On the Local Minima of the Empirical Risk