The Role of Implicit Memory in Controlling a Dynamic System

Abstract

The relationship between implicit memory and implicit learning is explored. Dienes and Fahey (1995) showed that learning to control a dynamic system was mediated by a look-up table consisting of previously successful responses to specific situations. The experiment reported in this paper showed that facilitated performance on old situations was independent of the subjects' ability to recognize those situations as old, suggesting that memory was implicit. Further analyses of the Dienes and Fahey data replicated this independence of control performance on recognition. However, unlike the implicit memory revealed on fragment completion tasks, successful performance on the dynamic control tasks was remarkably resilient to modality shifts. The results are discussed in terms of models of implicit learning and the nature of implicit memory.

Full text

1
The role of implicit memory in controlling a dynamic
system
Zoltán Dienes and Richard Fahey
University of Sussex, England
Running Head: IMPLICIT MEMORY AND SYSTEM CONTROL
Corresponding author: Zoltán Dienes, Experimental Psychology,
Sussex University, Brighton, Sussex, BN1 9QG. Fax: 01273
678611. E-mail: dienes@epunix.susx.ac.uk

Implicit Memory and System Control 2
Abstract
The relationship between implicit memory and
implicit learning is explored. Dienes and Fahey
(1995) showed that learning to control a dynamic
system was mediated by a look-up table consisting of
previously successful responses to specific
situations. The experiment reported in this paper
showed that facilitated performance on old
situations was independent of the subjects' ability
to recognize those situations as old, suggesting
that memory was implicit. Further analyses of the
Dienes and Fahey data replicated this independence
of control performance on recognition. However,
unlike the implicit memory revealed on fragment
completion tasks, successful performance on the
dynamic control tasks was remarkably resilient to
modality shifts. The results are discussed in terms
of models of implicit learning and the nature of
implicit memory.

Implicit Memory and System Control 3
Acknowledgements
This research was supported by the Medical
Research Council. Many thanks to Dianne Berry, Bob
Mathews, David Shanks, Michael Stadler, and Daniel
Willingham for valuable comments. Richard Fahey is
now at Computer Data Systems Inc., 8626 Tesoro
Drive, Suite 114, San Antonio, TX 78217, USA.

Implicit Memory and System Control 4
The role of implicit memory in controlling a
dynamic system
Two distinctions in the human learning
literature have become very influential: Implicit
versus explicit memory (e.g., Roediger & McDermott,
1993; Schacter, 1987); and implicit versus explicit
learning (e.g. Berry & Dienes, 1993; Reber, 1989;
Shanks & St John, 1994). Implicit rather than
explicit memory is shown on tasks that do not
require deliberate recollection of a past event
although the event influences performance. Implicit
rather than explicit learning is shown on tasks for
which the subject learns to make the right decision
(more often or more quickly) without being able to
justify the decision (Berry & Dienes, 1993). The
inability to justify is meant to indicate that
learning occurs without concurrent awareness of what
is being learned (Reber, 1989). The distinctions
between implicit and explicit memory and between
implicit and explicit learning are logically
orthogonal: For example, subjects may remember the
episode in which they learned to make a right
decision without being able to justify the decision.
The purpose of this paper is to explore the
relationship between implicit memory and the
implicit learning of a complex system (Berry &
Broadbent, 1984; Dienes & Fahey, 1995). Initially,

Implicit Memory and System Control 5
relevant evidence in the implicit memory literature
will be described. Then, recent findings and
theoretical proposals in the implicit learning
literature that indicate a relationship with
implicit memory will be overviewed. Next, an
experiment will be reported that explores the role
of implicit memory in controlling a dynamic system.
Finally, the data collected by Dienes and Fahey
(1995) will be re-analyzed in the same way to
establish the generality of the findings.
Previous research has determined a number of
characteristics of performance on implicit rather
then explicit memory tasks. A typical implicit
memory task is fragment completion: Subjects first
study a list of words (e.g., AARDVARK), and are then
shown fragments of words (e.g., _AR_VA__) and the
subjects are asked to complete the fragment with the
first word that comes to mind. The fragments of
words that have rather than have not been studied
previously are more likely to be completed. This
learning, or priming, has two important
characteristics that can be used to assess its
relevance to implicit learning.
First, the ability to complete the fragment (or
perform some other implicit memory task) is
stochastically independent of the ability to recall
or recognize the word (e.g., Hayman & Tulving, 1989;

