Posit: the problem of cognition has almost no overlap with supervised learning, what we're currently good at -- learning to map space x to space y given a dense sampling of x-cross-y
-
Definitely dense, since it runs on a convnet. By their very nature, convnets require to be exposed to a dense sampling of their data manifold. But that doesn't mean they need to be exposed to every possible input (e.g. every possible image, for an image classification model).
They are capable of generalization, but it is necessarily local generalization, i.e. pattern recognition.
