Beyond generalization: a theory of robustness in machine learning