Implicit Memory and System Control 6
Tulving, 1985; Tulving & Schacter, 1990; Tulving,
Schacter, & Stark, 1982). The interpretation of
this finding is controversial. Hintzman and Hartry
(1990) argued that the low level of dependence
merely reflects many sources of variance affecting
fragment completion in addition to priming; in
contrast Tulving (e.g., Tulving & Schacter, 1990)
argued that the low level of dependence reflects the
operation of distinct memory systems underlying
implicit and explicit memory tasks. The dispute is
as yet unresolved (see Flexser, 1991; Gardiner,
1991; Hintzman, 1991; Poldrack, 1996). One solution
to the problem of many sources of variance is to
show that the data are nonsignificantly different
from independence, and also that there was
sufficient sensitivity to detect a meaningful amount
of dependence (Ostergaard, 1992; Tulving & Hayman,
1993; Ostergaard, 1994 see Appendix A). This
paper will adopt this solution.
The second characteristic is that amnesics can
display unimpaired performance on implicit but not
explicit memory tasks (e.g., Warrington &
Weiskrantz, 1978). This finding suggests that
implicit memory may be based on a different memory
system to explicit memory (Roediger, 1990; Tulving,
1985; Tulving & Schacter, 1990; contrast Shanks,
1997). Finding a memory task that is spared in

Implicit Memory and System Control 7
amnesia would be suggestive evidence that the task
relies on the memory system or systems underlying
implicit memory tasks generally.
Turning now to the implicit learning
literature, one key paradigm is the control of
complex systems (e.g., Berry & Broadbent, 1984,
1987, 1988); other key paradigms are artificial
grammar learning (e.g., Reber, 1989) and sequential
reaction time tasks (e.g., Willingham, Nissen, &
Bullemer, 1989). In the "dynamic control tasks"
used by Berry and Broadbent, the subject controls
the level of one variable (e.g., their friendliness
towards a computer personality) in order to reach
target values on another variable (e.g., the
computer personality's friendliness to the subject).
Subjects acquire considerable knowledge about how to
control such systems, as indicated by their
progressive ability to reach and maintain target
values. This knowledge appears to be implicit
because subjects find it difficult to describe how
to reach and maintain target values (Stanley,
Mathews, Buss, & Kotler-Cope, 1989). Stanley et al
(1989) asked subjects after every 10 trial block to
give complete instructions on how to perform the
task. The informativeness of these instructions was
assessed by the performance of yoked subjects asked
to follow the transcribed instructions. Stanley et

Implicit Memory and System Control 8
al demonstrated that sudden improvements in
performance by the original learners were not
associated with simultaneous increases in the
informativeness of the instructions. In fact,
instructions helped the performance of yoked
subjects only if the instructions were taken at
least four blocks after the improvement in
performance. Note that there are other types of
task for which the instructions given by subjects
immediately and completely account for their
performance (e.g., Mathews, Buss, Chinn, & Stanley,
1988; Schwartz, 1966).
Recent theoretical views on implicit learning
suggest a relationship with implicit memory. Both
Broadbent, Fitzgerald, and Broadbent (1986) and
Stanley et al. (1989) argued for a mechanism for the
implicit learning of the control tasks that could
rely on implicit memory. They suggested that
whereas explicit knowledge is based on a mental
model of the system, implicit knowledge is based on
memory for specific events related to the control
task. This memory-based view of implicit learning
is consistent with theoretical proposals in other
domains of implicit learning (see, e.g., Brooks,
1978; Cho & Mathews, 1996; Neal & Hesketh, 1997;
Perruchet, 1994; Vokey & Brooks, 1992; Whittlesea &
Dorken, 1993), but the issue has aroused some

