A learning theoretic perspective on local explainability