Learning can occur without theory. I spent years researching and developing systems to do just that.
If you’re talking about unsupervised classification algorithms, don’t they kinda make their theory as they learn? At least, in the “model,” or “lossy compression” sense of “theory.” Finding features that cluster well in a data set is forming a theory about that data set.
If you’re talking about unsupervised classification algorithms, don’t they kinda make their theory as they learn? At least, in the “model,” or “lossy compression” sense of “theory.” Finding features that cluster well in a data set is forming a theory about that data set.
You still need a theory, a.k.a., a prior on the kind of data you expect to be compressing. Otherwise you run into the No Free Lunch Theorem.