Implicit Memory and System Control 9
controversy (e.g. Reber, 1967, 1989). At stake is
the question of how sophisticated unconscious
processing is. Reber argued that implicit learning
results in an abstract knowledge base that goes well
beyond storage of specific exemplars. Cleeremans
(1993) showed how the successful computational
models of implicit learning of sequential reaction
time tasks could produce knowledge that lay along a
continuum of abstractness depending on task
conditions: in this case, the mechanism underlying
implicit learning tasks can go beyond memories for
instances but it falls short of formulating explicit
rules. Following Broadbent et al, we have argued
(Dienes and Fahey, 1995) that, for the dynamic
control tasks, one does not need to postulate a
mechanism any more complicated than memory for
specific events.
Broadbent et al. (1986) suggested that in
learning a dynamic control task a subject could
construct a "look-up table" which would determine
the appropriate action by matching the current
situation to the most similar of the entries already
in the table. Using methodology introduced by
Marescaux, Luc, and Karnas (1989), Dienes and Fahey
(1995) tested and confirmed several predictions of
the look-up table approach in two experiments. In
Experiment 1, subjects were trained on a simulated

Implicit Memory and System Control 10
sugar production factory, in which they manipulated
the level of work force in order to reach target
values of sugar production. After training,
subjects were given a specific situation task in
which the subject was presented with hypothetical
situations (e.g., "if you had just employed 400
workers and if the sugar production was then 8000
tons, what should you do next to bring the sugar
production to target?"). Some of the situations
were new (i.e. the subject had not experienced them
in the training phase). Some of the situations were
old (i.e. the subject had experienced the situation
in training) and the subject had given a correct
response to them (i.e. a response that lead to
target or closeby) or an incorrect response; these
old situations were called 'correct situations' and
'incorrect situations', respectively. (Note that
calling a situation correct refers to whether the
response given in the context of the situation was
correct in the training phase, not to whether the
situation contained the target as a feature.)
Crucially, subjects' experience in controlling a
dynamic system did not allow subjects to perform
above chance on new situations; their knowledge
seemed only to apply to old correct situations.
Further, subjects tended to give the same response
to old correct situations as they had given

Implicit Memory and System Control 11
originally; the tendency to give the same response
to old incorrect situations was smaller. These
results were replicated in Experiment 2 with a
person interaction task in which subjects
manipulated the friendliness of a computer
personality. Further, detailed aspects of subjects'
performance in both experiments could be fitted with
a one-parameter computational model of a look-up
table (Logan, 1988).
Given that control of these tasks appears to be
mediated by memory of specific instances, one can
ask whether the memory is the sort that typically
underlies performance on explicit memory tasks or if
it is of the sort that underlies performance on
implicit memory tasks. Evidence consistent with the
latter alternative was provided by Squire and
Frambach (1990): They found that amnesics were not
initially impaired on the dynamic control tasks,
just as Warrington and Weiskrantz (1978), for
example, found that amnesics were not impaired on
implicit memory tasks. More direct evidence for
normal subjects can be provided by the data
collected by Dienes and Fahey (1995). They
presented subjects with a recognition test on the
same situations that had been used on the specific
situations task, allowing a test of stochastic
independence between recognition and performance on

Implicit Memory and System Control 12
the specific situations task. Dienes and Fahey did
not test the relation between recognition and
control performance: This paper will do so. First a
new experiment will be described that tests the
relation between recognition and control
performance, and then a reanalysis of the Dienes and
Fahey data will be reported. If there is a look-up
table based on explicit conscious recollection of
previous situations, there should be a dependence
between performance on the specific situations task
and recognition of the situation as old; if it's
based on implicit memory there should be no
dependence.
The dependency of the subjects' tendency to
respond correctly to a given situation on explicit
memory for that situation is important in assessing
whether the control tasks are learned unconsciously.
If correct responding is dependent on the subjects'
ability to recall the previously successful episode,
and subjects could justify their responding in terms
of the episodic memory, the claim for unconscious
learning (Hayes & Broadbent, 1988; Stanley et al.,
1989; contrast Sanderson, 1989; Shanks & St John,
1994) would need reconsidering.
Experiment 1
In Experiment 1 subjects were trained on the
sugar production factory task and then presented

