M L ] 2 7 D ec 2 01 6 Deep Learning without Poor Local Minima