Research map / Label-efficient and robust learning / Semi-supervised learning
BEM: Balanced and Entropy-based Mix for 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 BEM, a data-mixing plug-in for long-tailed semi-supervised learning that uses a class-balanced mix bank with CAM-based mixing and entropy-based sampling, selection, and loss to re-balance both data quantity and class-wise uncertainty.
Problem
Long-tailed semi-supervised learning (LTSSL) where in-batch data mixing fails to re-balance classes, and existing methods ignore class-wise uncertainty (entropy) when balancing performance across classes.
Approach
Replaces in-batch mixing with a class-balanced mix bank sampled by effective-number quantity estimates, uses CAM-based localization (CamMix) to paste class-relevant regions, and integrates class-wise entropy into the sampling, sample selection between labeled and unlabeled data, and loss weighting.
Key results
- BEM with FixMatch+ACR reaches 83.5% on CIFAR10-LT (gamma=100, N1=500, M1=4000) vs 81.6% for FixMatch+ACR.
- Adding BEM to FixMatch improves average accuracy by 11.8% (per introduction), with gains on FixMatch+LA of 4.4%, FixMatch+ABC of 1.4%, and FixMatch+ACR of 2.5%.
- On ImageNet-127 (32x32), BEM reaches 53.3% vs 29.7% for FixMatch; BEM+ACR reaches 58.0%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add BEM to FixMatch+ACR | FixMatch+ACR · CIFAR10-LT (γ=100, N1=500, M1=4000) | test accuracy: 81.6±0.19 -> 83.5±0.33 | Helped |
| Add BEM to FixMatch+ACR | FixMatch+ACR · CIFAR100-LT (γ=10, N1=50, M1=400) | test accuracy: 55.7±0.12 -> 55.8±0.32 | No clear effect |
| Add BEM to FixMatch+ACR | FixMatch+ACR · CIFAR10-LT (γl=100, γu=1/100, N1=500, M1=4000) | test accuracy: 85.0±0.09 -> 85.5±0.21 | Helped |
| Add BEM to ACR | ACR · ImageNet-127 (32 × 32) | test accuracy: 57.2 -> 58.0 | Helped |
| Add BEM to MeanTeacher | MeanTeacher · CIFAR10-LT | test accuracy: 68.6±0.88 -> 73.5±0.56 | Helped |
| Replace in-batch mixing with CamMix | FixMatch · CIFAR10-LT | test accuracy: 70.8 -> 71.9 | Helped |
| Add CamMix (ablation) | BEM on CIFAR10-LT · CIFAR10-LT | test accuracy: 74.0 -> 75.7 | Helped |
| Add CBMB (class-balanced sampling rather than random sampling) (ablation) | BEM on CIFAR10-LT · CIFAR10-LT | test accuracy: 72.1 -> 75.7 | Helped |
| Add ESS (ablation) | BEM on CIFAR10-LT · CIFAR10-LT | test accuracy: 74.7 -> 75.7 | Helped |
| Add ESM (ablation) | BEM on STL10-LT · STL10-LT | test accuracy: 65.3 -> 68.3 | Helped |
| Add ECB loss (ablation) | BEM on CIFAR10-LT · CIFAR10-LT | test accuracy: 74.9 -> 75.7 | Helped |
Surprising findings
- BEM is less effective than DASO under the reversed unlabeled-distribution setting on CIFAR10-LT, which the authors attribute to data mixing not achieving thorough re-balancing in challenging scenarios.
Limitations
- Less effective than DASO in the reversed unlabeled distribution setting on CIFAR10-LT.
- Relies on estimating unknown unlabeled class distribution and entropy on the fly via EMA, with hyperparameters (alpha, beta, EMA weights) that require tuning.
Experimental setup
Datasets: CIFAR10-LT CIFAR100-LT STL10-LT ImageNet-127
Benchmarks: CIFAR10-LT CIFAR100-LT STL10-LT ImageNet-127
Metrics: top-1 test accuracy
Baselines: FixMatch FixMatch+LA FixMatch+ABC FixMatch+ACR DARP CReST CReST+ DASO MixUp CutMix SaliencyMix MeanTeacher
Builds on: FixMatch SaliencyMix class-balanced loss DASO ACR
Closest papers by meaning
- SimPro: A Simple Probabilistic Framework Towards Realistic Long-Tailed Semi-Supervised Learning (ICML 2024 · similarity 0.87)
- Three Heads Are Better than One: Complementary Experts for Long-Tailed Semi-supervised Learning (AAAI 2024 · similarity 0.86)
- Long-Tailed Learning as Multi-Objective Optimization (AAAI 2024 · similarity 0.86)
- Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition (NeurIPS 2024 · similarity 0.86)
- LT-Soups: Bridging Head and Tail Classes via Subsampled Model Soups (NeurIPS 2025 · similarity 0.85)
- Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic Long-Tailed Semi-Supervised Learning (NeurIPS 2025 · similarity 0.85)
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.