A neural anisotropic view of underspecification in deep learning