Meta-Learning

Learning to Learn

Meta-learning learns to learn.

Approaches

Metric learning: learn similarity. Model-based: learn fast updates. Optimization-based: learn initialization (MAML).

Few-Shot Learning

Learn from few examples. N-way K-shot: N classes, K examples each.

Data augmentation helps. Transductive setting: test examples help.

Applications

Quick adaptation to new tasks. Language models quickly follow instructions.

Key Takeaways

1. Meta-learning learns learning algorithms
2. MAML learns good initialization
3. Enables few-shot learning