Meta-Learning

Training a model on many tasks so it gets good at quickly learning new tasks, not just one fixed task.

Meta-learning, also called “learning to learn,” does not train toward a single task; instead the model repeatedly goes through “see a few examples, adapt, get tested” across many related small tasks, and what it learns is an initialization, update rule, or memory mechanism that lets it adapt quickly to a new task. Common approaches fall into three families: metric-based (such as prototypical networks), model-based (using external memory or fast weights), and optimization-based. The best known is MAML, proposed by Chelsea Finn, Pieter Abbeel, and Sergey Levine in 2017, which directly trains a set of initial parameters so the model performs well on a new task after just a few steps of gradient descent on a small amount of data. Because robot data is expensive, meta-learning was long the main approach to learning skills from very few examples; one-shot imitation learning and meta-reinforcement learning both build on it, and some researchers view large models' in-context learning as a form of implicit meta-learning.

ExampleFinn and colleagues (2017) used MAML for “meta-imitation learning”: the robot is meta-trained on demonstrations from many tasks, and at test time it only needs to watch a single visual demonstration of a new task to learn it end to end.

Also called

Learning to Learn

Related

Few-shot · One-shot Imitation Learning · Meta Reinforcement Learning · In-Context Learning · Transfer Learning · Multi-Task Learning

Sources

Finn et al. 2017: Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks (MAML)
Lilian Weng: Meta-Learning: Learning to Learn Fast (2018)
Finn et al. 2017: One-Shot Visual Imitation Learning via Meta-Learning