Use the cross entropy cost function

[xanderdunn]

Outlined here

This intelligently learns quickly when the neural net is very wrong and slowly when the neural net is close to right.