Implicit Memory and System Control 13
with specific situations from the training phase.
Each subject was asked to both give a control
response to the situation (specific situations task)
and to indicate whether they recognized the
situation as old or not (recognition task). The
primary aim of the experiment is to assess whether
the look-up table used by subjects in controlling
the sugar production task was based on implicit
memory. This was achieved by assessing (1) the
degree by which the tasks differed from stochastic
independence; and (2) the degree to which the tasks
differed from a plausible model of dependence (the
Ostergaard, 1992, model, described in the results
section below and in more detail in Appendix A, and
also endorsed by Poldrack, 1996).
In order to conclude that subjects used an
implicit look-up table, the following three points
need to be established:
A. Subjects learnt the task by acquiring simply a
look-up table. Dienes and Fahey (1995) provided the
evidence for this claim.
B. The tasks do not differ significantly from
stochastic independence.
C. The tasks do differ significantly from the model
of dependence.
This paper will attempt to establish the
validity of points B and C. If B is established but

Implicit Memory and System Control 14
not C (i.e. the tasks do not differ significantly
from either stochastic independence or from the
model of dependence) then the data are consistent
with subjects using either an implicit or an
explicit look-up table, or some combination of the
two. If the data are significantly different from
independence but not significantly different from
the model of dependence then the data suggest
subjects are using an explicit look-up table.
Method
Subjects. The subjects were 20 paid volunteers aged
between
18 and 35 from the Sussex University subject
panel.
The task. Subjects in Experiment 1 were trained on
the sugar production task introduced by Berry and
Broadbent (1984). Subjects were asked to imagine
they were in charge of a sugar production factory.
They were told they could change the amount of sugar
produced by changing the size of the work force.
Their goal was to achieve and maintain a target
sugar output of 9,000 tons. The starting work force
was 600 workers and starting level of sugar output
was 6,000 tons. On each trial, subjects entered a
number between 1 and 12 on the computer keyboard to
represent the number of hundreds of workers they
wished to employ on that trial. The level of sugar

Implicit Memory and System Control 15
production on trial n was determined by the equation
Pn=2*W-P
n-1 + N, where Pnis the number of
thousands of tons of sugar output on trial n, W is
the number of hundreds of workers employed by the
subject, and N is noise (N could be -1, 0, or +1
with equal probability). If the equation resulted
in a sugar output of less than 1,000 tons, the
output was simply set at 1,000 tons; similarly, if
the equation resulted in an output of greater than
12,000 tons, it was set at 12,000 tons. Subjects
were aware of these lower and upper limits. Because
of the noise, N, in the equation, subjects'
responses were counted as correct if they resulted
in an output on target or one level off.
Optimal performance on this task requires the
subject to take account of the current level of
sugar production. In fact, Wn+1 should be exactly
halfway between Pnand target in order to reach the
target on the next trial.
Procedure. All subjects were trained for two blocks
of 40 trials of the sugar production factory. Each
block started with a sugar production of 6,000 tons
and a work force of 600 workers. On each trial,
subjects saw a graph indicating the level of
performance on all previous trials of that set. A
horizontal line indicated the target performance.
In addition, written information about the level of

Implicit Memory and System Control 16
work force and the level of sugar production for the
last trial was presented above the graph (see Figure
1).
----------------------------------------------------
Insert figure 1 about here
----------------------------------------------------
The computer kept a record of all situations
the subject came across. A situation was defined as
the current level of sugar production. The
situations were tabulated into those for which the
subject entered a workforce that resulted in the
target sugar output or only one level off (loosely
correct situations), and those that were followed by
an output more than one level from target (incorrect
situations). Note that calling a situation correct
refers to whether the response given in the context
of the situation was correct, not to whether the
situation contained the target as a feature.
Next subjects were exposed to specific
situations on the computer and performed either the
specific situations task introduced by Marescaux et
al (1989) or a recognition task. In both tasks
subjects were presented with a level of sugar
production: The statement 'The current level of
sugar production is X tons' appeared in the centre
of the screen in an identical way for both tasks.
All 12 levels of sugar production were presented in

