Nothing comes off the top of my head, and it's not like they get to specialize in compilers. They mostly end up doing one-off hacks to make an existing algorithm more performant.

What's fun about machine learning is that it touches so many other parts of computer science. You could be at a high level writing DSLs in Clojure to make it possible for statisticians to specify their models directly, or you could go to the low level and write GPU code.

The general rule is that if your boss thinks he can do your job, you lose. If he doesn't think that, you win. When you're a data scientist, your odds are much higher of coming out in the second category.