Neural networks - Signals, Processes, and Systems - page 441

On closer examination, back propagation is a so-called gradient descent method set in a
landscape, whose high and low points are formed by one or more errors.
Back propagation then uses the steepest way to the adjacent valley (the gradient exactly
defines the path water or a rolling ball would take on its way to the valley).
With each backpropagation run, the landscape changes by the alteration of the weightings
and errors. The low point comes closer, or rather the immediate vicinity is shown
enlarged. The stable condition, i.e. the condition of the lowest potential, is reached at the
valley bottom.
Creating neuronal networks with DASYLab
DASY Lab is a progam for data acquisition . DASY Lab is fed with data or signals via
interfaces or files that are processed, analyzed, evaluated, visualized, emitted etc. by the
virtual system that DASY Lab represents.
In most cases, these are time-dependent data/signals, which within a defined range show
a definite curve shape. This time-dependent curve shape contains the information,
properties or patterns that correspond to the target values at the output of the neural
network.
The parameters extracted from these signals can be allocated to the time range as well as
to the frequency range . Therefore, two parameter modules are provided for signal
preprocessing. One module evaluates the time course, the other evaluates the spectral
characteristics.
NB: Neuronal networks are also used outside the technological and scientific field
(see examples). The data material does not necessarily have a time axis and
therefore no spectral characteristics in a physical sense!
The module “Calculating parameters”
The goal of signal pre-processing is data reduction . The aim is to extract significant
parameters from the data or the signal which are typical of the distinguishing
characteristics of the various target values of the neural network. The parameter module
provides an array of mathematical functions for this purpose.
Illustration 284 describes the process of parameter selection. The generator here provides
a (periodic) sawtooth consisting of 256 discrete values. In the menu, various parameters
are selected, in this case 7 in all. At the output of the parameter module, the discrete
parameters appear in the sequence defined by the selection (see parameter window). The
reduction of the data by the factor 36 (256/7) can be clearly seen when the parameter
signal is compared with the sawtooth signal.