maml (model-agnostic meta-learning)

MAML (Model-Agnostic Meta-Learning) finds weight initialization enabling rapid adaptation to new tasks with gradient descent. **Core idea**: Learn θ such that few gradient steps on new task produce good task-specific parameters. Not learning final weights, but learning where to start. **Algorithm**: For each training task: compute adapted params θ' = θ - α∇L_task(θ), evaluate loss on query set with θ', update θ using gradient through adaptation (second-order). **Key insight**: Optimize for post-adaptation performance, not initial performance. Learns initialization sensitive to task-specific gradients. **First vs second order**: Full MAML uses Hessian (expensive), First-Order MAML (FOMAML) approximates (much cheaper, often works well), Reptile (even simpler approximation). **Model-agnostic**: Works with any differentiable model - vision, NLP, RL. **Challenges**: Computational cost (nested loops, second derivatives), requires many tasks for training, sensitive to hyperparameters. **Applications**: Few-shot image classification, robotic skill learning, personalized recommendations, fast NLP adaptation. Foundational meta-learning algorithm still widely used and extended.