Semi-supervised Few-shot Learning
Semi-supervised Few-shot Learning (SS-FSL) · Also known as: SS-FSL, semi-supervised meta-learning, few-shot learning with unlabeled data, low-label few-shot learning
Semi-supervised Few-shot Learning (SS-FSL) trains models to classify new classes from only a handful of labeled examples per class, while simultaneously leveraging a pool of unlabeled data to enrich class representations. By combining meta-learning episodes with soft pseudo-label assignment for unlabeled samples, it achieves notably higher accuracy than purely supervised few-shot methods when abundant unlabeled data is available.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use SS-FSL when labeled data is extremely scarce — typically 1 to 20 examples per novel class — but a larger pool of unlabeled in-domain data is accessible. It is well-suited to medical imaging, rare-species recognition, low-resource NLP, and any setting where annotation is costly. It outperforms pure few-shot learning when the unlabeled pool is genuinely from a related distribution. Do NOT use it when unlabeled data comes from a very different domain (distribution mismatch degrades prototypes), when moderate labeled data exists (standard supervised learning will outperform it), or when the number of novel classes at test time greatly exceeds the episode design used in training.
Strengths & limitations
- Leverages cheap unlabeled data to improve class representations, reducing annotation cost substantially.
- Higher accuracy than purely supervised few-shot methods in matched-distribution settings.
- Episodic training explicitly targets the few-shot generalization problem rather than assuming large-class coverage.
- Prototype-based decision rule is interpretable and computationally lightweight at inference time.
- Naturally extends standard few-shot frameworks (Prototypical Networks, MAML) without architectural overhaul.
- Sensitive to distribution mismatch: if unlabeled examples are out-of-domain, noisy pseudo-labels corrupt prototypes rather than refining them.
- Episode design (N-way, K-shot) must align with the intended test scenario; misalignment reduces transfer.
- Requires a meta-training base set with many labeled classes to learn a generalizable embedding, which may not be available in truly data-scarce domains.
- Soft-label assignment and prototype refinement add hyperparameters (temperature, unlabeled weighting) that need tuning.
- Evaluation is sensitive to the random sampling of episodes; reported accuracy should aggregate over many test episodes.
Frequently asked
How is this different from standard few-shot learning?
Standard few-shot learning uses only the small labeled support set to classify queries. SS-FSL additionally incorporates unlabeled examples during the episode, using soft pseudo-labels to refine class prototypes and thereby improve classification accuracy when a relevant unlabeled pool is available.
What if my unlabeled data contains classes not seen at training time?
SS-FSL is designed to handle distractors — unlabeled examples that may not belong to any of the N episode classes. The soft-label mechanism naturally down-weights or spreads probability across classes for such outliers, but very many distractors can still degrade prototype quality. Monitoring the ratio of distractors to true class unlabeled examples is advisable.
How many unlabeled examples per episode are needed?
Empirically, a moderate number (5–20 unlabeled examples per class) provides meaningful prototype refinement without introducing excessive noise. Too few offer little benefit over pure few-shot learning; too many can amplify the impact of mislabeled or out-of-distribution examples.
Can I use a pre-trained backbone like ResNet or ViT?
Yes, and in practice pre-trained backbones substantially improve performance by providing a strong initial feature space. Fine-tuning should be done carefully — typically with a low learning rate — to retain the general representations while adapting to the target domain.
What datasets are standard benchmarks for SS-FSL?
miniImageNet and tieredImageNet are the most common benchmarks, with results typically reported for 5-way 1-shot and 5-way 5-shot settings. tieredImageNet's class hierarchy allows a cleaner evaluation of generalization by ensuring train and test classes are semantically disjoint.
Sources
- Ren, M., Triantafillou, E., Ravi, S., Snell, J., Swersky, K., Tenenbaum, J. B., Larochelle, H., & Zemel, R. S. (2018). Meta-learning for semi-supervised few-shot classification. In International Conference on Learning Representations (ICLR 2018). link ↗
- Finn, C., Abbeel, P., & Levine, S. (2017). Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the 34th International Conference on Machine Learning (ICML 2017), PMLR 70, 1126–1135. link ↗
How to cite this page
ScholarGate. (2026, June 3). Semi-supervised Few-shot Learning (SS-FSL). ScholarGate. https://scholargate.app/en/machine-learning/semi-supervised-few-shot-learning
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Few-shot LearningMachine learning↔ compare
- Self-supervised LearningMachine learning↔ compare
- Semi-supervised LearningMachine learning↔ compare
- Transfer LearningMachine learning↔ compare