Digital Signal Processing Reference
In-Depth Information
random bin
time domain
network data
2
XOR
1
3
7
4
7
5
6
Training
Abbildung 287:
The „collect network data“ module
This module receives the training data, evaluates it and uses it to train the neural network. This creates the
decisive module. Again, the straightforward example of an XOR- network was used to make the internal
data of this module as plain and clear as possible.
The entire process can be described as follows: (1) A binary pattern generator (“black box module”)
provides a two- channel stochastic binary signal with the values 0 and 5 (TTL level). This binary pattern
reaches the “collect network data” module (2) as well as an XOR- gate (3). The latter provides the correct
XOR output signal. This signal is sent to the input N of the “collect network data” module. This input
receives the signals that will later be displayed at the
output
of the neural network with the now due binary
input combination, in this case 0 or 5.
he input and output data of the subsequent neural network are initially filed in an nnd file, clearly
structured (4), here illustrated here as an EXCEL file. The entire neural network structure has now to be
calculated from this (“learn” button (6)). A subprogram changes the weightings and minimizes errors
using bac kpropagation until these errors are below a predefined limit. The results of this calculation are
filed as a network file (nn file) (5). This network file (7) gives a complete description of the subsequent
neural network (see also Illustration 287).
The “collect network data” module
As Illustration 286 shows, this module not only collects data but is in a certain way the
actual basis for the creation of the neural network. The number of neurons of the input
layer of the subsequent neural network simply results from the number of parameters
appearing at the output of the parameter module or/and the frequency parameter module.
