Few-shot learning is the ability to learn new tasks or recognize new categories from a small number of examples, typically 1 to 10 labeled samples per class. Where traditional deep learning needs thousands of labeled images to learn a new category, a few-shot model adapts from minimal data. This is practical because collecting large labeled datasets is expensive, time-consuming, and sometimes impossible (rare defect types, endangered species, classified objects).

Few-shot learning approaches include metric learning (learn an embedding space where examples of the same class cluster together, then classify by nearest neighbor), meta-learning (train across many tasks so the model learns to learn quickly), and in-context learning (provide examples directly in the prompt, as with VLMs). Modern VLMs support few-shot learning naturally: show the model 3 labeled examples as part of the prompt, then ask it to classify or detect in a new image. Prototypical Networks, Matching Networks, and MAML are classic few-shot architectures, but VLM-based few-shot (via in-context learning) is increasingly the default approach.

Few-shot learning is especially valuable in manufacturing (new product lines with few examples of defects), medical imaging (rare diseases with limited training data), wildlife monitoring (rare species identification), and rapid prototyping of vision systems where teams need to test feasibility before committing to a full annotation campaign.