Memory - Computational Explorations in Cognitive Neuroscience - page 308

tence but fail Piagetian tasks due to performance lim-
itations. Under this assumption, researchers have de-
signed clever experiments that demonstrate earlier signs
of competence. For example, a reduced production of
perseverative errors has bee n observed in gaze and ex-
pectation variants of the AB task. Hofstadter and
Reznick (1996) found that when infants' looking and
reaching behaviors differ in this task, the looking re-
sponse is more accurate. An d, infants make fewer er-
rors in gaze variants of the AB task in which they ob-
serve hidings at A and B without ever reaching (Hof-
stadter & Reznick, 1996; Matthews, 1992; Lecuyer,
Abgueguen, & Lemarie, 199 2) . Finally, in violation-of-
expectation variants of the AB task, 8- to 12-month-old
infants look longer when a toy hidden at B is revealed
at A than when it is revealed at B , following delays at
which they would nonetheless search perseveratively at
A (Ahmed & Ruffman, 1998; Baillargeon & Graber,
1988; Baillargeon, DeVos, & Graber, 1989).
These demonstrations of sensitivity to an object's
new hiding location are commonly treated as eviden ce
that infants “know” where toys are hidden in the AB
task, but search perseveratively due to deficits external
to their knowledge representations, such as deficits in
inhibitory control. However, as we shall see, we can
understand these findings in terms of the competition
between activation- and weight-based memory in the
neural network framework without needing to invoke
notions of reified knowledge and ancillary deficits.

9.6 The Development and Interaction of Memory

Systems

In this section, we explore the development and interac-
tion of activation- and weight-based memory. Whereas
previous sections focused on these forms of memory
separately, here we consider cases where changes in the
weights and persistent activity interact in the same sys-
tem, and sometimes support competing responses. This
competition can arise when the weights have been en-
hanced (primed) for one response pathway in the net-
work (e.g., as a result of repeated practice), but the per-
sistent activity from recent processing favors a compet-
ing pathway. In addition to exploring these interactions,
this model provides an introduction to important issues
in the study of development. Neural network models
provide an important tool for understanding how expe-
rience and maturational/genetic factors can interact in
producing the patterns of changes that take place as an
organis m d evelops into adulthood (Elman et al., 1996).
The AB (“A-not-B”) paradigm (Piaget, 1954), which
has been studied extensively in human infants and sev-
eral animal species, provides a good example of com-
petition between activation- a nd weight-based memory
(Munakata, 1998). In the AB task, subjects watch an
experimenter hide a toy in one location ( A ). They are
typically allowed to search for the object after a short
delay, and this procedure is repeated. Subjects then
watch the experimenter hide the object in a new location
( B ). Following a short delay, human i nf ants often search
perseveratively at A ,makingthe AB error. Weight-
based memory may support perseverative reaching to
A while persistent activity — dependent on the devel-
opment of the prefrontal cortex — can direct correct
reaching to B . Consistent with this, lesions of the pre-
frontal cortex impair i nf ant and adult rhesus monkeys'
performance on the AB task

9.6.1 Basic Properties of the Model

In this model we explore the ways in which activation-
and weight-base d memory can support competing re-
sponses in the AB task. Weight-based memory is im-
plemented as standard CPCA Hebbian learning that oc-
curs as a function of the activity of the units in the
network. Activation-based memory is implemented
by having recurrent connections among representa-
tions in the network, much like those discussed in the
prefrontal-style active memory system in the previous
section. Note that in the original model (Munakata,
1998) and the exercises presented here, these recurrent
weights are increased by hand to simulate the effects
of development; however, the potential role of experi-

(Diamond & Goldman-

Rakic, 1989, 1986).
It is important to note that many r ese archers have crit-
icized Piagetian tests such as the AB task for underes-
timating infants' knowledge, because they require “per-
formance” factors (such as reaching) to demonstrate an
underlying “competence” (such as a concept of object
permanence); infants may have the underlying compe-