Empirical Measure of Multiclass Generalization Performance: the K-Winner Machine Case