The machine learning we are familiar with usually trains a model with a large amount of data for a specific task, and often has to start from scratch for another task. Meta Learning aims to solve this problem: to enable models to not only learn to 'do something', but also to 'learn how to learn new things faster'.
Its core idea is to treat the "learning process" itself as the training object. Traditional training optimizes model parameters, while meta learning optimizes at a higher level - it seeks a good initialization, optimizer, or learning rules that enable the model to achieve good results with only a small number of samples and a few updates when facing new tasks. This is also why it is often referred to as' learning to learn '.
The most commonly cited representative method is MAML (Model agnostic Meta Learning). Its approach is to repeatedly simulate "fast adaptation" on a large number of "small tasks", allowing the model to learn an initial parameter that is particularly sensitive to new tasks - starting from this point, with very few samples and gradient updates, it can quickly converge to good results. Due to its independence from specific network structures, MAML is referred to as' model independent '.
Meta learning is closely related to few shot learning: when a category only has a few annotated images, the ability of meta learning to "quickly adapt" comes in handy. It also inspired in context learning of later large models - completing tasks with just a few examples in the prompt without changing parameters, to some extent, bringing the adaptation process into forward computing.
Of course, it also comes at a cost: it requires constructing a large number of tasks, the training process is more complex, and it is also sensitive to task distribution. When the differences between tasks are too large, the effectiveness will be compromised. Therefore, it is more commonly used in scenarios where data is scarce but tasks are diverse, such as small sample recognition, rapid adaptation of robots, and drug and material screening.
In summary, if supervised learning is about "answering a question", meta learning is about "learning how to answer a class of questions". Understanding it helps to see an important path for modern AI to move from being a "specialist" to a "quick adopter".
[Reference source] Comprehensive compilation of publicly available literature on machine learning and meta learning research (such as classic works like MAML).