# Meta-Learning: MAML, Reptile, and Prototypical Networks
Status: public
Confidence: medium (0.78) (verified)
Last verified: 2026-05-30
Generation: ai_structured


## TL;DR

Meta-learning trains systems so they can adapt quickly to new tasks. MAML and Reptile focus on learning useful initial parameters, while Prototypical Networks learn an embedding space for few-shot classification.

## Core Explanation

The core setup is a distribution of related tasks. A meta-learner sees many training tasks, then is evaluated on new tasks from the same family with only a small support set. This differs from ordinary supervised learning because the objective is fast adaptation, not only high accuracy on one fixed dataset.

Gradient-based methods such as MAML and Reptile optimize an initialization that can move quickly after a few updates. Metric-based methods such as Prototypical Networks instead learn a representation where examples from the same class are close to a class prototype.

## Related Articles

- [Few-Shot Learning: Prototypical Networks, MAML, and In-Context Adaptation](../few-shot-learning.md)
- [Transfer Learning: Fine-Tuning and Feature Reuse](../transfer-learning.md)
- [Reinforcement Learning: Agents, Rewards, and Policies](../reinforcement-learning.md)