Neural Networks: An Overview - Neural Networks: Methodology and Applications - page 33

Fig. 1.16. Each sample is represented as a point in the area-reflectivity plane.
Capacitors are shown as x's and integrated circuits as +'s

explains why feedforward neural networks are e cient classifiers. The purpose
of the present section is to provide a general presentation of classification in its
relation to nonlinear regression. Chapter 6 provides a much more detailed view
of neural network classification and of techniques that evolved from neural
networks.

1.3.1 What Is a Classification Problem?

A classifier is an algorithm that automatically assigns a class (or category) to
a given pattern.
Before considering the specific case of “neural” classifiers, it is important
to understand the basic characteristics of classification problems. Consider
the following illustrative example: in an automatic sorting application, capac-
itors must be discriminated from integrated circuits, from a black-and-white
picture provided by a video camera, so that a robotic arm can grab either a
capacitor or an integrated circuit as requested. Roughly, capacitors appear in
the picture as bright, small rectangular objects, whereas integrated circuits
are large, dark objects. Therefore, the area A and the reflectivity R can be
considered as relevant features for discriminating the objects, i.e., for assigning
a given object either to the class “integrated circuit” or to the class “capaci-
tor”. Assume that samples of capacitors and of integrated circuits have been
collected, and that their areas and reflectivities have been measured: then
each sample can be represented by a point in a two-dimensional space, whose
coordinates are its area and its reflectivity, as shown on Fig. 1.16.

1.3.2 When Is a Statistical Classifier such as a Neural Network
Appropriate?

The above example shows that the ingredients of a classification problem are
• a set of N patterns;