Digital Signal Processing Reference
In-Depth Information
Hidden
layer
j
Input
layer
j
Output
layer
j
Input
vector
x
Output
vector
x
Weights
w
ij
Weights
w
jk
Number of
neur ons
N
k
N
i
N
j
Block diagr am of an
(artificial) neuron
Tr ansfer functi on at the
neuron exit
Source : http://images.google .com/
images?svnum=10& hl=en &q=Neuronale +Netze& btnG=Search
Illustration 282:
Visualization of natural "biological" and artificial neuronal networks
In the left half of the picture there is an attempt to visualize a tiny section of a biological neural network in
its spatial complexity. By contrast, as seen in the right hand half of the picture, artificial neural networks
have a rather simple two- dimensional structure that consists of an
imput layer
, one or more “
hidden”
layers
and an
output layer
. Usually, only adjacent layers communicate in only one direction. In
accordance with the biological antetype, each neuron here represents a fairly simple “module” in terms of
signal technology: the weighted incoming signals are added and limited in their size by
non- linear
transfer functions. This limit “stabilizes” the neural network by reducing the influence of dominant sums.
As a result, more training units are necessary until the neural network has finally optimized. With the help
of “massively parallel” signal processing, again as with the biological antetype, it is possible to carry out
extremely rapid pattern recognition even in the case of slow individual processes.. In contrast to a
computer, the
entire
neural network is always involved in pattern recognition. There is no division
between hardware and software in the neural network. The neural network is both things at the same time.
The ability to recognize patterns is hidden in the weighting of the different connections and in the network
structure of the trained neural network!!
NB: this chapter could not have been written before the DASY
Lab
- modules had
been extended by another process modeling group to generate neural networks.
This is the achievement of scientists and engineers at the Fraunhofer Institute für
Produktionstechnik und Automation (IPA) in Stuttgart. At that time, they were
looking for a new approach to avoid defective goods by the monitoring and
optimization of the production process ("zero error production"!).
