Higher-Level Cognition - Computational Explorations in Cognitive Neuroscience - page 387

ably, only reversals require the inhibition of learned as-
sociations between the target feature or target dimen-
sion and the response. They specifically dismissed the
notion that an active-memory (“on-line” processing) ac-
count of frontal function could explain the data. How-
ever, as we saw in the Stroop model, the top-down bi-
asing of correct information from active memory can in
fact serve to inhibit incorrect (previous, prepotent) in-
formation.
Dias et al. (1997) further argued that the orbital
area is important for “affective” inhibitory processing,
whereas the dorsolateral area is important for “atten-
tional selection.” Under their logic, the intradimen-
sional reversal requires a reversal of affective asso-
ciations, whereas the extradimensional shift requires
higher-order shifting of attentional focus to another di-
mension. This way of characterizing these different
frontal areas is partially consistent with other notions in
the literature, but also inconsistent with several others.
Our alternative interpretation of these findings con-
sists of two parts. First, the effects of these frontal areas
are seen only in reversal conditions because of the pre-
potent responses that must be overcome, through top-
down biasing from active memory. As discussed above,
this is essentially the same idea that we explored in the
Stroop task, except that here we must establish how
the frontal cortex can switch categorization rules more
rapidly than the posterior system to provide an appro-
priate bias in the reversal conditions.
Second, we posit that different frontal areas represent
information at different levels of abstraction. Specif-
ically, orbital areas contain more detailed representa-
tions of the featural information within a dimension,
and thus facilitate the intradimensional switching of tar-
gets, whereas the dorsolateral areas contain more ab-
stract representations of dimensions per se, and thus
facilitate switching to another dimension. There is
some empirical evidence consistent with such a distinc-
tion between orbital and lateral areas. For example,
a number of neural recording studies in monkeys sug-
gest that ventral (orbital) areas encode object or pattern
information (e.g. Mishkin & Manning, 1978; Wilson,
Scalaidhe, & Goldman-Rakic, 1993). Other researchers
have hypothesized based on a variety of data that the
dorsolateral areas are involved in more complex, ab-

stract processing, whereas the ventral (e.g., orbital) ar-
eas are used for simpler memory processes that require
maintaining specific information (e.g., Petrides, 1996).
The model we explore here implements this posited
distinction between detailed and abstract represen-
tations in different frontal areas, combined with
activation-based memory control mechanisms that en-
able the simulated prefrontal cortex to switch catego-
rization rules more rapidly than the posterior cortex.
We use the same dopamine-based gating principles
developed in chapter 9 to control the updating and main-
tenance of prefrontal representations. Recall that the
dopamine-based system is based on the temporal differ-
ences (TD) reinforcement learning mechanism, which
we explored in chapter 6. We can think of this mech-
anism as producing a form of trial-and-error search .
Specifically, prefrontal representations are initially ran-
domly activated, and persist until errors are made, at
which time they are deactivated by the negative TD
error signal, allowing other representations to be acti-
vated. Thus, in contrast with weight-based learning, the
prefrontal representations can be rapidly deactivated,
and new ones activated very quickly. This speed in dis-
carding a previous categorization rule and switching to
a new one produces the frontal advantage in our model.

11.4.1 Basic Properties of the Model

Figure 11.12 shows the structure of the model. The in-
put layer represents the stimuli in a simple format, with
separate units for the 2 different dimensions in each of
the 2 different locations (left and right). There are 4
units within each dimension, and features are encoded
using simple distributed representations having 2 out of
the 4 units active. The hidden layer is organized in the
same way as the input, but critically it is limited so that
it can only have 2 units active at the same time, so that
once a given target feature has been learned, the hidden
representations of the other features are naturally sup-
pressed. The output response is then produced via con-
nections directly from the hidden layer, which is con-
sistent with the top-down biasing model of frontal con-
tribution — all of the actual learning is performed via
weight-based associations between hidden and output
units.