---
id: few-shot-learning
title: "Few-Shot Learning: Prototypical Networks, MAML, and In-Context Adaptation"
schema_type: article
category: ai
language: en
confidence: high
last_verified: "2026-05-24"
created_date: "2026-05-24"
generation_method: ai_assisted
ai_models:
  - claude-4.5-sonnet
derived_from_human_seed: true
conflict_of_interest: none_declared
is_live_document: false
data_period: static
completeness: 0.85
atomic_facts:
  - id: af-few-shot-learning-1
    statement: >-
      Few-shot learning paradigms evolved through three eras: (1) Metric learning (2015-2018) -- Prototypical Networks (ProtoNet, NeurIPS 2017) learn distance metrics and classify via
      nearest-prototype comparison, achieving 56% on miniImageNet 5-way 1-shot; (2) Optimization-based (2017-2019) -- MAML (ICML 2017) learns a model initialization that adapts to new tasks in 1-5
      gradient steps; (3) Foundation model era (2022-present) -- GPT-3/Claude perform few-shot via in-context examples without parameter updates.
    source_title: ProtoNet (Snell et al., NeurIPS 2017) / MAML (Finn et al., ICML 2017) -- foundational few-shot learning methods / Foundation model in-context learning (2022-present)
    source_url: https://arxiv.org/abs/1703.05175
    confidence: high
  - id: af-few-shot-learning-2
    statement: >-
      Foundation model few-shot learning (2023-2025) demonstrated that large models achieve 70-90% accuracy on novel classification tasks given 5-10 demonstration examples in the prompt -- without any
      parameter updates -- effectively shifting research focus from meta-learning algorithms to in-context learning mechanisms and optimal demonstration selection strategies.
    source_title: GPT-3/4 in-context learning (2020-2024) / Brown et al. -- Language Models are Few-Shot Learners / Meta-learning survey (2025)
    source_url: https://arxiv.org/abs/1703.03400
    confidence: high
primary_sources:
  - id: ps-few-shot-learning-1
    title: Prototypical Networks for Few-shot Learning
    type: academic_paper
    year: 2017
    institution: NeurIPS / University of Toronto
    url: https://arxiv.org/abs/1703.05175
  - id: ps-few-shot-learning-2
    title: Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks (MAML)
    type: academic_paper
    year: 2017
    institution: ICML / UC Berkeley
    url: https://arxiv.org/abs/1703.03400
known_gaps:
  - Optimal demonstration selection -- which examples maximize in-context learning performance
  - Adapting foundation models to entirely novel domains with no training distribution overlap
disputed_statements: []
secondary_sources:
  - title: "Meta-learning Approaches for Few-Shot Learning: A Survey of Recent Advances"
    type: survey_paper
    year: 2024
    authors:
      - Gharoun, Hassan
      - Momenifar, Fereshteh
      - Chen, Fang
      - Gandomi, Amir H.
    institution: ACM Computing Surveys
    url: https://doi.org/10.1145/3659943
  - title: Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks (MAML)
    type: conference_paper
    year: 2017
    authors:
      - Finn, Chelsea
      - Abbeel, Pieter
      - Levine, Sergey
    institution: UC Berkeley / ICML
    url: https://arxiv.org/abs/1703.03400
  - title: Prototypical Networks for Few-shot Learning
    type: conference_paper
    year: 2017
    authors:
      - Snell, Jake
      - Swersky, Kevin
      - Zemel, Richard S.
    institution: University of Toronto / NeurIPS
    url: https://arxiv.org/abs/1703.05175
  - title: Matching Networks for One Shot Learning
    type: conference_paper
    year: 2016
    authors:
      - Vinyals, Oriol
      - Blundell, Charles
      - Lillicrap, Timothy
      - Kavukcuoglu, Koray
      - Wierstra, Daan
    institution: Google DeepMind / NeurIPS
    url: https://arxiv.org/abs/1606.04080
updated: "2026-05-24"
---
## TL;DR
Few-shot learning teaches AI to recognize new concepts from just a handful of examples -- the way humans learn (see one panda, recognize all pandas). From Prototypical Networks to MAML to in-context learning in foundation models, the ability to generalize from few examples is transforming AI from narrow specialists to flexible generalists.

## Core Explanation
The N-way K-shot classification problem: given N classes with K labeled examples each, classify a new query example. Meta-learning (learning to learn): train on many small tasks (episodes), each with N-way K-shot structure. Three paradigms: (1) Metric-based -- ProtoNet computes class prototypes as mean embedding of support examples, classifies query by nearest-prototype distance; (2) Optimization-based -- MAML trains parameters such that one gradient step on a new task's loss produces good performance; (3) Hallucination-based -- generate additional training examples via data augmentation or generative models.

## Detailed Analysis
ProtoNet (Snell et al., 2017): embedding function maps inputs to a metric space where Euclidean distance corresponds to semantic similarity. Compute prototype per class, classify query as closest prototype. Works well when classes are visually distinct, struggles with fine-grained categories. MAML (Finn et al., 2017): learns initialization sensitive to task-specific gradients. Inner loop: task-specific adaptation (few gradient steps). Outer loop: meta-update across tasks. Foundation model few-shot: rather than meta-training, simply prompt a frozen LLM with examples. Key observation (Min et al., 2022): demonstration labels need not be correct -- model learns task format, not specific answers. Ground-truth labels provide marginal gain over random labels. Applications: medical imaging (rare disease classification), personalized NLP, and robotics (adapting manipulation to novel objects).

## Further Reading
- Learn2Learn: PyTorch Meta-Learning Library
- Meta-Dataset: Few-Shot Learning Benchmark
- ChatGPT Few-Shot Prompting Guide