Newsgroups: comp.ai
Path: cantaloupe.srv.cs.cmu.edu!nntp.club.cc.cmu.edu!hudson.lm.com!news.math.psu.edu!ra.nrl.navy.mil!lamarck.sura.net!newsfeed.internetmci.com!tank.news.pipex.net!pipex!in1.uu.net!pipeline!psinntp!psinntp!psinntp!news.rlcn.rl.af.mil!fester.se.rl.af.mil!user
From: marcoj@ai.rl.af.mil (James D. Marco)
Subject: Re: some basic ad hoc nel stuff
Message-ID: <marcoj-2408951239360001@fester.se.rl.af.mil>
Sender: news@news.rlcn.rl.af.mil
Nntp-Posting-Host: fester.se.rl.af.mil
Organization: Rome Laboratory USAF
References: <Pine.SUN.3.91.950823181618.29646B-100000@sun2>
Date: Thu, 24 Aug 1995 17:39:36 GMT
Lines: 62

In article <Pine.SUN.3.91.950823181618.29646B-100000@sun2>, Eugene Leitl
<ui22204@sunmail.lrz-muenchen.de> wrote:

[text snipped...]
> Comments?
> 
> -- Eugene

Much of interest here. Couple of comments:
An article in SCIENTIFIC AMERICAN (SA August, pg.16,18,"It's All in the 
Timing", John Horgan) some interesting results were published concerning
the timing of neural transmissions.  It was indicated that signal
transmission 
times were fairly constant in response to the same stimulus(<1ms varience).
The indication is:  "temporal neural computation" is an available biological 
computational pathway.
     

These experimental results are not completely reflected in the paradigm 
you propose (see quote below).

   "Whether information encoding is the mean firing rate over a time
   interval or the delay between two subsequent spikes (limiting
   precision?) shall at this moment be of no interest to us. Fact
   is, the neuron has a state at t and it makes its private
   synapses quite aware of it. Since the frequency/delay dynamic
   range is quite limited and the neuron is a (relatively)
   low-quality analogue circuit, we can approximate its state at t
   by a few-bit cardinal integer (8-16 bit), 16 bit being surely a
   distinct_ overkill.

   I very much doubt that axons/neurits play a role in information
   processing (as I totally ignore glia here), though they
   introduce a delay (typically very small in relation to the
   synapse "switch" time scale (at least considering the neocortex
   pyramidal cells superclass), them being constrained by the
   postsynaptical cleft diffusion speed) and may frequency-filter
   the signals. In any case this is nothing that couldn't be
   bundled into the synapse instance, requiring only some few-bit
   integers."
 
Since "axons/neurits" cause temporal changes to the processing in a net,
based on the recent research quoted above, these effects cannot be
discounted. 
(the "beat of a butterflies wings", and all that).

The biological neuron contains several methodes of adjusting the various
timing parameters for stimulus, transmission and response (recovery time -
after firing, signal propagation along the neuran's cell membrane,
neurotransmitters of various types, etc). 

I recently modeled the delays in synaptic transmission as integer values
representing the recovery delay of a neuron. A set of 16 neurons was
trained by
using a GA to adjust the timing parameters to produce a particular string
in response to a particular string input.  Initial results indicated a
convergence towards the target after 100 iterations.  I hypothesize that
it may be possible to perform computations based entirely on delay
control.       
   

No flames intended here, simply some new info...
