Higher-Level Cognition - Computational Explorations in Cognitive Neuroscience - page 372

However, this relative ease can also lead symbolic mod-
els to over estimate the power and flexibility of human
symbolic reasoning.
For example, many symbolic models have no diffi-
culty with simple abstract logic problems that confound
people. In the neural network framework, higher-level
cognition instead emerges from interactions among the
kinds of more basic neural mechanisms covered in pre-
vious chapters, and it remains strongly constrained by
them. This framework clearly predicts that people
should have difficulty with abstract logic problems, and
that they should have less difficulty with more concrete,
familiar versions of such problems, which would en-
gage the rich semantic representations built up over ex-
perience in a neural network. Indeed, people are much
better at solving logic problems when they are stated
in more concrete and familiar terms (e.g., Cheng &
Holyoak, 1985; Johnson-Laird, Legrenzi, & Legrenzi,
1972).
Thus, we view the challenge of this chapter as one
of building a more flexible, dynamic cognitive system
out of neural mechanisms that use dedicated, content-
specific representations, without relying on implausi-
ble symbolic machinery (even as implemented in neu-
ral hardware; see chapter 7). The resulting cognitive
system should be capable of appropriate levels of de-
liberate, explicit processing involving relatively abstract
constructs, and it should be able to chain together a se-
ries of cognitive operations over a relatively long period
of time, all focused on achieving a specific goal or series
of goals.

a) Weight−based b) Activation−based

(no−match)

AB A B

{

Active
memory

red blue sqr tri

match

red blue sqr tri red blue sqr tri Input

Figure 11.1: Weight- versus activation-based processing in
the context of a categorization task. (a) weight-based solu-
tion involves associating discriminating feature (“red”) with
category A, other features (or just “not red”) with category B.
(b) activation-based solution maintains critical feature in ac-
tive memory, and does a match comparison with the current
input, such that if input matches this feature, the response is
A, otherwise it is B. The match unit could detect enhanced
activation from mutual support between input and maintained
feature. The activation-based solution is much more rapidly
updatable (i.e., flexible), and the categorization rule is easily
accessible to other parts of the system.

prefrontal working memory system developed in chap-
ter 9, is likely to be essential for achieving the flexi-
ble, dynamic style of processing that lies at the heart
of higher-level cognitive processes such as explicit, ab-
stract reasoning, planning, and problem solving.
To make the activation- versus weight-based process-
ing distinction more concrete, consider the example
(which we model later) of performing a simple cate-
gorization task (figure 11.1). The task is to catego-
rize stimuli as either A sor B s based on their distin-
guishing features (e.g., respond A for red stimuli, B for
any other color). A weight-based solution to this prob-
lem would involve learning associations (weights) be-
tween the critical input feature units and the category
response units. An alternative activation-based solution
involves maintaining in active memory a representation
of the discriminating feature(s) (e.g., “red”), and then
performing a matching operation that compares the sim-
ilarity of the maintained features with those of the cur-
rent input, and responds appropriately (e.g., if the input
matches the maintained feature “red,” respond A ,else

The Importance of Activation-Based Processing

The core idea behind our approach to this challenge
involves a distinction between activation- and weight-
based processing, which builds on the distinction be-
tween activation- versus weight-based memory as dis-
cussed in chapter 9. Activation-based processing is
based on the activation, maintenance, and updating of
active representations to influence cognitive process-
ing, whereas weight-based processing is based on the
adaptation of weight values to alter input/output map-
pings. An appropriately configured and dynamically
controlled activation-based memory system, like the

).
The activation-based solution, though more complex,
has a distinct advantage in the speed with which the cat-