Deep Learning without Poor Local Minima