K-shots

K-shots are the labeled examples available for each class in k-shot learning, where k is the exact number of examples a model gets to train on or adapt from for a new task. K-shot learning is a specific form of few-shot learning: one-shot means k=1, two-shot means k=2, and settings commonly go up to around ten examples per class.

The question k-shot learning tests is how well a model generalizes from that small, fixed amount of data. Meta-learning methods such as prototypical networks and model-agnostic meta-learning (MAML) let a model adapt quickly to new classes or domains without full retraining. Training often runs in episodes built from a support set of k examples per class and a query set to evaluate on, with gradient-based optimization tuned to learn efficiently from limited examples.

The quality and representativeness of the k shots strongly affect performance, because those few examples have to capture what defines each class. Selection strategies include random sampling, diverse subset selection and prototype-based approaches that cover as much of the class variation as possible. In practice k-shot methods are used for domain adaptation, personalization and handling new products, scenarios or customer segments, including by AI agents, when collecting a large training set is impractical. They usually trade some accuracy against a model trained on a full dataset.

Related terms

• K-shot learning
• Key differentiator between conversational AI and chatbot

Related services: Agentic AI consulting.

← Back to the AI Glossary