Implicit Memory and System Control 17
a different random order for each task. In the
specific situations task, subjects were told to
enter the level of work force they thought would
achieve or maintain the target level of output,
based on their previous experience of controlling
the factory. The same target level was used as in
the training phase. Subjects were told that after
each situation, the next situation to be shown would
be unrelated to the work force they had just
entered; it would simply be another possible
situation, and thus, they would get no feedback on
how successful they were being. In the recognition
task, subjects were told that some of the situations
will have been ones they came across in their first
task. Subjects responded "old" or "new" to each
situation depending on whether they could remember
seeing it in the training phase. The specific
situations and recognition tasks were presented in
counterbalanced order.
Results
Order had no effect on any measure (ps > .20),
including the measures of dependence and
independence, so this factor will not be discussed
further. Not all subjects produced data for which
all dependent variables could be calculated (for
example, if the subject had experienced all levels
of sugar production in the learning phase there

Implicit Memory and System Control 18
would be no new situations in the later phases, and
so the Ostergaard (1992) measure could not be
calculated for that subject), and so the degrees of
freedom of the subsequent analyses vary slightly as
a consequence.
Learning. The number of trials that were loosely
correct improved from the first block of trials
(8.2, SD = 4.9) to the second block of trials (13.0,
SD = 7.5), t(19) = 2.85, p = .01.
Specific situations task. Reliable priming was
obtained: Subjects performed better on old levels
of sugar production (.26, SD = .20) rather than on
new levels (.07, SD = .27), Wilcoxon p = .03.
Recognition. Mean recognition (64%, SD = 26%) was
significantly above chance, t(19) = 2.46, p = .02.
Stochastic dependence. The contingency table for
the specific situations and recognition tasks is
shown in table 1. The entries were summed over all
subjects and correct situations. The proportion of
situations expected to be both correctly recognized
and correctly responded to on the specific
situations task based on assumed independence
between the tasks was separately calculated for each
subject. The difference between the actual
proportion and that expected on the basis of
independence was, averaged over subjects, .001 (SD
= .062), which is non-significantly different from

Implicit Memory and System Control 19
zero, t < 1, analysing over subjects. That is,
there was no evidence that giving a correct response
to a situation in the specific situations task
depended on recognizing the situation.
A null result is only compelling to the extent
that the data were sensitive enough to detect a
reasonable effect size. Appendix A describes a
model based on Ostergaard (1992) that can be used to
estimate the amount of dependence expected if
learning consisted of a look-up table based on
explicit memory. The logic of the model is as
follows. Subjects can to some extent give correct
answers to each task (specific situations task and
recognition task) not because they have learnt to
respond to specific situations but because of the
pre-existing strategies they bring to the task or
because they simply guess. These additional sources
of variance increase the expected amount of
independence between the tasks; it is only to the
extent that learning has occurred that dependence
can emerge between the tasks. To take into account
these additional sources of variance, the proportion
of correct responses on new situations was
subtracted from the proportion of correct responses
on old situations to estimate the 'study effect' of
each task; i.e. the effect that exposure to
situations has on ability to perform the task above

