Pattern Recognition Techniques - Biomedical Signal Analysis: Contemporary Methods and Applications - page 212

With a linear separation, the perceptron learned a diagnosis differen-
tiation in 90% of the analyzed samples. The network showed a graduated
discrimination specificity for the diagnoses CB and ILD. The application
of the ANN to a larger number of samples and higher-dimensional data
sets, could prove the benefit of this artificial intelligence tool.
In conclusion, the combination of these artificial intelligence ap-
proaches could be a very helpful tool to facilitate diagnosis assignment
from immunological patient data where no diagnosis can be given or the
discrimination between diagnoses is dicult.

6.11 Overview of Statistical, Syntactic, and

Neural Pattern Recognition

The artificial neural networks techniques are an important part of the
field of pattern recognition. In general, there are many classification
paradigms which lead to a reasonable solution of a classification problem:
syntactic, statistical, or neural. The delimitations between statistical,
syntactic and neural pattern recognition approaches are fuzzy since
all share common features and are geared toward obtaining a correct
classification result.
The decision to choose a particular approach over another is based on
analysis of underlying statistical components, or grammatical structure,
or on the suitability of a neural network solution [173].
Table 6.5 and figure 6.22 elucidate the similarities and differences
between the three pattern recognition approaches [227].
Both neural and statistical classification techniques require that the
information be given as a numerical-valued feature vector. In some cases,
information is available as a structural relation between the components
of a vector. The important aspect of structural information forms the
basis of the structural and syntactic classification concepts. Thus, struc-
tural pattern recognition can be employed for both classification and
description.
Each method has its strengths, but at the same time there are also
some drawbacks: the statistical method does not operate with syntactic
information; the syntactic method does not operate based on adaptive
learning rules; and the neural network approach does not contain any
semantic information in its architecture [173].