Lots of linear functions with \x -> max(0, x) thrown in between.

(That's literally what neural nets with relu as activation unit do.)