Few-shot learning
The training goal of learning a new task from just a handful of labeled examples.
Think of it like
Like recognizing a new bird species after seeing two or three photos, because you already understand what makes birds differ.
Example
A model trained for few-shot learning classifies a new category correctly after seeing only five labeled samples of it.
How it actually works
As a training paradigm, few-shot learning targets sample efficiency: build a model that generalizes to new classes or tasks from very little data, usually via meta-learning or strong pretrained representations. It’s distinct from in-context few-shot prompting, which achieves a similar effect at inference without weight updates. Both chase the same prize — doing more with fewer examples.
For product teams
The ambition of “learn from almost nothing,” valuable wherever labeled data is scarce.
For engineers
Learning to generalize from k examples per class, typically via meta-learning or pretrained features; the training-time analog of in-context few-shot.
Related
- Meta-learning — Often achieved via meta-learning.
- Few-shot — The inference-time analog is few-shot.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome