Practical Neural Network Recipies in C++

348

Chapter 19

of failing to detect a member of class i. The total cost of using this

network is

COST -E^c, 19-4

i

Even more generalization is possible if cost can be defined for

particular types of misclassification. Computation of a confusion

matrix will be discussed in a later section. This matrix exhibits every

possible type of misclassification error. The cost associated with the

network would then be computed as the sum of the cost times

probability for each element of the confusion matrix. This is discussed

more fully on page 349.

The astute reader may wonder why, in case of disparate costs

and/or prior probabilities, the network is