meta-learning (learning to learn)

Meta-learning trains models to quickly adapt to new tasks with minimal examples - "learning to learn." **Goal**: Learn general adaptation strategy across many tasks, apply to new tasks with few examples. **Problem setup**: Training involves many tasks (each with support/query sets), model learns what transfers across tasks, evaluated on ability to adapt to held-out tasks. **Key approaches**: **Metric-based**: Learn embedding space where similar examples cluster (Prototypical Networks, Matching Networks). **Optimization-based**: Learn initialization for fast adaptation (MAML). **Model-based**: Learn model that directly produces new model weights or predictions. **Training**: Sample task → fine-tune on support set → evaluate on query set → update meta-parameters based on performance. **Few-shot classification setup**: N-way K-shot - classify among N classes with K examples each. **Applications**: Robotics (new skills quickly), drug discovery, personalization, low-resource languages. **Challenges**: Task distribution matters, computational cost, transferring to very different tasks. Foundation for few-shot learning research.