Meta-learning

A person who has learned five card games learns a sixth faster than someone who has never played. They have not seen the new rules before, but they know what to look for. Meta-learning tries to give a model that kind of head start. Ordinary training makes a model good at one task. Meta-learning trains across many tasks and rewards the model for how quickly it adapts to each new one.

Business example. A retailer opens a new product category with only a handful of labeled reviews. A model meta-trained on many earlier categories, each treated as its own small task, can adapt to the new category from those few examples instead of needing thousands.

How it works. The survey by Hospedales et al. describes the setup as two loops. The inner loop learns a single task from a small amount of data. The outer loop adjusts something shared across tasks, such as the starting weights, the learning rule or the way examples are compared, so the inner loop does better next time. The best-known example, MAML, learns starting model parameters from which a few gradient steps on a new task are enough.

Relation to LLMs. Few-shot learning is the goal meta-learning was built for. Large pretrained models now often reach it a different way: a large language model can follow a few examples in its prompt with no weight updates at all. Editorial synthesis: that has made explicit meta-learning less visible in product work, though the idea of training for fast adaptation, rather than for one fixed task, still shapes how researchers think about general models. Related approaches include transfer learning and multitask learning.

Why it matters. Most real teams have plenty of tasks with too little labeled data each. Meta-learning is one research answer to that problem.