FewShot

Few-shot learning is a machine learning paradigm in which a model must learn to perform a task from a small number of labeled examples per class. In typical N-way K-shot settings, the model is asked to classify among N classes after observing K labeled examples of each class, often called the support set, while performance is measured on a separate query set. This setup contrasts with standard supervised learning that requires large labeled datasets and with zero-shot learning that relies on transferable knowledge rather than labeled instances.

Approaches to few-shot learning include meta-learning, wherein the model is trained on many related tasks so it can adapt quickly to new ones; metric-based methods such as prototypical networks or Siamese networks that learn a similarity function between representations; and optimization-based methods like Model-Agnostic Meta-Learning (MAML) that enable rapid adaptation using a few gradient steps. Data augmentation, transfer learning, and memory-augmented models are also used to improve performance. In natural language processing, prompt-based few-shot learning uses large language models conditioned on a small support set to perform tasks without extensive fine-tuning.

Evaluation typically reports accuracy on held-out tasks drawn from the same N-way K-shot protocol, averaged across many tasks. Standard benchmarks in vision include Omniglot, MiniImageNet, and tiered-ImageNet; in NLP, few-shot relation classification and QA datasets are common.

Challenges include sensitivity to task distribution, overfitting to the small support set, reliance on high-pretraining quality, and computational demands of meta-learning approaches. Despite these, few-shot learning enables rapid adaptation to new tasks with limited labeled data.