[Talk] Beyond Empirical Risk Minimization: the lessons of deep learning