In article <3hu50m$ep4 at usenet.rpi.edu>, lohnen at marcus.its.rpi.edu (Nils
Lohner) wrote:
> I am trying to program a neural network Character Recognition scheme
> modeled as closely to the human brain as possible, albeit on a much smaller
> scale. I need information on how neurons and synapses are connected. Is
> there some sort of a ratio between the number of inputs to outputs of a
neuron?
> How are the initial synapses designed/decided upon at birth (this is for
network
> initialization)? I am basically looking for any connectivity information
> at various states of brain development.
> Any help or references regarding this topic will be greatly appreciated.
> Please email or post responses. Many thanks in advance.
>> Nils Lohner
> --
> - Nils Lohner internet: lohnen at rpi.edu Rensselaer Polytechnic Institute
Which part of the brain do you want to model? There is no standard ratio,
so it is difficult to answer your question. The number of inputs and
outputs a neuron has can vary from less than 10 to several hundred,
depending on the type of neuron. Here are some rough generalizations which
you might use as a starting point. Other readers may be able to give you
more accurate information.
1. many neurons synapse on each target neuron (hundreds?)
2. each neuron synapses on many neurons (perhaps hundreds) and may make
several synapses on each neuron it innervates (tens to hundreds).
3. in neonatal animals neurons tend to innervate more neurons than in
adults, so there is a reduction in the number of "inputs" and "outputs"
for each neuron during the later stages of development (of course
initially these are both zero, so we are talking about an initial increase
to a maximum, then a reduction to the final adult number).
To complicate things, while neurons are reducing the number of target
cells that they contact ("outputs") they tend to increase the number of
synapses they make on each of the neurons that they do remain in contact
with.
But in my humble opinion, any attempt to make your model "realistic"
will not succeed. The main problem is that this kind of data is available
for only a few kinds of neurons, and these tend to be relatively simple
cells. We just don't have precise data (or even rough data) for most
neurons. Without a real understanding of the circuits involved, how can
you hope to make a "realistic" model? But these general principles might
help you get on the right track.
--
Donald Wigston
Atlanta, GA