Implicit Memory and System Control 20
and beyond what subjects can do without learning. It
was assumed that subjects could recognize all the
situations that contributed to the study effect of
the specific situations task (this is the assumption
of an explicit look-up table); thus the proportion
of situations that subjects performed correctly on
both tasks will be the study effect for the specific
situations task, plus an additional component due to
correct performance on both tasks for independent
reasons. Thus, if the study effect for the specific
situations task were zero, the Ostergaard (1992)
model would make the same prediction as the model of
independence. As more learning occurs on the
specific situations task, the more the Ostergaard
(1992)model departs from the model of independence.
For more explanation see Appendix A. The difference
between the actual proportion correct on both tasks
and that expected on the basis of Ostergaard's model
of dependence was, averaged over subjects, -.11 (SD
= .15), t(13) = 2.74, p = .017. That is, the
dependence between the tasks was less than would be
expected if correct performance on the specific
situations task was based on recognition of the same
situations (see Appendix B for a further analysis
showing the robustness of this result). One measure
of the effect size of the difference of subjects'
data from the model of dependence is Cohen's (1988)

Implicit Memory and System Control 21
d, defined as the mean difference divided by the
standard deviation, i.e. .11/.15 = 0.73, close to
what Cohen calls a large effect (d = 0.80).
The key result of Experiment 1 was that
performance on the specific situations task did not
depend on the subjects' ability to recognize old
situations as old. The analysis of Experiment 2 of
Dienes and Fahey (1995), presented next, tests the
generality of this finding.2
Experiment 2 of Dienes and Fahey (1995)
Experiment 1 used a simulated sugar production
factory as the control task. Another commonly used
task is the person interaction task. Berry and
Broadbent (1988) and Hayes and Broadbent (1988)
employed two versions of a person interaction task
that they claimed were learned in distinctive
implicit and explicit modes. Experiment 2 reported
by Dienes and Fahey (1995) employed both versions of
the person interaction task. The major purpose of
this analysis of Experiment 2 was to determine the
relationship between recognition and performance on
the specific situations task for the two versions of
the person interaction task.
A secondary aim of this analysis of Experiment
2 was to determine if there was any effect of
modality shift on control task performance. One of
the characteristics of implicit rather than explicit

Implicit Memory and System Control 22
memory is its frequent sensitivity to changes in
surface features, and in particular to shifts
between the auditory and visual modalities (e.g.
Bassili, Smith, & MacLeod, 1989; Berry, Banbury, &
Henry, 1997; Roediger & Blaxton, 1987). Dienes and
Berry (1997) also argued that the knowledge acquired
in a number of implicit learning paradigms is
inflexible and perceptually bound. For example, in
the artificial grammar learning paradigm, subjects
trained on colour patches show a decrement in
performance when tested on colour names, and vice
versa (Dienes & Altmann, 1997). Berry and Broadbent
(1988) found there was no transfer between two
control tasks with the same underlying equation when
the cover story changed (e.g. from being a transport
task to a person interaction task), even when
subjects were informed of the critical relationship
between the tasks.
Schacter (1990) presented one interpretation of
the finding of modality specificity in implicit
memory tasks. According to Schacter, performance on
implicit memory tasks is mediated by a set of
modular perceptual representation systems (PRSs)
that process information about the form and
structure of objects but do not represent
associative information about them. Different types
of objects have their own PRS; for example, there is

Implicit Memory and System Control 23
a word form system underlying performance on
fragment completion, and a structural description
system underlying the form processing of common
visual objects. According to Schacter, priming on
many implicit memory tasks reflects the
establishment of highly specific representations
within an appropriate PRS; it is these
representations that underlie the sensitivity of
priming to modality shifts. However, not all
implicit tasks rely on a PRS. For example, in
conceptual priming, after studying the word
'limerick', subjects may be given a conceptual cue
('What name is given to a lighthearted five line
poem?'). The priming obtained in this paradigm is
stochastically independent of recognition but also
insensitive to modality shifts (Challis, Chiu, Kerr,
Law, Schneider, and Yonelinas, 1993). Priming on
such tasks may be due to associative links being
formed in a semantic memory system, and not in a
PRS. Variability in performance on the dynamic
control tasks is not a measure of variability in the
subject's perception of the structure of any type of
object. Perceiving the situation is trivial in the
dynamic control tasks; the problem is purely in
learning the right response to a given situation,
that is, in learning the associative links. Thus,
Schacter's (1990) model does not predict any effect

