Task Distribution (Meta-Learning)
Ordinary machine learning sees one task and many examples. Meta-learning turns that around: many tasks, each with only a handful of examples. The task distribution is the population those tasks are drawn from, and it does for a meta-learner precisely what a training set does for a normal model.
What counts as a task
A task here is a complete miniature learning problem, not a single data point. Each one carries:
- a small labelled support set the model is allowed to adapt on;
- a query set it is then scored on;
- its own label space — "these five handwritten characters", "these five bird species".
Sampling episodes from it
Meta-training draws a fresh task every episode. In MAML and Reptile an inner loop adapts to the sampled task with a few gradient steps, and an outer loop updates the shared initialisation so that adaptation gets cheaper on the next draw. Benchmarks make the distribution concrete: Omniglot samples N-way character tasks, miniImageNet samples N-way image tasks, and Meta-Dataset deliberately spans ten different image sources to widen it.
Why the distribution is the experiment
A meta-learner transfers to a new task only to the degree that task resembles the ones it trained on. Narrow, near-identical tasks produce a system that looks like few-shot learning but has effectively memorised one problem; a distribution stretched too wide leaves the shared initialisation with nothing useful to encode. Task augmentation and multi-source benchmarks exist to attack exactly this, which is why the task distribution is the first thing to read in any few-shot result.
A task distribution is the population of related learning problems a meta-learning system draws from — each task its own miniature dataset, with a support set to adapt on and a query set to be scored on. It fills the role a training set fills in ordinary learning: methods such as MAML sample episodes from it and generalise only to new tasks that resemble it. A few-shot result is uninterpretable without knowing it.
Think of the difference between a student who has crammed one subject and a student who has sat exams in twenty different subjects. The second one has picked up something extra: they have learned how to study.
Meta-learning tries to build that second student, and to do it the computer needs a large pile of different-but-related mini-problems to practise on. That pile is the task distribution. Each mini-problem comes with only a handful of examples, so there is no way to brute-force it — the computer has to acquire general habits instead. The catch is that those habits only carry over to new problems that look something like the practice ones.