Are Saddles Good Enough for Deep Learning?