Machine Learning is fundamentally a computing UI innovation, much like desktop GUI + mice, or spreadsheet software. A new interface between what people want, and the production of computer programs that implement it. A new medium.
This new interface has two facets: data collection & annotation processes on one hand, and loss function engineering on the other hand -- a way to specify the abstractions you want to encode, and what you need to do with them.
What everyone is so focused on -- the big piles of nonlinear transforms trained with gradient descent -- is just an implementation detail in the background.
