Higher-Level Cognition - Computational Explorations in Cognitive Neuroscience - page 384

a top-down source of activation that can support weak
forms of processing (e.g., color naming). The model
simulated several important features of the behavioral
data, and, importantly, simulated the effects of frontal
damage, thereby more directly supporting frontal in-
volvement in this kind of controlled processing.
There have been a number of follow-ups to the origi-
nal Cohen et al. (1990) Stroop model. The Cohen and
Huston (1994) model improved on the original by us-
ing an interactive (bidirectionally connected) network
instead of a feedforward one. Other papers have ad-
dressed other variations of the task, and other aspects
of the fit to behavioral data (Schooler, Neumann, &
Roberts, 1997; Kanne, Balota, & Faust, 1998; Cohen,
Usher, & McClelland, 1998). However, the somewhat
strange effects of early word SOA's on color naming re-
ported by Glaser and Glaser (1982) have yet to be suc-
cessfully accounted for by a model. Thus, many details
remain to be resolved, but the general principles behind
the model remain sound.

based processing, instead of just pure activation-based
processing. This third alternative is also more consis-
tent with the role of the prefrontal cortex in the Stroop
model, as it involves a top-down biasing of the relevant
features in the posterior cortex, so that it is easier for
weight-based learning to associate these features with
the relevant response.
This top-down biasing is specifically relevant if a
new categorization rule must be learned that conflicts
with a prior rule. Here, the top-down biasing can fa-
cilitate the activation of representations appropriate for
the new rule and enable them to better compete against
the strengthened representations from the prior rule, just
as top-down biasing from the prefrontal cortex enabled
the weaker color-naming pathway to compete with the
stronger word-reading pathway in the Stroop model.
However, if there is no existing imbalance in the poste-
rior representations (e.g., when learning an entirely new
categorization rule), then this top-down biasing would
not make that much of a difference, and learning would
proceed according to the rate of weight-based learning.
The A-not-B model from chapter 9 also has this general
characteristic — the frontal cortex is specifically impor-
tant for overcoming prior biases in where to reach.
As we will see in a moment, the pattern of data
we will be modeling shows this characteristic that new
categorization learning is unaffected by prefrontal le-
sions, while reversals of existing categorization rules
(which involve conflict from prior learning) are af-
fected. In contrast, if we were to adopt the kind
of more purely activation-based categorization solution
discussed in the introduction (involving a comparison
between maintained information and the sensory in-
put), we would expect that both initial acquisition of
rules and reversal learning would be faster because no
weight-based learning at all would be involved. Inter-
estingly, this pattern of data has been found in humans
for easily verbalizable categorization rules (i.e., those
that can be easily maintained in activation-based mem-
ory), which were found to be more rapidly learned than
other nonverbalizable (but equally reliable and salient)
categorization rules

11.4 Dynamic Categorization/Sorting Tasks

In the Stroop model, we simulated the robust mainte-
nance abilities of the prefrontal cortex by simply clamp-
ing the units with appropriate external input. In the
present model, based on O'Reilly, Noelle, Braver, and
Cohen (submitted), we retain the principle of top-down
biasing, and turn our focus to the mechanisms of main-
tenance and updating of prefrontal representations. The
task we explore is a version of the Wisconsin card sort-
ing task, which can also be regarded as a dynamic cat-
egorization task where the rules change periodically.
As discussed in the introduction, categorization tasks
can be solved in two qualitatively different ways — ei-
ther by using weight-based associations between dis-
tinctive features and responses, or by using activation-
based maintenance of the distinctive features that can be
compared with the inputs to generate a response. The
activation-based approach has the distinct advantage of
flexibility — a new categorization rule can be adopted
simply by activating a new set of critical features.
There is actually another way that activation-based
processing can contribute to a categorization task that
involves a combination of activation-based and weight-

(Ashby, Alfonso-Reese, & Wal-

dron, 1998).
Thus, we suggest that the purely activation-based ap-
proach may be something that only humans have devel-