Higher-Level Cognition - Computational Explorations in Cognitive Neuroscience - page 379

reaction times for word reading and in a strength-based
competition between color naming and word reading
that results in the conflict slowing effect.
Another illuminating piece of behavioral data comes
from a training study in which subjects learned to label
novel shapes with color names (e.g., to respond “blue”
to a squiggle shape) (MacLeod & Dunbar, 1988). Sub-
jects with different amounts of this shape-naming train-
ing were tested on naming the shapes (the trained task)
and the ink colors (the Stroop task), under congruent
and conflicting conditions. For relatively brief amounts
of shape-naming training (one day), conflicting color
interfered with shape naming, just like word reading
usually interferes with color naming in the standard
Stroop task. Meanwhile, color naming was essentially
unaffected by the shapes (again, like word reading nor-
mally is). After more extensive training (20 days), the
shape naming task acted like word reading in the stan-
dard Stroop task — it interfered with color naming,
and was itself relatively insensitive to color. Thus, au-
tomatic versus controlled processing is a relative dis-
tinction, depending on where the competing stimuli lie
along a continuum of strength. All of these aspects of
Stroop behavior are important for understanding the na-
ture of the phenomenon in normals, and they provide
important constraints for the model.
Additional evidence from frontal patients and
schizophrenics speaks directly to the role of frontal cor-
tex in the Stroop task (e.g., Cohen & Servan-Schreiber,
1992; Vendrell, Junque, & Grafman, 1995). Both of
these populations exhibit differentially impaired per-
formance on the color naming conflict condition, in-
dicating an impairment in overriding the prepotent re-
sponse of word reading based on task instructions to
name the color. The frontal patient data clearly sup-
port the idea that frontal cortex is important for this
controlled-processing function. The schizophrenic data
are also supportive, because schizophrenia apparently
involves an impairment of frontal dopamine function
(among other things).
Cohen and Servan-Schreiber (1992) showed that the
schizophrenic data could be modeled by assuming that
dopamine affects the gain of the activity of frontal neu-
rons (e.g., ￿ in equation 2.19 from chapter 2). A re-
duction of gain in the frontal task units in the Stroop

cn wr

gr GR

0 0.751 0.262 0 0.95 0

Hidden PFC

gr GR gr rd

0 0.95 0.95 0 0 0.701

Colors Words Output

Figure 11.5: The Stroop model, where the PFC (prefrontal
cortex) represents the task context (cn = color naming, wr =
word reading). Activity pattern corresponds to color naming
conflict condition, where the color input is red (r), the word
input is green (G), and the task context activated in the PFC
is color naming (cn), which biases the color naming hidden
units (the left two units in the hidden layer) and enables the
network to respond “red” (rd) in the output.

model, corresponding to lower dopaminergic gain in
schizophrenics, produces a differential impairment in
the color naming conflict condition. The net effect of
the gain reduction is to lower the amount of top-down
support to the color naming pathway, supporting the
general idea that the critical contribution of frontal cor-
tex in this task is to provide this top-down support.

11.3.1 Basic Properties of the Model

Figure 11.5 shows the Stroop model. Two different col-
ors and color words are represented, red and green. The
color and word inputs feed into a hidden layer with sep-
arate hidden units for processing color and word infor-
mation. These hidden units, which are intended to rep-
resent posterior-cortical processing pathways, are com-
bined into a single layer to enable them to compete
amongst each other under a kWTA constraint — it is
this competition that produces the slowed reaction times
in the color naming conflict condition. The kWTA con-
straint is set so that there is more stringent competition
within each processing pathway than between pathways
(achieved by setting the k parameter to 1 within sub-