Higher-Level Cognition - Computational Explorations in Cognitive Neuroscience - page 385

Figure 11.8: Example stimuli used in the Wisconsin Card
Sorting Test. Cards vary in three dimensions: shape, color,
and number, with four values along each dimension, as shown
(different hatching patterns indicate different colors). When
matching according to shape, for example, all cards contain-
ing a triangle should be placed under the first card shown in
the figure, regardless of the number or color of the triangles.
Similarly, when matching according to number, all cards hav-
ing two stimuli, regardless of their color or shape, should be
placed under the second card.

Figure 11.9: Example stimuli from the feature & dimension
switching categorization task studied by Dias et al. (1997).
Each row represents one trial. Subjects were rewarded for
choosing the target (indicated by the surrounding box, which
was not presented to the subjects). The stimuli were created
by combining one of two possible line-shapes and one of two
possible filled-shapes (all four possible test combinations are
shown for these line-shape and filled-shape dimensions). The
target was determined by one feature from one dimension, in
this case, the triangle feature from the filled-shape dimension.

oped the strategies for, whereas the top-down biasing
approach is something that all animals with a frontal
cortex could readily apply. Given that the present model
focuses on data from monkeys, it is appropriate that
the top-down biasing approach is used here. Never-
theless, we expect that if similar tasks were run with
people, the purely activation-based approach would be
employed, resulting for example in more rapid overall
learning with frontally intact people.

gued, the active representation can be relatively rapidly
switched when the sorting rules changes, whereas
a weight-based solution must unlearn the previous
weights and learn the new ones. The perseveration
observed in frontal patients on the WCST can be ac-
counted for by the loss of the more flexible, frontally
mediated activation-based facilitation to a less flexible
weight-based one.
The actual task that our model implements, called the
ID/ED task, has been explored in monkeys, neurolog-
ically intact humans, frontal patients, and Parkinson's
patients (Dias, Robbins, & Roberts, 1997; Roberts,
Robbins, & Everitt, 1988; Owen, Roberts, Hodges,
Summers, Polkey, & Robbins, 1993). Instead of sort-
ing, the task involves a two-alternative choice decision,
which simplifies things somewhat. Also, as we will see
in a moment, Dias et al. (1997) found an intriguing
pattern of behavioral deficits associated with lesions of
different frontal areas — this may shed some light on
the topographic organization of frontal representations.

The Dynamic Categorization Task

The task we explore is similar to the Wisconsin card
sorting test (WCST). In the WCST, participants sort se-
quentially presented cards according to one of three dif-
ferent stimulus dimensions (shape, color, and number;
see figure 11.8). They must induce the correct sorting
dimension based on feedback from the experimenter,
and the experimenter periodically switches the catego-
rization rule without notifying the subject. Normal sub-
jects will pick up on this switch relatively quickly, and
adopt the appropriate rule. In contrast, frontal patients
tend to perseverate and retain the current sorting rule
even in the face of mounting errors on the task.
In the top-down biasing model, frontal cortex con-
tributes to the WCST by maintaining a representation of
the currently relevant dimension in frontal active mem-
ory, which focuses top-down activation on the corre-
sponding perceptual processing pathway, thereby facil-
itating sorting along this dimension.

As we have ar-