Implicit Memory and System Control 24
of modality shift for the dynamic control tasks, in
contrast to some other implicit tasks (e.g. Servan-
Schreiber & Anderson, 1990, regarded learning in the
artificial grammar learning paradigm as an example
of perceptual learning). The current reanalysis of
Experiment 2 tested whether knowledge acquired about
the person interaction tasks behaved like stem
completion and artificial grammar learning
(perceptually bound) or like conceptual priming (not
perceptually bound).
Method
In Experiment 2 of Dienes and Fahey (1995),
each of 48 subjects was randomly allocated to one of
the cells of a 2X2X2(Person [Person S vs
Person U] by learning modality [visual vs auditory]
by testing modality [visual vs auditory]) between
subjects design. Equal numbers of subjects were
allocated to each of the cells. Subjects were told
that they would be meeting a computer person called
Ellis, and that they could communicate their level
of friendliness to Ellis through the keyboard.
Ellis would respond with how friendly he is to the
subject. The aim of subjects was to move Ellis to a
target value of friendliness and keep him there.
Friendliness varied along a 12 point scale: Very
Rude, Rude, Very Cool, Cool, Indifferent, Polite,

Implicit Memory and System Control 25
Very Polite, Friendly, Very Friendly, Affectionate,
Very Affectionate, and Loving.
Berry and Broadbent (1988) and Hayes and
Broadbent (1988) employed two equations for
controlling Ellis' behaviour, a "salient" equation
and a "nonsalient" equation. The salient equation
was En=S
n-2+N,where Enis a number between 1
and 12 representing Ellis' behaviour on the 12 point
scale on trial n, Snis the subject's behaviour on
the 12 point scale on that trial, and N is noise (-
1, 0, or +1 with equiprobability). The nonsalient
equation was En=S
n-1 -2+N. Because of the
noise, a response of Ellis up to one off target was
counted as correct. Hayes and Broadbent (1988)
called the personality controlled by the salient
equation "Person S" and that controlled by
nonsalient equation "Person U". Note that for the
nonsalient equation Ellis' behaviour depends not on
how the subject just responded on that trial, but on
how the subject responded one trial back. The
target in Experiment 2 was Polite for both Person S
and Person U, and so the optimal strategy when
interacting with either would be to always enter
Friendly.
In the learning phase, subjects interacted with
Ellis for one block of 30 trials for Person S and
one block of 50 trials for Person U. Subjects in

Implicit Memory and System Control 26
the visual learning condition saw two bar charts,
one representing Ellis' behaviour on the previous
trial, and the other their behaviour on the previous
trial. A horizontal line in Ellis' chart indicated
the target behaviour (Polite). The bars moved up
and down according to Ellis' and the subject's
respective behaviours (see Figure 1). Subjects in
the auditory learning condition did not see the
computer screen. The experimenter simply said "Your
behaviour was X; Ellis' behaviour was X". The scale
of possible behaviours was placed in front of the
subjects to remind them. All subjects entered their
response by typing in the corresponding initials
(e.g., VA for Very Affectionate). The computer kept
a record of all situations the subject came across.
A situation was defined as what the subject could
see, i.e. the current level of Ellis' behaviour and
the subject's behaviour on the last trial.
Next subjects were exposed to specific
situations and performed either the specific
situations task or the recognition task on the
computer. In the specific situations task, subjects
were shown possible situations consisting of Ellis'
and the subject's behaviours on the preceding trial.
Subjects in the visual testing condition saw the
information displayed as bar charts; subjects in the
auditory testing condition heard the experimenter

