Comment by gnipgnip

The only "math" in deep learning is given by reverse mode AD (or if you're into fancy stuff, "efficient computing of pullbacks"). The rest of it plain old hacking, and empirical tricks with the occasional variational doodads.