A learning setting in which a model is required to generalise to a new task or class from only a small number of labelled examples, typically by leveraging prior knowledge or inductive biases learned across many related tasks.
Semantic Classification
Content
- Few-shot learning addresses tasks where labelled data is scarce, contrasting with conventional supervised learning that assumes many examples per class. Approaches include metric learning, optimisation-based meta-learning and the use of pre-trained representations that transfer to new tasks.
- Large language models exhibit a related capability through in-context learning, where examples supplied in the prompt guide behaviour without parameter updates. Evaluation typically uses episodes that specify a small support set and a query set drawn from previously unseen classes.