Conclusions - Computational Explorations in Cognitive Neuroscience - page 413

be captured as in the models mentioned earlier. Nev-
ertheless, the explicit forms of cognition are obviously
real and are largely unaccounted for by present neural
network models. We revisit this issue next.

been viewed as relatively removed from empirical ap-
proaches, for example because of the biological implau-
sibility of the backpropagation algorithm. Furthermore,
any empirical researchers have felt they can get by with
“common sense” theoretical approaches (e.g., simple
box-and-arrow process models and introspection).
Throughout this topic, we have demonstrated how the
computational/theoretical approach can play an essen-
tial role in cognitive neuroscience, because of the com-
plexity and subtlety of the brain, behavior, and their in-
teractions. We have shown how this approach can actu-
ally speak to the practical concerns facing empiricists.
Here we highlight some of the main contributions using
the same categories as in the introductory chapter.

12.4.4 Capturing Higher-Level Cognition

Initial forays into the challenging task of capturing
higher-level cognition (temporally extended tasks, plan-
ning, task-switching, “executive function,” etc.) were
described in chapter 11, but this is clearly a frontier
area where much more work needs to be done. We
are nevertheless optimistic that the distinction between
activation- and weight-based processing captures an es-
sential aspect of what differentiates explicit/declarative
from implicit/procedural processes. It is clear that neu-
ral network models are just beginning to make progress
in the activation-based processing domain, but we think
that such models will soon provide an important alter-
native to traditional symbolic models for understanding
higher-level cognitive phenomena.
An important question that must be faced is whether
higher-level cognition necessarily requires large-scale
models, or whether important aspects of it can be cap-
tured in smaller more manageable models. In other
words, does higher-level cognition happen only when
a large critical mass of cortex gets going, or can more
basic principles be instantiated in scaled-down models?
We clearly believe the latter to be at least partially true,
suggesting that continued progress can be made with
models of the scale of those presented in chapter 11.

12.5.1 Models Help Us to Understand Phenomena

In general, theories are important for making sense of
behavior. It is one thing to observe patterns of neural
responding in various cortical areas, or to observe dif-
ferent patterns of spared and impaired performance with
brain damage, but quite another to actually make sense
of these findings within a coherent overall framework.
The computational approach can relate cognitive neuro-
science data to a set of functional principles that help us
understand why the brain subserves the behaviors that it
does.
One example is in the explanation of why oriented
bars of light constitute our basic representations of the
visual world. The model in chapter 8 (based on the work
of Olshausen & Field, 1996) shows that these bars of
light constitute the basic statistical structure of the vi-
sual environment, and thus provide a principled basis
for visual representations.
In memory, constructs like repetition priming and se-
mantic priming have traditionally been considered sep-
arate entities. Repetition priming is typically identi-
fied with long-term effects, whereas semantic priming
is typically identified with transient effects. From a
computational/mechanistic basis, the space of priming
effects can be accommodated with three dimensions
(weight versus activation based, content, and similar-
ity, see chapter 9). Within this framework, one can
have both activation-based and weight-based semantic
and repetition priming. Researchers working from the

12.5 Contributions of Computation to Cognitive

Neuroscience

In this section, we highlight some of the main contri-
butions of the computational approach to the broader
field of cognitive neuroscience. In other areas of sci-
ence, computational and formal models are used by
the theoretical branch of the discipline (e.g., theoreti-
cal physics, computational chemistry), and their advan-
tages are clear to all involved. This general appreci-
ation of computational modeling is not as widespread
in cognitive neuroscience. One possible reason for this
lack of appreciation is that computational models have