real neurologists

[Brian MacWhinney]

Syd,
   True enough.  Real neurologists don't know or care about connectionism
or thinking, learning, or the human mind.  Real neurologists spend their
time studying individual cells or mapping regions.  They typically regard
information-processing theories about mental functioning as untestable,
given current knowledge.  Fortunately, there is a large and growing group
of neurologists who are willing to transcend these traditional limitations
and who have begun to address learning and processing.  For an example of
this type of work, you may wish to look at books such as "Learning and
COmputational Neuroscience: Foundations of Adaptive Networks" Michael
Gabriel and John Moore (Eds.).  Here you will find articles by people such
as Barto, Kehoe, Schmajuk, Desmond, Sutton and others that represent the
bridge area between real neurologically faithful models and ones that
attempt to form higher level processing abstractions, while still remaining
faithful to the neurological facts.  All of the models require a good
understanding of neurology, as well as a facility with mathematical
modelling.  None of them look like the vanilla connectionism that you might
find in books such as "Parallel Distributed Processing" by Rumelhart and
McClelland.  However, everywhere you will find links and hooks back and
forth between the simple vanilla models and the real
neurologically-grounded models.

   Models that take the details of neuronal functioning seriously and which
capture the intricacy of neuroanatomical patterns are going to be tough to
build.  One area where modelling and real neurological facts seem to be
coming into good contact is in regard to the details of the wiring of local
map topology.  For example, models of Kohonen self-organizing feature maps
closely echo facts of lateral inhibition that are important in setting up
neuronal fields.  Even more interesting is the way in which the neuronal
evidence for the importance of minimizing axon length can be represented as
a useful constraint in neural network models.

   In general, network models vary greatly in the degree to which they
attend to neurological details and known facts.  Given this, it is
important to be careful when declaring that all neural network models are
egregiously out of accord with know facts of neural functioning.  All
models are abstractions and hence somewhat out of accord with details, but
some network models are really quite close to what we know about neural
functioning.  Check out Gabriel and Moore, for example.

--Brian MacWhinney