Neural Networks: An Overview - Neural Networks: Methodology and Applications - page 34

• n variables (or features) that describe the patterns and are relevant to

the classification task at hand, the set of descriptors of a given pattern
building the representation of the pattern;

• asetof C classes to which the patterns should be assigned (one of the

classes may be a rejection class in which all patterns that cannot be as-
signed to the other classes will be classified).
Therefore, solving a classification problem requires finding an application of
the set of patterns to be classified into the set of classes.
It is important to realize that statistical classifiers such as neural networks
are not appropriate for solving all classification problems: many alternative
classification methods are available. The following (more or less academic)
examples (from [Stoppiglia 1997]) illustrate the area of application of neural
networks in classification. For each example, the following questions will be
asked:

• Does prior knowledge suggest relevant features?

• Are those features measurable (or can they be computed from measure-

ments)?

• What is the role of the rejection class?

Any vending machine can recognize the coins automatically, and reject fake
or foreign coins. The answers to the above questions are

• Relevant features can easily be found: the coin diameter, its weight, its

thickness, the alloy composition, etc.; there is a small number of such fea-
tures, and new coins are actually designed in order to facilitate automatic
discrimination.

• The features are measurable quantities.

• In feature space, the classes are small hyper-parallelepipeds defined by the

manufacturing tolerances; the rejection class is the rest of feature space.

In such circumstances, a simple decision tree that operates with simple logical
rules, derived from the analysis of the problem, can readily solve the classifi-
cation problem. In such a case, statistical tools such as neural networks are
not appropriate.
Vehicle comfort assessment can be viewed as a classification problem. In
order to anticipate the reactions of customers to a new vehicle, car manu-
facturers resort to panels of customers, who are asked to express an opinion.
Comfort is an ill-defined concept, which depends on many factors such as
noise, seat design, etc. Assessing the comfort, for instance, by classifying it
into three classes (very good, fair, poor), is a process that is di cult to for-
malize because it is based on feelings rather than on measurements.

• The relevant features are not necessarily known and clearly expressed by

the customers; even if features could be defined, the assessments might be
di cult to relate to the features; two customers, under the same conditions,
could give very different assessments.