no-modularity-at-all

[Brian MacWhinney]

Dear FunkNet,
   Sorry about the delay in replying to Tom Givon's note from last week
about the inadequacy of no-modularity-at-all connectionism.  He rightly
emphasizes the extent to which most extant connectionist models tend to
minimize the modular organization of the human mind.  In particular, Givon
notes that "the real bridge between the level we are working at and
neurology is most like not at the level of individual neurons and their
connections" but rather at the level of interconnected modules.  Givon
furthermore argues that the patterns at the lower level are comparatively
universal and therefore could not be the source of the unique human
language capacity.
  I think Givon's point is important for the readers of this list.  I agree
that connectionism has not properly demonstrated its
relevance to functionalist linguistics.  At the best, connectionists
(myself included) have presented a system for neural-like notation
that allows one to believe that the problems of symbolic representation,
language processing, and language acquisition might eventually be solved
from an emergentist perspective.  The value of this demonstration is simply
that it provides functional linguists (additional) license to dismiss
claims regarding the psychological reality of generative grammar.  However,
in practice, functional linguists often have their own empirical reasons
for
constructing alternative accounts of language processing.
   When a functional linguist looks to psycholinguistics for insight, it is
perhaps a bit disappointing to find psycholinguists obsessed with modelling
fine details of the acquisition of the past tense.  When functionalists
look to neurology, it is equally discouraging to find no anatomical level
grounding of the human-primate discontinuity.
   Givon argues that the problem is that connectionists need to be looking
at larger models that connect more modules.  I agree completely.  In
fact, I have recently contrasted models that look at "local maps" with
models that look at "functional neural circuits", much along lines that
match Givon's points.  I have argued that models now need to look at both
levels.  Where I disagree with Givon is on my assessment of
the current status of connectionist work.  First, I can point to many
recent
connectionist models that look at the functional level.  Consider Prahlad
Gupta's connectionist model of Baddeley's phonological memory loop.  The
memory loop is a classic example of a functional neural circuit that
involves the superior temporal, motor cortex, hippocampus, and possibly
additional lexical areas.  By looking at how people manage to refresh
concepts in this loop, we learn about memory for discourse, sentence
production, and other things Givon and the rest of us care about.
    Second, I am a bit surprised that Givon has not realized that his
interest in evolution means that he must pay attention to local anatomical
patterns.   We know there are subtle neuroanatomical differences between
man and primate in frontal cortex and elsewhere.  My understanding of brain
evolution is that changes in patterns of connectivity emerge as a result of
subtle changes on the microscopic level.  It is my understanding that it is
exactly these low-level differences that support the higher-level
functional abilities Givon refers to.  Givon chides connectionists for
reductionism, but Nature herself is a reductionist.

--Brian MacWhinney