Research map / Label-efficient and robust learning / Semi-supervised learning
Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition
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 probabilistic framework unifying class-balanced contrastive learning methods via Gaussian kernel density estimation, and extends it to long-tailed semi-supervised learning as CCL with reliable and smoothed continuous pseudo-labels plus energy-based selection.
Problem
Long-tailed semi-supervised learning where labeled and unlabeled data have different (unknown) label distributions, causing biased pseudo-labels; existing methods neglect representation learning.
Approach
A unified probabilistic view of long-tail contrastive losses via Gaussian kernel density estimation, extended to unlabeled data with continuous pseudo-labels (energy-filtered reliable labels and label-propagated smoothed labels) and an EMA-estimated unlabeled class prior, with a dual-branch balanced/standard classifier.
Key results
- On ImageNet-127, CCL reaches 61.5% at 32x32 and 67.8% at 64x64, outperforming ACR by 4.3% and 4.2%.
- In the consistent setting, CCL improves over the prior state of the art ACR by 2.8% on average.
- Without predefined anchor distributions, CCL gains an average of 1.6% over ACR in inconsistent settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CCL instead of ACR | FixMatch on CIFAR10-LT and CIFAR100-LT · Consistent, uniform, and reversed unlabeled distributions | test accuracy: 2.8% on average (consistent); 1.6% on average (inconsistent) | Helped |
| Use CCL instead of ACR | FixMatch on ImageNet-127 · 32 × 32 and 64 × 64 | test accuracy: 4.3%; 4.2% | Helped |
| Add a dual-branch classifier (ablation) | CCL · CIFAR10-LT and CIFAR100-LT, consistent setting | test accuracy: 1.5% | Helped |
| Add continuous reliable pseudo-label loss (ablation) | CCL · CIFAR10-LT and CIFAR100-LT | test accuracy: 0.8% | Helped |
| Add continuous smoothed pseudo-label loss (ablation) | CCL · CIFAR10-LT and CIFAR100-LT, consistent, uniform, and reversed | test accuracy | Helped |
| Use energy-based selection rather than confidence-based selection | CCL unlabeled-data selection · CIFAR100-LT | class-prior estimation L1 distance | Helped |
Limitations
- Matrix inversion for smoothed pseudo-labels has O(|B^u|^3) cost.
- Relies on energy-score selection and hyperparameters (beta, thresholds) whose sensitivity is not fully characterized in the text.
Experimental setup
Datasets: CIFAR10-LT CIFAR100-LT STL10-LT ImageNet-127
Benchmarks: CIFAR10-LT CIFAR100-LT STL10-LT ImageNet-127
Metrics: test accuracy expected calibration error class prior estimation L1 error
Baselines: FixMatch DARP CReST+ DASO ACR CPE CoSSL TRAS DARP+cRT CReST++LA ABC LA
Builds on: FixMatch supervised contrastive learning balanced contrastive learning DASO ACR
Closest papers by meaning
- Decoupled Contrastive Learning for Long-Tailed Recognition (AAAI 2024 · similarity 0.91)
- Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic Long-Tailed Semi-Supervised Learning (NeurIPS 2025 · similarity 0.90)
- Three Heads Are Better than One: Complementary Experts for Long-Tailed Semi-supervised Learning (AAAI 2024 · similarity 0.88)
- Contrastive Credibility Propagation for Reliable Semi-supervised Learning (AAAI 2024 · similarity 0.88)
- Long-tailed Object Detection Pretraining: Dynamic Rebalancing Contrastive Learning with Dual Reconstruction (NeurIPS 2024 · similarity 0.88)
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · 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.