> My understanding is that there are not very strong theoretical underpinnings to most of the machine learning techniques in use today.

Yes, particularly with regards to generalization. There's a large number of theorists racing to explain that.