Few-Shot Learning
Classifying a new category correctly after seeing only a handful of labeled examples of it, instead of the thousands an ordinary classifier needs per class.
An n-shot k-way task gives a model n labeled examples for each of k classes, called the support set, and asks it to classify new query examples from those same classes. Prototypical networks average each class's support examples in embedding space into one prototype vector, then classify a query by nearest prototype.
In-context learning is the same problem solved by large language models at inference time, conditioning on labeled examples placed directly in the prompt instead of any parameter update.