Neural networks - Signals, Processes, and Systems - page 461

Neuronal networks: perspective and limitations
Neuronal networks have the ability to learn in a certain sense. This fact alone represents
a change of paradigms in technical signal processing. This results in new applications, as
even by relatively small, unidirectional working networks with back propagation become
possible.

The question arises, if significantly bigger neuronal networks of this type basically can
generate new applications, making them roughly comparable to biological neuronal
networks. This is very likely to be negated, as too many structural characteristics of
biological neuronal networks are missing, especially regarding the brain as being the
antetype.

In the nearer future, it will be crucial to understand these missing structural characteristics
and to implement them bit by bit into neuronal networks. In this case, glorious
perspectives can be prognosticated for this technology. It shall be allowed to draft a
possible direction of development.

Feedback arguably plays a decisive role in biological neuronal networks, especially in the
so- called neo cortex (phylogenetically the youngest, most differentiated part of the
human cerebral cortex, i.e. 90 % of it), the domicile of intelligence. Here, most “circuits”
are downright dominated by feedback mechanisms. There´s no doubt exactly this
feedback is the recipe for success of nature regarding the brain. Nevertheless, there is no
approved theory for the why out there so far.

Hint:

It shall be mentioned in this context, that backpropagation represents no real
feedback. It only happens in the learning phase , not in the working phase. Any
information here only flows one- directional from entrance to exit..

Furthermore, our neuronal network doesn´t own a sense for time . Biological neuronal
networks process fast changing information streams. The neuronal network applied here
in contrast processes a static input pattern and transforms this into a static output pattern.
Then, another input pattern is presented.

In the unidirectional working neuronal network with
backpropagation used here, there is no history or chronicle of
what happened a short time before. It is impossible to compare a
new pattern or partial pattern with a pattern that had been stored
before, a profound and system defining characteristic of the
biological neuronal network. This can pricipally be organized by
the means of feedback.

Considering this, an interesting alternative to the neuronal network is the auto associative
store . The fact that the output of the neuron is fed back to the input is the main difference
to the neuronal network applied here. This may appear strange, but the feedback loop
leads to a couple of interesting characteristics. If an activity pattern is imposed on these
neurons, they store this by feedback. In a certain way, the auto associative network
associates with itself, therefore the name.