 |
|
| Computers Forum Index » Computer Artificial Intelligence - Neural Nets » Into the principle of neural network operation... |
|
Page 1 of 1 |
|
| Author |
Message |
| Oleg Volkov... |
 Posted: Mon Nov 02, 2009 10:28 am |
|
|
|
Guest
|
Hi, all!
I want to share with you ideas I have got of how a neural network
within a biological system can operate. I do not know whether the
ideas are new, however.
I am not well informed about current trends in development of
artificial neural networks. From the information that catches my eye
when I'm surfing the Internet I came to conclusion that modern
algorithms of training of neural networks often imply the process of
training can not be performed by neural network itself and requires
computational facilities outside the network. Some training algorithms
even require the use of function gradient, so the question arises
whether these complexities can really be encountered in biological
systems. The complexities may exist, but they should rest upon simple
and sound principles, which are mostly used by nature.
I came across with the use of function gradient when reading the
abstract of an article in which Hidden Markov Models (HMMs) were
suggested to be among the concepts biological neural networks are
based on. The idea of state model of some sort represented by
biological neural network is not new and is proved by experiments in
which individual neurons in human or animal brain are activated when
specific pictures are exposed to living beings or when those beings
find themselves in specific situations.
So one can consider the objective of operation of biological neural
network is the building of environment state model and acting
according to that model. The whole meaning of life is inconceivable,
but at the beginning, we can think the task living being should
certainly solve is to obtain food from the environment, or some kind
of resource, or energy.
To model the environment as a state system a deterministic finite
automaton can be taken as initial approximation. The neural network
has to communicate with the automaton by means of automaton input and
output signals. To some transitions between automaton states numbers
can be assigned which indicate the amount of food, resource, or energy
living being receives when corresponding transition is made.
Therefore, the neural network should explore the model of automaton
states and make the automaton to execute transitions with greater
numbers more often.
I do not know whether the algorithm occurred to me is new. Regardless
of that fact, the software project I am giving a link to is of
indispensable benefit it contains the algorithm implemented
programmatically. The basic idea of the algorithm is quite simple and
therefore can be discovered in biological neural networks some day.
Roughly speaking, the main part of the algorithm is dedicated to
solving two tasks simultaneously: state model training and a kind of
task usually performed by Viterbi algorithm.
But... The algorithm description is still draft and possibly contains
wrong, ungrounded, or unclear statements. Nevertheless, I hope reading
the description will be helpful to you.
QSMM Project pages:
http://qsmm.sourceforge.net/
http://sourceforge.net/projects/qsmm/
Oleg Volkov |
|
|
| Back to top |
|
|
|
|
|
All times are GMT
The time now is Sun Dec 06, 2009 3:50 am
|
|