastive 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 BaCon, a balanced feature-level contrastive learning plug-in (class-wise feature-center positives, reliable negative selection, and distribution-aware temperature) that improves the instance-level CISSL method ABC on imbalanced semi-supervised benchmarks.
Problem
Class-imbalanced semi-supervised learning, where biased pseudo-labels and biased backbone representations under long-tailed labeled and unlabeled distributions degrade accuracy on balanced test sets.
Approach
Directly regularizes the feature distribution with a contrastive loss whose positives are class-wise feature centers from a memory bank of confident samples, whose negatives are chosen by a reliable negative selection rule, and whose temperature is scaled per class by the class size ratio and decayed over training.
Key results
- On CIFAR10-LT (gamma=100, beta=20%), FixMatch-based BaCon reaches 84.46% balanced accuracy vs. 83.25% for ABC and 84.09% for CoSSL.
- On CIFAR100-LT (gamma=20, beta=40%), FixMatch-based BaCon achieves 57.96%, 0.63 points above CoSSL.
- On CIFAR10-LT with inversely proportional labeled/unlabeled imbalance (gamma_L=1/gamma_U=100), BaCon reaches 83.80% vs. 71.99% for CoSSL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add BaCon to FixMatch | FixMatch · CIFAR10-LT | balanced accuracy: 84.09±0.16 -> 84.46±0.15 | Helped |
| Add BaCon to FixMatch | FixMatch · CIFAR100-LT | balanced accuracy: 57.33±0.05 -> 57.96±0.26 | Helped |
| Add BaCon to FixMatch | FixMatch · STL10-LT | balanced accuracy: 71.23±0.04 -> 71.55±0.09 | Helped |
| Add BaCon to ReMixMatch | ReMixMatch · CIFAR10-LT | balanced accuracy: 84.93±0.02 -> 85.05±0.09 | Helped |
| Add BaCon to ReMixMatch | ReMixMatch · CIFAR100-LT | balanced accuracy: 60.46±0.15 -> 60.15±0.05 | Hurt |
| Add BaCon to ReMixMatch | ReMixMatch · STL10-LT | balanced accuracy: 68.73±0.77 -> 69.26±0.83 | Helped |
| Add contrastive loss with RNS (ablation) | FixMatch + ABC on CIFAR10-LT · CIFAR10-LT | accuracy: 83.95 -> 84.30 | Helped |
| Use naive BTA without RNS (ablation) | BaCon on CIFAR10-LT · CIFAR10-LT | accuracy: 84.30 -> 83.87 | Hurt |
| Use naive BTA without iteration decay (ablation) | BaCon on CIFAR10-LT · CIFAR10-LT | accuracy: 84.30 -> 84.15 | Hurt |
| Add iteration decay to BTA (ablation) | BaCon on CIFAR10-LT · CIFAR10-LT | accuracy: 84.15 -> 84.61 | Helped |
| Replace 32-D linear projection with identity mapping (ablation) | BaCon on CIFAR10-LT · CIFAR10-LT | accuracy: 84.61 -> 82.86 | Hurt |
| Replace 32-D linear projection with nonlinear projection (ablation) | BaCon on CIFAR10-LT · CIFAR10-LT | accuracy: 84.61 -> 83.95 | Hurt |
Surprising findings
- Training the auxiliary classifier on a frozen balanced representation gave 88.60% accuracy versus 82.48% on an imbalanced representation, suggesting the biased backbone representation is a major bottleneck.
- CoSSL drops by 11.95 points when labeled and unlabeled imbalance ratios are inversely proportional, while BaCon stays stable.
Limitations
- Gains over the strongest baselines are small on several settings (e.g., below 1 point on CIFAR10-LT and STL10-LT with ReMixMatch, and CoSSL is better on CIFAR100-LT with ReMixMatch).
- Several hyperparameters (threshold 0.98, n=3, projection dimension 32, lambda, eta) are fixed by default and tuned on the evaluation setting, with limited sensitivity analysis.
- Nonlinear projection and identity mapping underperform the linear 32-D projection, with the explanation given only as a conjecture.
Experimental setup
Datasets: CIFAR10 CIFAR100 STL10 SVHN
Benchmarks: CIFAR10-LT CIFAR100-LT STL10-LT SVHN-LT
Metrics: balanced accuracy
Baselines: FixMatch ReMixMatch DASO DebiasPL CReST CReST+PDA DARP Adsh SAW ABC CoSSL
Builds on: FixMatch ABC ReMixMatch InfoNCE USB
Closest papers by meaning
- A Tale of Two Classes: Adapting Supervised Contrastive Learning to Binary Imbalanced Datasets (CVPR 2025 · similarity 0.88)
- ConR: Contrastive Regularizer for Deep Imbalanced Regression (ICLR 2024 · similarity 0.88)
- Twice Class Bias Correction for Imbalanced Semi-supervised Learning (AAAI 2024 · similarity 0.87)
- CDMAD: Class-Distribution-Mismatch-Aware Debiasing for Class-Imbalanced Semi-Supervised Learning (CVPR 2024 · similarity 0.87)
- Three Heads Are Better than One: Complementary Experts for Long-Tailed Semi-supervised Learning (AAAI 2024 · similarity 0.86)
- LCGC: Learning from Consistency Gradient Conflicting for Class-Imbalanced Semi-Supervised Debiasing (AAAI 2025 · similarity 0.86)
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.