Courses on discrete and convex optimization should be required in every CS program. I don't know why they aren't.

Most of modern ML just extends convex optimization techniques anyway.