Large-Scale Brain Area Functional Organization - Computational Explorations in Cognitive Neuroscience - page 213

the hippocampus learns rapidly to form separated repre-
sentations that minimize the interference between even
similar memories. Taken together, these interacting sys-
tems, together with the basic principles outlined ear-
lier, constitute a description of the cognitive architec-
ture of the brain. Although aspects of this architecture
are speculative, much of its general character is well
supported by available data.
This cognitive architecture provides a framework for
explaining the cognitive phenomena explored in detail
in the subsequent chapters. In chapter 8, we will ex-
plore models of sensory processing in the posterior cor-
tex, that exploit overlapping distributed representations
and slow, integrative learning to shape these represen-
tations according to the structure of the sensory envi-
ronment. Chapter 10 also mostly explores models of
the posterior cortex, but in the context of language pro-
cessing and mappings between different kinds of repre-
sentations (principally orthographic, phonological and
semantic). Chapter 9 focuses on memory, and explores
the main tradeoffs and specializations that underlie the
major components of the cognitive architecture. Finally,
chapter 11 explores models that incorporate interacting
posterior and frontal components of this cognitive ar-
chitecture in the service of understanding how complex,
temporally extended cognitive processing can emerge
therefrom.
In the sections that follow, we elaborate some of the
properties of this cognitive architecture by comparing
it with more traditional cognitive architectures, and by
mapping some of the main architectural distinctions in
the literature onto this architecture.

1977), which is characterized as a simpler, more direct
association of a stimulus input with a response. Another
way of stating this distinction is that controlled process-
ing is “higher level” cognition, whereas automatic pro-
cessing is “lower level.”
As a general characterization, the more traditional
cognitive models based on production systems and the
computer metaphor (e.g., Anderson, 1983; Newell,
1990) have been concerned with controlled processing
(e.g., problem solving, mathematical theorem proving,
logical reasoning), whereas neural network models have
been concerned with automatic processing (e.g., per-
ception, input-output mappings). Thus, it is important
to see how our neural network-based cognitive archi-
tecture gives rise to controlled processing, and compare
this with more traditional models. Although this is the
primary topic of chapter 11, we will briefly sketch some
of the main ideas here, because they are important for
framing the entire endeavor.
Perhaps the best way to contrast the traditional ap-
proach to controlled processing from our own is in
terms of centralized versus distributed processing. In
many traditional models there is a centralized, control-
ling agent of some sort, one that is surrounded by a
number of more automatic processing systems. Thus,
we see a kind of Cartesian dualism here between the
controller and the controlled, that probably reflects the
compelling and widely shared intuition that a central
soullike entity operates the levers of the complex bio-
logical apparatus of our brains. Needless to say, this
approach begs the question as to how this internal ho-
munculus (“little man”) got to be so smart.
This centralized model is very clear in Baddeley's
(1986) framework, where there is a “central executive”
responsible for all the intelligent controlled process-
ing, while a number of relatively “dumb” slave systems
carry out more basic automatic processes. In Fodor's
(1983) model, the mind is construed as having a central
realm of intelligent, controlled processing, surrounded
by a large number of highly encapsulated modules that
automatically carry out sensory and motor processing.
Perhaps the best example of this kind of architecture is
the source metaphor itself — the standard serial com-
puter. A computer has a centralized processing unit
(CPU) that is surrounded by a number of “dumb” pe-

7.5.1 Controlled versus Automatic Processing

Some of the main differences between more tradi-
tional cognitive models and the cognitive architecture
sketched above can be highlighted by considering the
distinction between controlled versus automatic pro-
cessing. Generally speaking, controlled processing has
to do with the ability to flexibly adapt behavior to differ-
ent task demands, and generally act like an “intelligent”
agent in the world. This kind of processing has classi-
cally been described in contrast with automatic process-
ing (Schneider & Shiffrin, 1977; Shiffrin & Schneider,