Meta-learning, often called “learning to learn,” is the idea of training a model not just to solve one task but to become good at acquiring new tasks quickly. Where ordinary machine learning fits parameters to a single dataset, a meta-learning system is trained across many tasks so that, faced with a new one, it can adapt from only a handful of examples or a few updates.
The most influential modern formulation is Model-Agnostic Meta-Learning (MAML), introduced in 2017 by Chelsea Finn, Pieter Abbeel, and Sergey Levine. MAML looks for an initial set of model parameters such that a small number of gradient steps on a new task’s limited data produces good performance. Because it only assumes a model trained by gradient descent, it applies to classification, regression, and reinforcement learning alike. The approach connects closely to few-shot learning and transfer learning, which also aim to reuse prior experience on new problems.
Meta-learning has been especially important in robotics and few-shot image classification, where collecting large labeled datasets for every new task is impractical. The broader intuition also appears in large language models, where “in-context learning” lets a single pretrained model adapt to new tasks from examples in its prompt, without any weight updates at all.
Why business readers should care: meta-learning is part of why modern AI can be pointed at a new problem with little task-specific data, lowering the cost and time of adapting models to narrow business needs.