It appears that it may be necessary to acquire a very large number of tasks for meta-learning , because MAML for example says that each task is analogous to a single training example in regular learning.
This is slightly confusing to me because it appears that outside of automated techniques like N-way classification where you randomly sub-select classes (& training examples) to include in a given task.
Adding tasks seems to be quite laborious right? I mean it sounds like you would need to get a different micro dataset for each task? And if meta-learning needs so many tasks then how do you satisfy that need?
