What's wrong with Hebbian synapse models?

Stephan Anagnostaras stephan at psych.ucla.edu
Tue Apr 4 09:14:34 EST 1995

In article <3lmgib$1gt at crl5.crl.com>, rising at crl.com (Hawley K Rising) wrote:

> I am trying to get through a book on large scale neuronal theories (Koch 
> and Davis, Large Scale Neuronal Theories of the Brain,1994 MIT Press).  
> In the introduction they say that although they've had a pervasive effect 
> on models, most theories about brain function don't last a decade.  An 
> example they give is the Hebbian synapse (synapse strength grows in 
> proportion to correlated activity of the pre and post synaptic neurons).  
> What was wrong with this theory and can I get a source to read about 
> whatever replaced it as acceptable?
> Thanks in advance,
> Hawley Rising
> rising at crl.com

Well, the basic hebbian model is totally flawed.  If the synaptic strength
grew based on correlated activity, a single most correlated association
quickly comes to dominate the entire network.  However, models that
call themselves "Hebbian" nowadays always put in some way to limit
the growth and spread of a single association to avoid this problem.
There are other problems, in that this model does not properly
account for learning phenomena, where the idea of a single 
associative strength is flawed, but this is a more complicated issue.

None of the neural network models are "accepted" within the field
in general; connectionism is largely an outsider group, however, a 
powerful and vocal minority.  Even among connectionists, there is
poor agreement on what types of connections and associative networks
to build, so there is no "replacement," so to speak. This will depend
on whose book you read.

Many neural network models of the past emphasized arbitrary theoretical
networks which were clearly untenable positions if you are trying to
create a theory of the brain (rather than a theory of artificial
intelligence).  As such a number of people are attempting to limit
which connections exist (as opposed to not exist) within a particular
system. This requires a great deal of lab work with actual animals,
which makes this position tediously slow.  However, you could try reading
Churchland & Sejnowski's book, the Computational Brain, for a view of
this position.



More information about the Neur-sci mailing list