Digital Signal Processing Reference
In-Depth Information
Exercises on chapter 14
Exercise 1
The use of neuronal networks in computer- assisted digital signal processing is often
described as “change of paradigms”. What reasons can be given for this in comparison
with “conventional” analogue and digital signal processing?
Exercise 2
(a) Why does “conventional”
signal preprocessing
appear suitable with the use of
neuronal networks? What is achieved with it in the end?
(b) What knowledge is required, what support is reasonable, to perform best possible
signal preprocessing?
Exercise 3
The IPA modules each provide a parameter module for the
time range
and the
frequency
range
(frequency analysis).
(a) For the
time range
, the parameter module includes numerous special mathematical
functions. Here, the
mathematical
understanding of the user is especially asked for
when it comes to the selection of functions that could be used in a practical case.
During which practical problem respectively example would you use e.g. the
function “standard deviation" or “median”?
Does anything speak against selecting as many functions as possible to guarantee
those pattern- information filtered out you might not see in the first place??
(b) The use of the parameter module for the
frequency range
requires a special
physical
understanding in practical appliance. Please name some case studies in this context
and try to find a generalizing statement to its use.
Exercise 4
For the realization of neuronal networks, exactly two modules are at hand:
collect network
data
(respectively collect training data) and
network prognosis
.
(a) With the help of the module
collect network data
, at first the *.nnd file is created.
Hereby, the entrance- (parameter-) patterns are allocated to certain target patterns.
How can you combine these entrance patterns with the target patterns in practical
use?
(b) The neuronal network is included in the *.nn file which we got over the menu of the
*.nnd file (“learning”). Explain the strategy how the network structure is found
through the entrance data.
(c) The menu “network training” includes a three- layer
default
for the neuronal network.
What possibilities are left (e.g. with the “XOR- network”) if this doesn´t provide a
functioning network?
