Neural networks - Signals, Processes, and Systems - page 446

• The data included in the *.nnd file is divided in the ratio of 2 to 1 into
training data and test data. The latter serve to verify and assess the network
quality. With consecutively recorded series of measurements, there is often
an accumulation of test series with smaller values of learning patterns
followed by an accumulation of test series with higher values or vice versa.
To avoid training with different data from that used in the test, the data is
scrambled in advance and divided into about 10 to 20 classes. These
classes are then divided by class in the ratio of 2 to 1. Thus low as well as
high values can be found with great probability both in the training data
and the test data.

• The number of epochs shows how often the back propagation process with
the learning patterns of the training data is operated. The test data does not
run through the back propagation process, because they have a control
function. In order to create real test conditions, the data that was involved
in the configuration of the neural network should not be used.

• The learning and test errors are calculated every five epochs for the graphic
error display of the learning and test pattern in the menu. Thus with 5000
epochs, for example, no more points than necessary are calculated for
visualization, e.g. for a monitor resolution of 1024 < 768 ( XGA ).

• The learning factor is the weighting factor with which the momentary
“error” of a neuron is incorporated into the next weighting modification.
Logically, the number of learning epochs depends directly on the learning
factor, the learning factor again in a complex way also depends on the
network structure and training data. With a smaller learning factor, with
less modification of the weightings from epoch to epoch, the time before
the final creation of the network is greater due to the higher number of runs.

• Correlation describes the similarity between patterns. The correlation
factor here is a degree for identicalness of target value (nominal value) and
prognosis value. In the menu “train neuronal network”, the correlation
factors between the target values and the prognosis values provided by the
network are calculated and displayed each for the learning pattern, the
testing pattern and the whole of all patterns. These correlations allow
conclusions about the respective error size. The curve shape in the menu
shows the amount of epochs the error size is smallest.

• Standard values : Generally, by standardization (at the range 0…1) of the
entrance data and the parameters at the exit, the possible dominating
influence of an input- or output value shall be “cushioned”. All entrance
values thus have approximately the same chance to influence the network
design in the training. The absolute size of the data at the entrance says
indeed nothing about the relevance for pattern recognition at the exit,
because the parameters at the entrance don´t necessarily stand in any
relation to each other.