Humans learn new categories from a handful of examples. Most machine learning models need thousands. Few-shot learning narrows that gap by training models to recognize new classes from only a few labeled samples.

Approaches vary. Metric learning trains networks to compare examples, so new classes can be matched by similarity. Meta-learning trains models to learn quickly across many tasks. Large language models achieve few-shot behaviour through prompting alone, without weight updates.

Common settings

• One-shot, a single example per class
• Five-shot, five examples per class
• N-way K-shot, N classes with K examples each
• Zero-shot, no examples, only descriptions

Few-shot performance still lags full supervision on most benchmarks. The gap is closing, especially for language tasks, but the technique works best when combined with strong pretraining and careful prompt design.