What is one-shot learning?

Definition

One-shot learning tests whether a model can use a single example to infer the important structure of an unfamiliar task. The example may be provided during training, adaptation, or directly in the inference context.

Success requires more than copying the demonstration. A useful one-shot learner identifies the task intent and transfers it to changed objects, states, or conditions while avoiding irrelevant details from the example.

Frequently asked questions

How does one-shot learning differ from zero-shot learning?

One-shot learning receives one example, while zero-shot learning attempts the task from instructions or prior knowledge without an example.

Is one-shot learning the same as copying?

No. Genuine task learning should transfer the demonstrated intent to a new situation rather than reproduce the exact observed motion.

Videos explaining one-shot learning