metatraining

Metatraining is the training process used in meta-learning where a model is trained across a distribution of tasks to acquire a prior that enables rapid adaptation to new tasks. The goal is for the model to learn not just a solution to individual tasks, but a learning strategy that can be applied quickly to unseen tasks with limited data.

In a typical metatraining setup, tasks are sampled from a task distribution. Each task provides its own data split into a support set used for inner-loop adaptation and a query set used to evaluate performance. During training, the model undergoes inner-loop updates on each task to adapt to that task, and an outer-loop optimization updates meta-parameters to improve performance across tasks. This episodic training encourages the model to generalize to new tasks with minimal data.

Common metatraining approaches include gradient-based methods, such as Model-Agnostic Meta-Learning (MAML) and its variants (e.g., ANIL, Reptile), which perform fast adaptation via gradient steps. Other families are metric-based methods like ProtoNet and memory-augmented or probabilistic meta-learning models. The choice of method depends on the problem domain, data availability, and desired adaptation speed.

Metatraining is central to few-shot learning and rapid task adaptation, with evaluations typically on benchmarks like few-shot image classification or reinforcement learning settings. Challenges include distribution shift between training and test tasks, computational cost of inner-loop optimization, and ensuring robust generalization to truly new tasks.

In summary, metatraining trains a model to learn how to learn, enabling efficient adaptation to new tasks by leveraging experience across a diverse task distribution.