To make the data more suitable for training a neural net, fuzzification can be used. Fuzzy class memberships are of proper magnitude and if the classes are suitably overlapping, the resulting membership vectors are also normalized for each original variable.
Figure 5: Fuzzification of discrete data and the resulting new variables.
Figure 5 presents a case, in which the original variable can have four discrete values. The output is three resulting class memberships, of which the middle class represents uncertainty. This kind of fuzzification results in three times n new variables, where n is the number of original variables. These new variables are used in training of the neural network, but the original values can be used while calculating statistics. Also, once the neural network is trained, the fuzzy class memberships can be converted to original values by defuzzification.
In figure 6, and are the class memberships to "Disagree" and "Don't know" classes. The defuzzified value can be calculated as the center of area, in which the value represents the center of gravity of the grayed area. An easier way of calculating the value is to sum all the areas together regardless whether they overlap or not. This is called the center of sums method.