Neural networks and principal component analysis: Learning from examples without local minima