Zero-Shot Learning

Definition

Zero-shot learning is a machine learning capability where a model performs a task it was never explicitly trained or shown examples for, relying only on its general pretrained knowledge and the task instructions.

Overview

Zero-shot learning is the extreme end of the generalization spectrum: instead of providing labeled training data or even example demonstrations, the model is simply given a natural-language instruction and expected to produce a correct response. This is possible because large models pretrained on massive, diverse datasets learn broad patterns of language, reasoning, and world knowledge that transfer to tasks they were never specifically optimized for — a capability closely related to transfer learning. Modern large language models exhibit zero-shot capability for a wide range of tasks, from translation and summarization to sentiment classification, simply because these skills were implicitly learned during pretraining on internet-scale text. Performance is typically weaker than few-shot learning, where a few examples help the model calibrate exactly what output format and behavior is expected, but zero-shot prompting requires far less setup. Zero-shot learning is a core reason general-purpose foundation models are so useful in practice: a single deployed model like the ones underlying ChatGPT can handle an enormous variety of unanticipated user requests without task-specific retraining for each one.

Key Concepts

• Performs a task with no labeled examples or task-specific training
• Relies entirely on knowledge learned during broad pretraining
• Enabled by natural-language instructions describing the desired task
• Generally less accurate than few-shot prompting for the same task
• Core to why general-purpose foundation models generalize so broadly
• Widely used for rapid prototyping before investing in fine-tuning

Use Cases

Frequently Asked Questions

Related in AI Fundamentals