Neural networks - Signals, Processes, and Systems - page 436

By contrast, the following is true of neural networks:
Neural networks are optimized for specific pattern recognition
by being trained with appropriate data. This is done until the
input training data produce the desired output. In a certain sense
neural networks learn by training!
This ability results from the structure of the neurons and the
network taken from natural neural networks. The possibilities
resulting from this clearly constitute a change of paradigms in
signal processing.
In the introductory literature it is often stated that “neural networks do not require the use
of mathematical rules. They only need examples from which they can learn”. This is,
however, in most cases only true in principle or is only a half-truth:
Generally speaking, signal pre-processing makes sense,
selecting essential information from the training data and
leading to a preferably simple, robust and successful solution for
the neural network.
Optimal signal pre-processing in the field of technology and
natural sciences mostly requires an exact analysis of the physical
context and finally leads to a number of parameters referring to
the temporal process, frequency characteristics or, for example,
also the frequency of occurrence. The information selected by
appropriate parameters is usually fed into the neural network.
Example: A (simple) neural network is to recognize, whether a signal (e.g.
triangle-, sawtooth-, rectangle- or sine signal) is present at the input and report the
result.
Even when using neural networks, at least in the field of technology and natural sciences,
conventional signal processing taking into account the time- and frequency range as well
as statistics is indispensable. The “simple” example above shows this: is a solution
strategy in the time- and frequency range to be preferred? What might an optimal solution
look like if the frequency and amplitude of the signals mentioned vary within certain
limits and/or the input signal is noisy?
In the following the term neural networks always refers to artificial neural networks
unless the term biological neural networks is used.
What are the applications of neural networks?
Neural networks are a powerful new instrument for solving problems in any field in any
possible way connected to pattern recognition .