Category : Word Processors
Archive   : HYPELNK.ZIP
Filename : NEURAL

ÚÄÄÄÄÄÄÄÄÄÄÄ¿
³Neural Nets³
ÀÄÄÄÄÄÄÄÄÄÄÄÙ

Similarity maps clearly resemble a Neural Network topology. Related nodes
(items) are connected by pathways (links) and tend to be activated in
neighborhood groups during search or browsing operations.

From our perspective, neural nets seem more inclined to mimic the
amorphous features of a brain than its organizing capabilities, especially
in the "hidden layer." A properly constructed similarity map is more likely
to resemble the old phrenology maps of the mind, or the various cartoons
showing "doctor," "lawyer," "computer nerd" brains.

It may be worth considering the use of a neural-net routine to act as a
monitor (a teacher or a coach) that tracks a path through a similarity map,
possibly using Hopfield* Associations:

"Stop returning to the CONFUSED Map, it's time to tackle NETWORK
Analysis!"

A routine that would keep tabs on node-to-node transitions during specific
applications, could tag "synactic activity" pathways, in accordance with
Hebb's theory.* More frequently used connections can be viewed as
representing more probable or more desirable transitions. An expert user
could serve to teach the system how to react in given circumstances. The
process could obviously be speeded up by incorporating parallel processors.

Possibly, the pattern recognition ability of Neural Networks could be
utilized to help create better InfoMaps, or vice versa. These considerations
go substantially beyond any immediate plans for the process. Any reader more
knowledgeable in the neural net area is encouraged to ! the author.
Any other comments, critique or suggestions would be welcome.

----------------------------------------------------------------------------
* There exist many books, reports and magazines that deal with the topic of
Neural Networks. In the tradition of educator John Dewey and epistemologist
Jean Piaget, we would recommend doing, rather than reading. Try the
BrainMaker program from California Scientific Software, (818) 355-1094.