Skip to contentScholarGate
LibraryBookshelfDeskReview StudioAssistant
Sign in
On this page
IntuitionHow it worksWhen to use itStrengths & limitationsCommon pitfallsApplicationsFrequently asked🔒 Read the full methodSourcesRelated methods
Cite this pageSpotted an issue on this page? Report or suggest a fix →
Home›Machine learning›Self-supervised Few-shot Learning
Machine learningMachine learning

Self-supervised Few-shot Learning

Self-supervised Few-shot Learning (SSL-FSL) · Also known as: SSL-FSL, self-supervised meta-learning, unsupervised few-shot learning, self-supervised prototypical learning

Self-supervised Few-shot Learning (SSL-FSL) combines self-supervised pretraining on large unlabeled corpora with few-shot meta-learning so that a model can recognize new categories from only a handful of labeled examples. By learning rich, transferable representations without expensive annotation, SSL-FSL addresses the fundamental bottleneck of supervised few-shot methods: the need for labeled support data at scale.

ScholarGate
  1. Machine learning
  2. v1
  3. 2 Sources
  4. PUBLISHED
Cite this page →
Tools & resources
Download slides
Learn & explore

Read the full method

Members only

Sign in with a free account to read this section.

Sign in

Method map

The neighbourhood of related methods — select a node to explore.

Self-supervised Few-shot Learning
Siamese NetworkTransfer LearningSelf-supervised Transfer…

When to use it

Use SSL-FSL when labeled data is extremely scarce (5–20 examples per class) but unlabeled data is plentiful, such as in medical imaging, rare-event detection, or novel-category recognition. It is ideal when annotation cost is prohibitive. Avoid SSL-FSL when you have abundant labeled data — standard supervised learning will outperform it. Also avoid it when computational resources for large-scale pretraining are unavailable, or when the unlabeled pool is small, out-of-domain, or low-quality relative to the target task.

Strengths & limitations

Strengths
  • Leverages abundant unlabeled data, drastically reducing dependence on expensive labeled examples.
  • Self-supervised pretraining produces rich, generalizable representations that transfer well across novel categories.
  • Naturally compatible with contrastive and generative SSL frameworks (SimCLR, BYOL, MAE), enabling plug-and-play backbone upgrades.
  • Reduces overfitting risk in few-shot regimes by initializing from a well-regularized pretrained state.
  • Applicable across modalities — images, text, audio, and graphs — wherever unlabeled data is abundant.
  • Can be combined with meta-learning algorithms (Prototypical Networks, MAML) for further gains.
Limitations
  • Pretraining on large unlabeled corpora is computationally expensive and memory-intensive.
  • Performance degrades when the unlabeled pretraining domain is mismatched with the target few-shot domain.
  • Choosing effective SSL auxiliary tasks for non-vision modalities is non-trivial and task-specific.
  • Evaluation is sensitive to episode construction protocol; results can be hard to compare across papers.
  • Offers no interpretable coefficients; explaining individual predictions requires additional post-hoc tools.

Frequently asked

How does SSL-FSL differ from standard transfer learning?

Standard transfer learning fine-tunes a pretrained model on a (typically larger) labeled target dataset. SSL-FSL specifically targets scenarios where only a handful of labeled examples per class exist at test time, and it explicitly trains or evaluates using episodic protocols that simulate this constraint.

Which self-supervised objectives work best for few-shot learning?

Contrastive objectives (SimCLR, MoCo, BYOL) generally produce the strongest transferable representations for visual few-shot tasks. For language, masked language modeling (BERT-style) backbones are standard. The best choice depends on the modality and available compute.

Do I need a meta-learning algorithm on top of self-supervised pretraining?

Not always. Several works show that a strong self-supervised backbone with a simple nearest-centroid classifier can match or exceed sophisticated meta-learners. Meta-learning adds value when the distribution shift between base and novel classes is large.

How should I report results for SSL-FSL experiments?

Report the backbone architecture, SSL objective, pretraining dataset, episode configuration (N-way K-shot), number of test episodes, and 95% confidence intervals. Comparing against a supervised pretrained baseline is essential to isolate the contribution of self-supervision.

Is SSL-FSL suitable when I have no unlabeled data at all?

No. If no unlabeled data is available from the target domain or a related domain, the SSL pretraining step provides no benefit. In that case, standard few-shot methods with a supervised pretrained backbone on a large external dataset (e.g., ImageNet) are a more practical starting point.

Sources

  1. Gidaris, S., Bursuc, A., Komodakis, N., Perez, P., & Cord, M. (2019). Boosting Few-Shot Visual Learning with Self-Supervision. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 8059–8068. DOI: 10.1109/ICCV.2019.00815 ↗
  2. Su, J.-C., Maji, S., & Hariharan, B. (2020). When Does Self-Supervision Improve Few-Shot Learning? European Conference on Computer Vision (ECCV), Lecture Notes in Computer Science, vol 12371, 645–660. DOI: 10.1007/978-3-030-58571-6_38 ↗

How to cite this page

ScholarGate. (2026, June 3). Self-supervised Few-shot Learning (SSL-FSL). ScholarGate. https://scholargate.app/en/machine-learning/self-supervised-few-shot-learning

Related methods

Siamese NetworkTransfer 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.

  • Siamese NetworkDeep learning↔ compare
  • Transfer LearningMachine learning↔ compare
Compare side by side →

Referenced by

Self-supervised Transfer learning

Similar methods

Semi-supervised Few-shot LearningSelf-supervised Image ClassificationSelf-supervised Transfer learningFew-shot LearningSelf-supervised LearningSelf-supervised Active LearningSelf-supervised Metric learningRegularized Few-Shot Learning

Related reference concepts

Self-Supervised and Representation LearningUnsupervised LearningSupervised LearningDeep LearningMachine LearningReinforcement Learning

Spotted an issue on this page? Report or suggest a fix →

ScholarGate — Self-supervised Few-shot Learning (Self-supervised Few-shot Learning (SSL-FSL)). Retrieved 2026-07-21 from https://scholargate.app/en/machine-learning/self-supervised-few-shot-learning · Dataset: https://doi.org/10.5281/zenodo.20539026
Quick facts
Originator
Gidaris, S. et al.; Su, J.-C. et al. (concurrent seminal works)
Year
2019
Type
Hybrid learning paradigm (self-supervised pretraining + few-shot adaptation)
DataType
Image, text, or multimodal data with abundant unlabeled and scarce labeled examples
Subfamily
Machine learning
Related methods
Siamese NetworkTransfer Learning
ScholarGate

A content-first reference library for research methods — what each one is, how it works, and where it comes from.

Open data (CC-BY)

Explore

  • Library
  • Search the library…
  • Browse by field
  • Fields
  • Journey
  • Compare
  • Which method?

Reference

  • Subjects
  • Atlas
  • Glossary
  • Methodology
  • Philosophy

Your tools

  • Bookshelf
  • Desk
  • Chat

Company

  • About
  • Pricing
  • Contact
  • Suggest a method

Entries are compiled from published sources for reference. Verifying the accuracy and suitability of any information for your own use remains your responsibility.

© 2026 ScholarGate · A research-method reference library
  • Privacy
  • Cookies
  • Terms
  • Delete account