Higher-Level Cognition - Computational Explorations in Cognitive Neuroscience - page 375

In this section, we discuss a number of different ideas
about the general nature of frontal activation-based pro-
cessing representations. Unlike representations in the
visual system, for example, where it is likely that hu-
mans are similar to monkeys and relatively much is
known (chapter 8), human frontal representations may
be qualitatively different than those in monkeys, be-
cause our capacities for higher-level activation-based
processing are qualitatively different. Therefore, we
have no solid empirical data about what kinds of things
neurons in human frontal cortex actually encode, so we
must draw on a diverse set of more functionally driven
ideas, including inspiration from symbolically based
production systems, performance on Stroop and cate-
gorization tasks, and language-based representations.
Because traditional symbolically based production
systems have been successful at modeling higher-level
cognitive tasks such as problem solving (e.g., Ander-
son, 1983; Newell, 1990), it is useful to consider the
idea that frontal representations may resemble the kinds
of productions used in these systems. A production is
an elemental cognitive action that is activated when a
specific set of conditions is met. The effect of produc-
tion activation is ultimately to activate subsequent pro-
ductions, typically by manipulating the contents of ac-
tive memory to trigger the enabling conditions of other
productions. Thus, productions control the flow of pro-
cessing in a goal-driven manner via these enabling con-
ditions — one could imagine a similar weight-based
mechanism for detecting the enabling conditions of a
production. However, symbolic production systems
take liberal advantage of arbitrary symbol binding op-
erations, whereas the goallike active memory represen-
tations in the neural system may have a more content-
specific, dedicated nature. Thus, we are not advocating
the use of neural mechanisms that simply reimplement
existing symbolic processing frameworks (see chapter 7
for discussion). Nevertheless, the general notion of a
production might be a useful way of conceptualizing the
nature of the representations used in activation-based
processing.
Another relatively simple idea is that some of
the activation-based processing representations encode
stimulus-level information. In the categorization exam-
ple, correct task performance can be achieved by sim-

ply maintaining active representations of the discrimi-
nating features (e.g., “red”). At a somewhat more ab-
stract level, the Stroop task can be solved by maintain-
ing an active representation that specifically favors pro-
cessing in the color-naming pathway — this could be a
representation corresponding to the sensory perception
of color, for example.
Language-based representations also provide a po-
tential source of insight into the nature of frontal rep-
resentations, in two different senses. In a strictly
metaphorical sense, frontal representations may consti-
tute a “vocabulary” that can be systematically combined
to produce an effectively infinite combinatorial space of
context, goal and task representations. In a more literal
sense, the considerable time and effort spent develop-
ing the ability to combine flexibly verbal representa-
tions into complex, meaningful structures is undoubt-
edly directly leveraged by the frontal cortex.
Thus, the frontal cortex may have a strong tendency
to rely on linguistic representations in the service of
complex task performance. Support for this idea is
the nearly ubiquitous finding of activation localized
around Broca's area during working memory or other
frontal tasks (e.g., Awh, Jonides, Smith, Schumacher,
Koeppe, & Katz, 1996; Braver, Cohen, Nystrom,
Jonides, Smith, & Noll, 1997; Paulesu, Frith, & Frack-
owiak, 1993; Fiez, Raife, Balota, Schwarz, Raichle, &
Petersen, 1986). This area is commonly thought to sub-
serve the phonological loop function engaged in verbal
rehearsal (Baddeley, 1986). Thus, a simple, intuitively
appealing idea is that people engage in verbal rehearsal
of task-relevant information during the performance of
complex tasks.
One overarching characterization of frontal represen-
tations, which is consistent with the above ideas, is
that they are more discrete in nature, whereas posterior
cortical representations encode information in a more
continuous fashion (O'Reilly, Mozer, Munakata, &
Miyake, 1999b; O'Reilly et al., 1999a). For example,
productions are discrete because they either fire or not,
and language is based on discrete symbols (words). In-
terestingly, discreteness can confer increased noise tol-
erance for working memory representations, as we ex-
plored in chapter 9, suggesting a synergy between a
number of constraints.