Neural networks - Signals, Processes, and Systems - page 438

As the above examples and the closer definition/generalization of the capabilities of
neural networks show, the term pattern recognition is fairly abstract and can describe a
huge variety of phenomena. After all, it is the key term for any form of communication!!
In order to be able to assess the realization or the practical application of neural networks,
the following is an attempt to define more abstract subdivisions of this term:
Generalization
Generalization is to be understood as the ability to reason when confronted by
unknown things. This on condition that the unknown information may be different
but at least similar to the known information. This sounds complicated but will
become clear with examples:
• speech recognition
• valuation of real estate (location, transport links, insulation etc.)
Trend forecast
Because of the time axis inside the data material the future performance can be
interpolated. Examples:
• timely replacement of worn parts
• forecast of natural disasters
Evaluation
There is an expected behavior . The evaluation may be calculated by the variance
between actual and targeted data. Examples:
• quality differences in production
• approval/refusal of a loan
Tolerance
There are almost always tolerances in the case of comparable real data. Here, the
term tolerance refers to unclean, incorrect and incomplete data. Examples:
• complete or replace missing symbols, texts or image fragments
• decision to intervene in the system or stop the production process
Filtering
Select relevant information (to avoid redundancy) to minimize the effort of pattern
recognition. Examples:
• clean signals or data of “noise” in the broadest sense
• segmentation of important image contents in modern imaging processes
(e.g. ultra sound and NMR tomography)