Deep learning without poor local minima