Few-Shot and Zero-Shot Learning

Handling new tasks or classes from a handful of examples (few-shot) or from a description alone (zero-shot), by reusing knowledge from pretraining.

Also zero-shot · few-shot · one-shot learning · meta-learning

Fig. / How it works

Domain 02: Learning Paradigms
Role · kind: Core ideasParadigm
Level: Intermediate
Map: Show on the map

You have never seen a zebra, but someone tells you 'it's a horse with black and white stripes'. You recognize the first one you meet. That's zero-shot. Seeing two zebra photos first would be few-shot.

Where it fits

+All relationships as a list

Is a kind of / part of

Learning Paradigms

Kinds and parts

In-Context Learning

Siblings

Uses / built on

Transfer Learning

See it on the full map

How it works

Classic supervised learning needs many labeled examples of every class. People do not: you can recognize a new animal from one picture, or from a description.

Large pretrained models bring machines closer to that:

Zero-shot: no examples at all. CLIP puts images and text in one embedding space, so you can classify an image by comparing it to sentences like “a photo of a capybara”. A language model can follow the instruction “classify this review as positive or negative” without any examples.
Few-shot: a handful of examples. You can fine-tune a pretrained model on them, or simply put them in the prompt, which is in-context learning.

Both depend on transfer learning: the model has already learned general features during pretraining, so the new task needs only a little extra information.

Everyday use cases

Case 01
Photo search for anything
Typing "a red bicycle at night" finds matching photos even though no one trained a "red bicycle at night" classifier.
Case 02
Classifying support tickets
An LLM sorts tickets into new categories just from their names and two examples each.

Strengths, weaknesses and when to use it

Strengths

Removes or shrinks the need for labeled data for every new task.

Weaknesses

Usually less accurate than a model trained on many labeled examples for the same task.
Sensitive to how classes or prompts are worded.

When to use itWhen you need a quick working system for a new task or label set with little or no labeled data.

Compare with

Check yourself

Q01
CLIP classifies an image of a capybara correctly without capybara training labels. What is this?
Show answer
A: Zero-shot learning