Introduction
Show a child one picture of a zebra and they'll recognise zebras from then on. Most ML models need hundreds or thousands of examples per class. How do humans do it? Partly because we've already learned how to learn new animals — we know what kinds of features matter (shape, stripes, size) from years of seeing other animals.
Meta-learning — "learning to learn" — tries to give models this ability. Instead of training on one big task, the model trains on many small tasks, and gets better at adapting quickly to new ones. The goal is few-shot learning: learning a new task from very few examples.
Where it's useful:
- recognising new product types or new faces from a handful of photos;
- personalisation — adapting to a new user from a few interactions;
- drug discovery and rare diseases, where each task has very little data;
- robots that adapt to a new object or environment quickly.