Why We Don't Use The Mean Squared Error Loss In Classification