Most machine learning starts every task from zero: random weights, then thousands of gradient steps to fit one specific problem. Meta-learning, often summarized as "learning to learn," asks a different question: can we find a starting point for a model's weights that is specifically good at adapting quickly to brand-new, related tasks, using only a handful of examples?
Meta-training versus ordinary training
During meta-training, a model is exposed to many different but related tasks, not to master any single one, but to discover an initialization from which a small number of gradient steps reliably produces a good task-specific model. Approaches like MAML (Model-Agnostic Meta-Learning) explicitly optimize for this "quick adaptation" property rather than for performance on one fixed dataset.
Why it matters for few-shot problems
In domains where labeled data is scarce, such as personalizing a model to one user's handwriting or diagnosing a rare condition from a handful of scans, there simply is not enough data to train from scratch. A meta-trained starting point can reach usable performance from five or ten labeled examples, where a random initialization would need hundreds.
What "adapting" looks like
Adaptation itself still uses ordinary gradient steps. The difference is entirely in the starting point: a meta-trained initialization sits in a region of weight-space where those few gradient steps move the model a long way toward the right answer, while a random initialization wanders much more before it gets anywhere close.
Try it yourself
The Meta-Learning Lab compares a meta-trained initialization against a random one, both adapting to a brand-new sine-wave task from just five labeled points, so you can watch the gap in real time.
🧪 Try it yourself: the Meta-Learning Lab simulation lets you experiment with everything described above directly in your browser.