Conclusions - Computational Explorations in Cognitive Neuroscience - page 405

Language, explored in chapter 10, requires a number
of specialized processing pathways and representations,
that interact with (and build from) perceptual and other
pathways and representations. These specialized path-
ways operate according to the same principles as any
other pathway. We saw that a distributed, multipath-
way model of word representations can account for var-
ious patterns of dyslexia when damaged. Focusing on
the direct pathway from orthographic input to phono-
logical output, we saw how a network captured reg-
ularities that enabled it to generalize pronunciation to
nonwords in much the same way humans do. Focusing
on the pathway from semantics to phonology, we saw
how a network could learn the regularities and excep-
tions of inflectional morphology, and that it produced a
U-shaped overregularization curve as it learned, much
as children do. We also explored a model of semantic
representation learning based on co-occurrence statis-
tics of words over large samples of text. The resulting
semantic representations capture many relevant aspects
of word similarity. Finally, we saw that a network can
perform multiple constraint satisfaction across both se-
mantic and syntactic constraints in sentence processing,
producing a distributed representation that captures the
overall meaning or gestalt of the sentence. This model
provides a good example of how sequential information
can be integrated over time.
Chapter 11 took on the challenge of applying the bio-
logically realistic neural network mechanisms explored
in previous chapters to modeling higher-level cogni-
tion. We saw that activation-based processing, as im-
plemented in the frontal cortex, is a critical factor in
enabling the kind of dynamic, flexible, and verbally ac-
cessible processing that is characteristic of higher-level
cognitive function. Focusing on the role of the pre-
frontal cortex, we saw that a very simple model can ac-
count for normal and patient performance on the Stroop
task. An important next step was to see how dynamic
control over maintained prefrontal activations via gating
(implemented by the neuromodulator dopamine) leads
to more flexible task performance compared to simple
weight-based task learning. We explored this in the con-
text of a simple dynamic categorization task based on
the Wisconsin Card Sorting task. One major outstand-
ing challenge in the domain of higher-level cognition

is to understand how appropriate prefrontal representa-
tions can be learned over experience to enable this flex-
ible task performance without invoking something like
a homunculus.

12.3 General Challenges for Computational

Modeling

From the brief summary above, it is clear that compu-
tational models have a lot to say about many different
aspects of cognitive neuroscience, and that the current
framework can address an exciting range of phenom-
ena. However, a number of important challenges re-
main for future work. In this section, we revisit the
general problems that computational models face that
we outlined in the introductory chapter, and see how
the models we have explored have addressed these chal-
lenges, and where future work needs to continue to
make progress. In the subsequent section, we address
more specific issues that are faced by neural network
models and the framework adopted here.
First, we note that the history of neural network mod-
eling has been dominated by periods of either extreme
hype or extreme skepticism. In the past, the entire ap-
proach has been rejected based on limitations that were
subsequently overcome (e.g., the limitations of delta-
rule learning). Readers taking a similarly skeptical ap-
proach may find this list of future challenges so long,
or some of the issues here (or others we don't cover)
so damaging, that they question the validity of the en-
tire enterprise. On the other hand, other readers may
be so enamored with existing successes that they ignore
important limitations.
We encourage all readers to strike a balance between
rejecting the approach outright, and simply ignoring re-
maining challenges. Indeed, we feel that in recent years,
the hype that resurged in the '80s has leveled off into
a productive balance of skepticism and optimism (al-
though not usually within the same researcher). Hope-
fully, this balance will continue along with the progress
in future years.