Meta-learning

Meta-learning, or “learning to learn,” seeks to have a system, after training on many different tasks, acquire the ability to adapt quickly to new tasks using very few examples. Instead of learning a single task from scratch, it learns a method of learning. Its goal, as the authors of a landmark work put it, is “to train a model on a variety of tasks such that it can solve new tasks using only a small number of examples.”

Learning with few examples

Meta-learning is the usual approach for few-shot learning, where the model must recognize new concepts from one or a few examples per class. Training is organized into episodes: in each one, the model receives a small support set with labeled examples and must classify a query set. It thus rehearses, over and over, the situation of learning fast.

Three families of methods

The techniques are usually grouped into three. The optimization-based ones, such as MAML (Finn, Abbeel and Levine, 2017), learn an initialization of the parameters such that a few gradient steps suffice to adapt to a new task. The metric-based ones, such as prototypical networks (Snell and colleagues, 2017) or matching networks (Vinyals and colleagues, 2016), learn a space where classifying means measuring distances to a representative of each class. And the model- or memory-based ones use architectures with external memory to store and reuse what has been seen.

It is not the same as transfer learning

It is worth distinguishing it from neighboring concepts. In transfer learning, the knowledge of a model already trained on one task is reused for a related one, typically with abundant data. Meta-learning, by contrast, teaches how to learn from many tasks, and shines when the new task's data are scarce. Nor is it the same as hyperparameter optimization, although it can be applied to it.

Its echo in large models

There is a bridge to more recent AI. The in-context learning ability of large language models—solving a task from a few examples included in the prompt, without retraining—is often interpreted as a form of meta-learning: the model would have learned, during its training, to learn from the examples it is shown. It is a fertile conceptual reading, more than a formal equivalence.