Skip to content

Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks (MAML)

Authors: Chelsea Finn, Pieter Abbeel, Sergey Levine Year: 2017 Venue: ICML 2017 Citations: 15,000+

Summary

Introduces MAML, a model-agnostic meta-learning algorithm for fast adaptation to new tasks. Learns initial model parameters that can be quickly fine-tuned to new tasks with few gradient steps. Foundation for few-shot learning research.

Key Concepts

  • Meta-Learning: Learning to learn from few examples
  • Few-Shot Adaptation: Quickly adapt to new tasks with minimal data
  • Inner Loop: Gradient steps on few examples per task
  • Outer Loop: Meta-update to improve adaptation
  • Model Agnostic: Works with any differentiable model
  • Bi-Level Optimization: Nested gradient computation

Impact

  • 15,000+ citations
  • Foundation of modern few-shot learning
  • Inspired MAML variants (Reptile, ProtoMAML, etc.)
  • Practical algorithm for rapid model adaptation
  • Applications in robotics, computer vision, NLP
  • Theoretical understanding of meta-learning
  • Influenced few-shot learning in modern LLMs

Key Results

  • Few-shot learning with 5-shot adaptation
  • Faster adaptation than standard fine-tuning
  • Works across diverse tasks
  • Better generalization to new tasks
  • Prototypical Networks (Snell et al., 2017)
  • Relation Networks (Sung et al., 2018)
  • Reptile (Nichol et al., 2018)
  • In-Context Learning (Brown et al., 2020)