One of the best remarks from my prof on ML: you can prove things, but the real question is how it performs once the assumptions don't hold.