Higher-Level Cognition - Computational Explorations in Cognitive Neuroscience - page 373

egorization rules can be switched. By simply activating
a different active memory representation and applying
the same generic matching operation, an entirely differ-
ent categorization rule can be applied. This speed pro-
duces flexibility, for example enabling one to rapidly
adapt to changing situations (e.g., different categoriza-
tion rules), or rapidly search through a number of dif-
ferent possible rules. In contrast, the weight-based solu-
tion requires slow, constant learning and unlearning to
switch categorization rules. The categorization simula-
tion later in the chapter explores this flexibility property
of activation-based processing.
Activation-based processing also has important ad-
vantages in terms of the accessibility and impact of the
representations. When information is actively main-
tained, it is immediately accessible to other parts of
the system and constantly influences the activation of
other representations. In contrast, encoding informa-
tion by changing weights requires that those weights be
reused for the information to influence behavior, and
the effects tend to be more specific and less acces-
sible to other parts of the system (Munakata, 1998,
in press). One important consequence of accessibil-
ity is that activation-based processing representations
are likely to be more easily verbalized (i.e., accessed
by verbal output pathways) than weight-based process-
ing representations. Going in the other direction, it is
also likely that verbal inputs (e.g., task instructions) can
more easily affect activation-based representations than
weight-based ones.
In short, accessibility may play a defining role in
the distinction between explicit, consciously driven
(“declarative”) cognitive processing and more implicit,
automatic processing. In the categorization example,
the activation-based strategy would make it easy to state
what the defining feature is (“red”), whereas the weight
changes underlying the weight-based strategy are not
directly accessible. Furthermore, the categorization rule
in the activation-based case can be easily influenced by
verbal or other inputs, which need only activate a dif-
ferent feature (e.g., “blue”) to impose a different rule.
The ability of actively maintained information to con-
stantly influence processing elsewhere in the system
is important for focusing and coordinating processing
around a specified goal, as opposed to simply reacting

based on learned input/output associations. This dis-
tinction is typically characterized as one of controlled
versus automatic processing (Schneider & Shiffrin,
1977; Shiffrin & Schneider, 1977). A paradigmatic
example requiring controlled processing is the Stroop
task. In this task, color words (e.g., “red”) are pre-
sented in different colors, and people are instructed to
either read the word or name the color of ink that the
word is written in. Because we have so much experi-
ence reading, we naturally tend to read the word, even
if instructed to name the color — effortful controlled
processing is needed to overcome the prepotent bias to-
ward word reading. As demonstrated by Cohen et al.
(1990) (and explored later in this chapter), the ability to
override the prepotent word-reading response in favor
of color naming can be explained in terms of sustained
goallike activation that favors the color-naming process.
This account is consistent with the finding that frontal
patients, who are thought to be impaired in activation-
based processing, make disproportionately many errors
in the Stroop conflict conditions.
In addition to the activation- versus weight-based
processing distinction, one particularly interesting
source of intuition into the nature of the distinction be-
tween frontal and posterior cortex comes from consider-
ing the nature of dreams. Recent data suggests that the
frontal cortex (and other areas including the thalamus)
is inactivated during REM sleep (Braun, Balkin, & Her-
scovitch, 1998). Thus, the spreading activation of asso-
ciations and generally fuzzy, graded nature of dream-
state cognition is just what we would expect from a
system operating mostly on the basis of the posterior
cortex. The absence of long-term continuity and focus
is also what one would expect without the frontal in-
volvement in controlled processing. Although there are
many other things going on during REM sleep besides
frontal inactivation, it may nonetheless provide a useful
and universally accessible source of intuitions.
Before the kind of activation-based processing
sketched here can produce anything approaching
human-like higher-level cognition, two central issues
must be solved: (a) the control of activation-based pro-
cessing, and (b) the nature of activation-based process-
ing representations. We consider each of these prob-

lems in turn.