What is zero-shot learning?
The Harvard Chan AI Glossary
In our classes, we often pause to explain the ideas behind AI and the language used to describe it. These posts grew out of those conversations. They’re for anyone who would like to understand AI better, with no technical background assumed.
Zero-shot learning means applying an AI system to a task or category without giving it labeled training examples for that particular task or category. A labeled example pairs some information with the answer to be learned, such as a photograph identified as a dog. “Zero-shot” refers to the absence of those task-specific examples, not to an absence of prior learning.
Imagine a system that has learned relationships between pictures and written descriptions. You ask it to choose whether a new picture shows a bicycle or a motorcycle. It may compare the picture with descriptions of those categories, rather than needing a new collection of pictures labeled specifically for that comparison.
In traditional research, zero-shot learning often concerns categories excluded from the relevant training examples. The system uses other information, such as a description of an animal’s appearance, to recognize an unfamiliar category. In broad, internet-trained systems, the claim can be narrower: no additional labeled examples were supplied for the test. It may be impossible to establish that the system never encountered similar material earlier.
Zero-shot prompting is a related but different expression. It means asking an AI system to do something without including demonstrations in the prompt, which is the instruction and information supplied to it. “Translate this sentence into French” is a zero-shot prompt. Providing several sentence pairs first would make it few-shot prompting, meaning prompting with a few examples.
The distinction matters because a system can receive no examples in your message while still having extensive relevant experience from its earlier training. Training is the process through which software adjusts its internal settings using data. Zero-shot use draws on that earlier learning; it does not start from nothing.
A systematic-review team might ask a language model to screen study abstracts against written inclusion and exclusion criteria without showing any worked examples. That is zero-shot prompting, not proof that the model has never encountered those trials during training. The team would still need to measure missed eligible studies, false inclusions and ambiguous abstracts against a dual-reviewer standard. “Zero-shot” describes how the prompt was written; it does not certify that the screening is reliable.
Sources: Radford and colleagues — Learning Transferable Visual Models from Natural Language Supervision; Brown and colleagues — Language Models Are Few-Shot Learners.
If you would like to find out more about AI and improve your leadership skills, consider the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health. You can take an individual course or work toward the full certificate.
AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.
For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.
Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy