Conclusions - Computational Explorations in Cognitive Neuroscience - page 415

tracting the relevant semantic information requires sam-
pling over very large bodies of text. It is very difficult
to imagine anyone making a convincing verbal or oth-
erwise purely intuitive argument that this complex and
subtle information could be as useful as it appears to
be. Thus computational models can be essential for ap-
preciating and taking advantage of the complexity of
environments, as well as behaviors.

system in the brain, for example in explaining the func-
tion of the prefrontal cortex. Instead, inhibitory effects
can emerge due to competing activation elsewhere. The
A-not-B model in chapter 9 (Munakata, 1998) and the
Stroop model in chapter 11 (Cohen et al., 1990) both

make this point.
Another important benefit of the explicitness of mod-
els is that they can directly generate predictions. We
have discussed a number of such predictions through-
out the text. Although we think these predictions are
important, we caution against an apparent tendency to
overemphasize them, as if they were the only real test
of a model's value. It should be clear from all of the im-
portant contributions discussed in this section that this
is far from the truth — predictions are one of many dif-
ferent contributions that models can make.

12.5.3 Models Are Explicit

A major consequence of the fact that models force
one to be explicit is that they can help to replace
vague, functionally defined constructs with more ex-
plicit, mechanistically based ones. Cognitive neuro-
science has adopted a number of psychological con-
structs — such as attention, memory, working mem-
ory, and consciousness — that are vague and unlikely
to be the product of single mechanisms. Neural network
models can help the field progress beyond these some-
what simplistic constructs by “deconstructing” them,
and introducing in their stead more mechanistically
based constructs that fit better with the underlying bi-
ology.
For example, we saw that attention can emerge
through the interaction of brain areas with inhibitory
competition mechanisms. The overall process of mul-
tiple constraint satisfaction operating within these in-
hibitory constraints dictates which representations will
dominate in a given situation. Something close to this
view of attention has been proposed from a non neural-
network perspective (e.g., Allport, 1989; Desimone &
Duncan, 1995). The naturalness of this idea from the
neural network perspective both lends considerable sup-
port, and provides an explicit mechanistic basis for it.
Another deconstruction example addressed the vari-
ety of verbal characterizations of distinct memory types
(declarative, explicit, episodic, procedural, semantic,
etc.). We saw that principles of neural computation pro-
vided a more precise and fundamental characterization
of the properties of cortical and hippocampal memory
(O'Reilly & Rudy, 1999; McClelland et al., 1995).
Another example of a deconstructed box is the “dis-
engage” mechanism described in the previous section.
A similar example is the notion of a specific “inhibitor”

12.5.4 Models Allow Control

In virtually every exploration in this text, we have poked
and prodded the networks in ways that experimentalists
can only dream about. With a few simple clicks we
can directly visualize the synaptic connectivities that
underlie neural firing patterns in the models, while at
the same time presenting these models with stimuli and
observing the entire state of activation in response. Hav-
ing this kind of access to the mechanics of the models
leads to levels of understanding that would be impos-
sible to achieve in the intact system. We can leverage
this understanding to establish better and better corre-
spondences between the models and reality, so that our
detailed and sophisticated understanding of the artificial
system can further inform our understanding of the real
one.

12.5.5 Models Provide a Unified Framework

We have stressed the importance of developing and us-
ing a coherent, integrated set of principles for compu-
tational cognitive neuroscience. The broader field of
cognitive neuroscience should also benefit from such
a coherent framework, for a number of reasons, some
of which were alluded to in previous sections (e.g., in
making sense of the data). Another benefit is the ability
to relate two seemingly disparate phenomena by under-