Research map / Label-efficient and robust learning / Semi-supervised learning
OwMatch: Conditional Self-Labeling with Consistency for Open-World 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 OwMatch, combining conditional self-labeling (Sinkhorn-Knopp with labeled-data constraints) and open-world hierarchical thresholding with consistency, for open-world semi-supervised learning on CIFAR-10/100, ImageNet-100 and Tiny-ImageNet.
Problem
Open-world semi-supervised learning, where unlabeled data contains novel classes absent from labeled data, causing seen-class confirmation bias and misclassification of novel classes.
Approach
Conditions self-label assignment on ground-truth labels of labeled samples within an optimal-transport (Sinkhorn-Knopp) formulation to reduce confirmation bias and clustering misalignment, and uses a hierarchical, group- and class-wise confidence threshold to balance seen and novel learning paces; theoretically argues the conditional estimator is unbiased with lower chi-square deviation.
Key results
- On CIFAR-100 (50% novel, 50% label), OwMatch+ reaches 71.9% all-class accuracy and 63.9% novel accuracy, vs TRSSL's 64.7% all-class and 49.3% novel (OwMatch+ 63.9 novel vs 49.3 for TRSSL, about +14.6 points).
- On ImageNet-100 (50% novel, 50% label), OwMatch+ reaches 85.5% all-class accuracy, versus 83.8% for OpenCon.
- Ablation on CIFAR-100 shows conditional self-labeling raises novel accuracy from 56.7 to 58.5 and adding consistency and hierarchical thresholding brings it to 63.9.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add conditional self-labeling (ablation) | OwMatch base objective · CIFAR-100 | novel accuracy | Helped |
| add pseudo-label consistency regularization (ablation) | OwMatch with conditional self-labeling · CIFAR-100 | all accuracy | Helped |
| add open-world hierarchical thresholding (ablation) | OwMatch with conditional self-labeling and consistency · CIFAR-100 | all accuracy: 69.6 -> 71.9 | Helped |
| remove the supervised objective (ablation) | OwMatch+ · CIFAR-100 | all accuracy: 71.9 -> 70.6 | Hurt |
| exclude labeled data from online clustering (ablation) | OwMatch+ · CIFAR-100 | novel accuracy: 63.9 -> 61.2 | Hurt |
| replace OwHT with self-adaptive thresholding (ablation) | OwMatch+ · CIFAR-100 | novel accuracy: 63.9 -> 60.5 | Hurt |
| replace OwHT with static threshold 0.8 (ablation) | OwMatch+ · CIFAR-100 | all accuracy: 71.9 -> 71.7 | Hurt |
| increase local crops from 2 to 4 (ablation) | OwMatch+ · CIFAR-100 | all accuracy: 70.2 -> 71.9 | Helped |
| increase local crops from 4 to 6 (ablation) | OwMatch+ · CIFAR-100 | all accuracy: 71.9 -> 69.8 | Hurt |
| increase Sinkhorn-Knopp iterations from 3 to 10 (ablation) | OwMatch+ · CIFAR-100 | all accuracy: 67.6 -> 71.9 | Helped |
| add multi-crop augmentation | OwMatch · CIFAR-100 | all accuracy: 65.1 -> 71.9 | Helped |
| use OwMatch+ instead of OpenLDN | CIFAR-10 · CIFAR-10 | all accuracy: 95.4 -> 96.8 | Helped |
Surprising findings
- Hierarchical thresholding alone without conditional self-labeling is slightly better on CIFAR-10 seen accuracy, while the full combination is needed for best novel accuracy.
- Removing the supervised objective lowers seen accuracy substantially (80.1 to 76.8) on CIFAR-100, while removing labeled data from clustering reduces novel accuracy.
Limitations
- Assumes a known class prior for conditional self-labeling; adaptive estimation without prior loses up to about 3% all-class accuracy.
- Requires the number of novel classes to be known or estimated; estimation errors of up to 11% are present and sensitivity is studied.
- Comparison with GCD methods using ViT-B/16 is not direct because backbones differ.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-100 Tiny-ImageNet
Benchmarks: CIFAR-10 CIFAR-100 ImageNet-100 Tiny-ImageNet
Metrics: seen accuracy novel accuracy (Hungarian-matched clustering) all-class accuracy
Baselines: FixMatch DS3L CGDL DTC RankStats SimCLR UNO ORCA NACH OpenLDN TRSSL OpenCon
Builds on: FixMatch TRSSL SwAV FreeMatch ORCA
Closest papers by meaning
- Scaling Up Semi-supervised Learning with Unconstrained Unlabelled Data (AAAI 2024 · similarity 0.89)
- Open-World Semi-Supervised Learning (ICLR 2022 · similarity 0.89)
- AlphaMatch: Improving Consistency for Semi-Supervised Learning With Alpha-Divergence (CVPR 2021 · similarity 0.89)
- Boosting Semi-Supervised Learning by Exploiting All Unlabeled Data (CVPR 2023 · similarity 0.89)
- Class-Aware Contrastive Semi-Supervised Learning (CVPR 2022 · similarity 0.89)
- 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.