Robust AI Models by Learning from the Hard-to-Learn