🚀 go-pugleaf

In article <342fb9ae.4111800@news.earthlink.net>,
Louis Savain <savainl@earthlink.net> wrote:
>In article <60f4dc$d4j@andromeda.rutgers.edu>
>           ohaneef@andromeda.rutgers.edu "Omar Haneef" writes:
>  S[utton] & B[arto] have indeed done excellent work.  More recently however, the
>more interesting news are coming from neurophysiology and those in the
>field of computational neuroscience (has little to do with traditional
>ANNs) who are attempting to model the newest findings.
>
>  There is a recent discovery (little known in AI
>circles) by Henry Markram that strongly suggests that a "differential
>Hebbian" (A term I borrowed from Dr. Leslie Smith of the University of
>Stirling who may have borrowed it from A. H. Klopf) form of learning
>is taking place at the single neuron level.  In our humble opinion,
>this is a truly revolutionary finding.  What Markram and his team at
>the Weizmann institute for Science have found (see January 10, 1997
>issue of Science) is that the precise order in which a spike arrives
>at a synapse and a retrograde or antidromic spike is generated by a
>target neuron, has a dramatic effect on the strengthening or weakening
>of the synapse.  If the synaptic potential arrives first (within 10
>millisecond), the synapse is strengthened otherwise it is weakened.
>Our own model, which is based on causality, predicted as much.  What
>neurophysiologists (and everyone else for that matter) are in the dark
>about is the precise input conditions that lead to the firing of a
>neuron.  There is nothing wishy-washy about them.  The temporal or
>causal nature of these conditions is extremely important to learning
>and is directly related to Markram's discovery.  We think we know what
>these conditions are, as you will all soon be able to judge for
>yourselves.
>
>> But here is a question: if it does
>> causal relations, it won't be able to do logical form, right? I mean,
>> if you represent solely temporally distinct events, what do you do 
>> about representing simultaneous properties?
>
>  First of all, let me say that we try to refrain from using the word
>"representation".  Representation usually connotes "symbol
>manipulation" and we think that the fundamental underpinnings of
>intelligence and learning have little to do with symbols.  To answer
>your question, note that a network is a distributed, parallel system
>and synchronous events are automatically taken care of by the
>parallelism.  In our model, simultaneity is handled *laterally* and
>sequentiality (deep causality) is extracted *longitudinally*.  As far
>as logic is concerned, it is our firm conviction that it is not a
>fundamental prerequisite to learning.  It is learned just like
>everything else.  IOW, logic emerges.
>
>> (I have the greatest respect
>> for Koch and Sejnowski, of course, but I won't take their word -- if it
>> exists -- for it).
>
>  What Sejnowski and Koch have brought to the surface more than anyone
>else, IMO, is the role that precise signal timing plays in the
>operation of the brain.
>
>Regards,
>
>Louis Savain
>
>President, Marengo Media, Inc.
>
>
>Here are some relevant sites:
>
>http://www.hhmi.org/communic/bulletin/oct95/crack.htm
>http://sanjuan.cs.wwu.edu/mobus/
>http://www.cs.stir.ac.uk/~lss/AntiDromic/AntiDromic.html
>http://www.sciam.com/explorations/020397brain/020397explorations.html
>http://www.sat.t.u-tokyo.ac.jp/javier/wsc1.html
>http://www.klab.caltech.edu/
>http://www.cnl.salk.edu/CNL/
>
About neural timing-- 
It has been known for a long time (cf. work on binaural hearing, the centerline
neurons of the goldfish, and other matters of neural timing) that the dendrite
is an important computational unit. Using Wilfred Rall's model, think of a
dendrite as a cascade of RC circuits, with each synapse able to modulate the
conductance (gain) of its local circuit. This gives a nice "bilinear control
system" model for neurons that is probably consonant with the Volterra series
models used by mathematical biologists like Marmorelis. 
If an excitatory synapse is distal to an inhibitory synapse,
its input is effectively blocked while the inhibiting effect is exercised.
The usual interpretation of learning in this context is that the diameter
of the dendrite, hence its local conductance, can be made to grow by
the increase of tubulin proteins. However, some dendrites show no such
gnarled structure, and one may suppose that in that case, initially many
synaptic connections are made, and most weeded out in the process of learning.

As to Souvain's comments on the irrelevance of 
symbols and logic in modeling the brain: these
points are well taken. IMHO we got off on the wrong track in 1960, trying
to use Shannon Theory to model how we think. 

David





-- 
 David L. Elliott, Vis. Sr. Research Scientist, Inst. for Systems Research,
 University of Maryland, College Park, MD     | delliott@ISR.umd.edu   
 "A pig cannot write if he                    | http://www.isr.umd.edu/~delliott
  has no oink in his pen."  --Col. Lemuel Q. Stuelpnagel