Meta-Learner
In ordinary machine learning there is a learner and there is data. Add a second level — many tasks, each with data of its own — and you need something that learns across them. That something is the meta-learner.
What it actually learns
A base learner fits one task. A meta-learner fits the pattern across tasks, and what it produces depends on the design: an initialisation that adapts fast, which is the role the weights play in MAML; an update rule itself, where a recurrent network emits parameter updates in place of hand-written gradient descent; an embedding in which comparing a few labelled examples is enough to classify; or a predictor mapping a dataset's description straight to a promising configuration. The pairing always has the same shape — an inner loop that solves a task, an outer loop that improves how the inner loop goes.
The other meta-learner
The word does double duty, and the two senses are worth keeping apart.
- In meta-learning — the outer-loop learner that generalises across tasks.
- In stacked ensembles — the second-level model trained on base models' out-of-fold predictions to decide how far to trust each one.
Why it matters
A meta-learner is where an AutoML system keeps its experience. Without one, every dataset is the first dataset; with one, past runs become a prior over what to try — which is why frameworks that log their experiments get faster at the same class of problem over time.
A meta-learner is the component that learns across tasks rather than within one. It observes many learning episodes and outputs something that makes the next one easier: a fast-adapting initialisation, a learned update rule, or a mapping from dataset properties to good configurations. It matters because it is where an AutoML system's accumulated experience lives. The term also names the second-level model in a stacked ensemble, which learns how to weight base models' predictions.
A new teacher learns their subject. An experienced head of department learns something different: which teaching methods suit which kind of class. They are not better at the lesson — they are better at choosing the approach.
A meta-learner is that second role, in software. One part of the system learns a particular task, like spotting spam or predicting prices. The meta-learner watches many such tasks and works out what tends to succeed, so the next task can be handled with less data and less trial and error. It is how automated tools get better at getting better.