Conclusions - Computational Explorations in Cognitive Neuroscience - page 411

that goes along with being a human being? There are
two extreme positions to this issue that are typically
staked out, but we think the truth, as usual, lies some-
where in between. At one extreme are essentially dual-
istic beliefs that amount to the idea that there is some-
thing ineffable about the human brain/mind that can
never be captured in a purely mechanistic system. Even
if people do not explicitly think of themselves as dual-
ists, they view the standard notions of mechanistic sys-
tems, inspired by present day computers and machines,
as so far removed from the sublime complexity of hu-
man experience that a mechanistic reduction seems im-
possible, even repulsive. At the other extreme is the
reductionistic notion that someday psychology will dis-
appear as a field of study and human cognition will
be discussed in terms of neurons, or neurochemicals,
or...where does it end?
As emphasized in the introductory chapter, it is es-
sential to pursue the complementary process of recon-
structionism while engaging in the reductionist enter-
prise. We think that computational models are uniquely
well suited for developing reconstructionist understand-
ing, and we hope that this has become clear in the course
of this text. For example, we spent much of chap-
ter 2 discussing ion channels and their effects on elec-
trical conductances within individual neurons, but we
quickly found in chapter 3 that a whole new terminol-
ogy was necessary to understand the emergent proper-
ties of networks of interacting neurons. The introduc-
tion of learning mechanisms then led to another whole
new set of emergent phenomena. In subsequent chap-
ters, we then relied on these emergent network phenom-
ena to explain complex cognitive phenomena like atten-
tion, object recognition, processing of regularities and
exceptions in language, and so on.
Getting more to the heart of the matter is the issue
of consciousness. Although we have not focused much
on this issue, it has received a considerable amount of
attention lately from computational modelers and other
theoreticians. One interesting theme emerging from this
work is that one can usefully characterize the properties
of things that are within the scope of conscious aware-
ness in terms of the duration, persistence, stability, and
level of influence of representations (e.g., Kinsbourne,
1997; Mathis & Mozer, 1995).

properties that are within the purview of a reconstruc-
tionist modeling endeavor, and we anticipate that future
models will inevitably continue to make progress in un-
derstanding the “ineffable.”

12.3.5 Modeling Lacks Cumulative Research

Neural network models are often criticized for their lack
of cumulative research, even very recently (e.g., Gaz-
zaniga, Ivry, & Mangun, 1998). As we noted in the in-
troduction, this criticism might be applied to any new
field where there is a lot of territory to be covered —
indeed, it would be easy to level this same charge at the
current explosion of neuroimaging studies. We hope
that this topic helps to allay this criticism in the do-
main of computational cognitive neuroscience. We have
revisited, integrated, and consolidated a wide range of
ideas from many years of computational modeling re-
search. It should be very clear that the field as a whole
has developed a set of largely consistent ideas that pro-
vides a solid basis for further refinement and explo-
ration, as in any maturing scientific discipline.

12.4 Specific Challenges

Having discussed challenges that apply generally to the
computational modeling endeavor, we now turn to a set
of more specific challenges faced by neural network
models and the specific modeling framework adopted in
this topic. As with the more general challenges, some
of these have been met by existing models but are listed
here because they continue to be leveled at neural net-
work models, while others remain for future models to
resolve.

12.4.1 Analytical Treatments of Learning

Mathematically based analyses of neural networks are
an important part of the overall endeavor. These pro-
vide in-general proofs that our algorithms are likely to
achieve something useful, and they also provide insights
into the behavior of these algorithms. In developing the
Leabra algorithm, we included such analyses wherever
we could.

These are emergent