On the Infeasibility of Training Neural Networks with Small Squared Errors