🧠Few-Shot Learning

NEW

Learning to perform tasks accurately from only a small number of labeled examples per class.

DESCRIPTION

Few-shot learning addresses data scarcity by enabling models to generalize from very limited labeled samples. Approaches include meta-learning, metric learning (prototypical and Siamese networks), and leveraging large pretrained models. In LLMs, in-context few-shot prompting conditions models with examples at inference time. Practical success often depends on high-quality examples, augmentation, and careful evaluation. It reduces labeling costs and accelerates adaptation to new tasks.

ALTERNATIVE DEFINITIONS

Wikipedia

Few-shot learning is a machine learning problem where the training dataset contains very few examples for the target classes.

Defines the setting as limited examples per class.

Source

Google Developers ML Glossary

A training regime in which the model must generalize from a very small number of examples.

Generalizes the concept across model types and tasks.

Source

Hugging Face

Few-shot learning uses a small number of labeled examples to guide a model to perform a task or adapt to a new domain.

Focus on practical adaptation with modern transformers.

Source

RELATED TERMS

One-shot Learning Meta-Learning Prompting In-Context Learning Prototypical Networks Siamese Networks Data Augmentation Foundation Model

COMPETING TERMS

Zero-shot Learning Fine-Tuning

KNOWLEDGE AREAS

AI Data Science

TARGET AUDIENCE

AI Engineer Data Scientist MLOps Engineer AI Product Manager

EXTERNAL SOURCES

• Wikipedia - Few-shot learning
• OpenAI - GPT-3 few-shot capability paper
• arXiv - Prototypical networks paper
• Google Developers - Glossary definition and context
• Hugging Face - Practical guides for few-shot tasks

TOP QUESTIONS

Why is Few-shot Learning important?

It reduces labeling costs and accelerates adaptation to new tasks, enabling ML in domains with scarce labeled data.

Category: Relevancy

How does Few-shot Learning differ from Transfer Learning?

Transfer learning fine-tunes pretrained models on more data; few-shot aims to generalize with minimal labeled samples, sometimes without full fine-tuning.

Category: Misunderstanding

What are common Few-shot Learning methods?

Meta-learning, metric learning (prototypical/Siamese networks), and prompt-based conditioning for large language models are popular approaches.

Category: Overview

How do LLMs perform Few-shot Learning?

By in-context learning: providing a few examples in the prompt so the model infers patterns at inference without parameter updates.

Category: Example

What are risks in Few-shot Learning?

Overfitting to tiny samples, sensitivity to example quality, and evaluation variance. Robust validation and augmentation help.

Category: Risks

How many examples count as ‘few’?

Typically 1–5 (one-shot to five-shot) per class, though definitions vary by field and task complexity.

Category: Misunderstanding

What improves Few-shot performance?

High-quality support examples, data augmentation, good initialization from large pretrained models, and carefully crafted prompts for LLMs.

Category: Requirements

How to evaluate Few-shot Learning reliably?

Use multiple episodes/splits, report variance, and test across tasks/domains. For LLMs, randomize example order and templates.

Category: Testing

What are alternatives to Few-shot Learning?

Active learning to select labels efficiently, semi-supervised learning, and transfer learning with modestly larger labeled sets.

Category: Alternatives