How can I use the cost of the error in output of my classifier?

Question

I mean that during classification I want to use the vector of points, which NN will receive/lost during classification by choosing one of N classes. F.e. {9, -10, -2, 5} instead