Zero-Shot Learning
A model capability to correctly handle tasks or classify data it was never explicitly trained on by leveraging generalized knowledge representations.
Definition
Techniques (e.g., prompt engineering in large language models, attribute-based classifiers) that enable models to infer relationships between known and novel classes based on shared semantic or feature embeddings. Governance considerations include validating zero-shot performance on representative hold-out categories, monitoring for unexpected misclassifications, and establishing fallback procedures when confidence is low on unseen inputs.
Real-World Example
A customer-support chatbot built on a large language model uses zero-shot classification to route “subscription upgrade” questions—even though it was never trained on that label—by matching semantic intent to available categories. The system logs low-confidence zero-shot routings for human review, ensuring correct handling of novel inquiries.
.svg)
