Meta-learning
Training a model to get good at learning new tasks quickly, not just at one task.
Think of it like
Like a student who has learned how to study, so each new subject clicks faster than the last.
Example
A meta-learned model adapts to a brand-new classification task from just a handful of examples, having practiced adapting many times.
How it actually works
Meta-learning — “learning to learn” — trains across many tasks so the model acquires a strategy for fast adaptation, rather than mastering a single task. Approaches range from optimization-based methods that find easily-fine-tuned initializations to models that adapt in-context. It’s the formal cousin of what large models do when they pick up a task from a few prompt examples.
For product teams
The idea behind models that adapt fast to new tasks with little data — flexibility as a trained skill.
For engineers
Optimize across a task distribution for rapid adaptation; kin to in-context few-shot behavior in large models.
Related
- Few-shot learning — Related to few-shot learning.
- In-context learning — Also related to in-context learning.
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