Implicit Memory and System Control 27
read the information out. Subjects were told to
enter the behaviour they thought would achieve or
maintain the target level of Ellis' behaviour, based
on their previous experience interacting with Ellis.
The target level was the same as used in the
training phase. Subjects were told that after each
situation, the next situation to be shown would be
unrelated to the behaviour they had just entered; it
would simply be another possible situation, and
thus, they would get no feedback on how successful
they were being. In the recognition task, subjects
were told that half the situations will have been
ones they came across in their first task, and half
would be new situations. Subjects responded "old"
or "new" to each situation. The old situations for
both tasks were all situations experienced in the
training trials.
The specific situations and recognition tasks
were run in counterbalanced order using the design
employed by Tulving, Schacter, and Stark (1982).
Specifically, for each subject, the computer
randomly generated a number of new situations equal
to the number of old (i.e. half the situations were
not "study primed" - the new situations - and half
were study primed - the old situations). Half of
the new situations and half of the old situations
were randomly assigned to one stimuli set (set A),

Implicit Memory and System Control 28
the other half to set B. Half the subjects were
run, first, on the specific situations task with set
A stimuli; then the recognition task with sets A and
B; and finally, the situations task with set B. The
other half of the subjects were run on, first, the
recognition task with set A stimuli; then the
specific situations task with sets A and B; and
finally, the recognition task with set B. (Sets A
and B constituted different situations for different
subjects.)
Results
Learning. Subjects were scored for the number of
trials correct in the first set of 10 trials and in
the last set of 10 trials. A2X2X2X2(Block
[first set of 10 trials vs last set of 10 trials] by
Person [Person S vs Person U] by learning modality
[visual vs auditory] by testing modality [visual vs
auditory]) analysis of variance on the number of
trials correct indicated a significant effect of
person, F(1,40) = 10.47, p < .005, and of block,
F(1,38) = 14.29, p = .0005. That is, subjects
scored more trials correct for Person S (6.5) than
for Person U (4.9). Also, subjects performed better
on the second block (6.5) rather than the first
(5.0). The interaction of block with person was not
significant, p > .10. The improvement from the
first to the second block was 1.3 for Person U

Implicit Memory and System Control 29
(t(23) = 2.65, p < .05) and 1.7 (t(23) = 2.92, p <
.01) for Person S.Specific Situations Task.A
response was scored as correct on the specific
situations task if it would lead to target, or at
most one level off, at least two thirds of the time.
A2X2X2X2(person [Person S vs Person U] by
learning modality [visual vs auditory] by testing
modality [visual vs auditory] by priming [old versus
new]) mixed model analysis of variance on proportion
of correct responses indicated a significant effect
for person, F(1,40) = 13.62, p < .001, and priming,
F(1,40) = 50.62, p < .0001. That is, subjects
performed better with Person S (62%) rather than
Person U (43%). Also, subjects performed better on
situations which they had (63%) rather than had not
seen before (41%) in the person interaction task.
An important result is that a change in
modality between learning and testing had no effect
on performance. Table 2 shows the means for Persons
S and U for subjects in the same and different
modalities. In fact, subjects performed numerically
better in the different rather than the same
modality for both Persons S and U. Could the
experiment have suffered from a lack of power to
detect the effect? In the case of implicit memory
as revealed by stem completion, Bassili et al.
(1989) found that priming reduced from 32% in the

Implicit Memory and System Control 30
same modality groups to 20% in the different
modality groups (averaged across their experiments
one and two); i.e., a proportional reduction of 37%.
Brown, Neblett, Jones, & Mitchell (1991) reviewed
the effect of modality shift on implicit memory in
10 other studies, and in all cases the proportional
reduction was greater than 37% (their own study was
an exception). The power of the current experiment
to detect a proportional reduction in performance of
37% in both Persons S and U was greater than .99;
the power to detect such a reduction in Person U
alone was .89; and the power to detect an
interaction in which the drop occurs for Person U
but not Person S was .65.
Recognition. Overall recognition performance was
74%, significantly greater than chance, t(40) =
16.32, p < .0001. A2X2X2(person [Person S vs
Person U] by learning modality [visual vs auditory]
by testing modality [visual vs auditory]) analysis
of variance indicated a significant effect of
person, F(1,40) = 6.33, p < .05. That is, subjects
had a greater recognition performance with Person S
(78%) rather than Person U (70%).
Stochastic Dependence. The contingency table for
the specific situations and recognition tasks is
shown in table 3 for Persons S and U. The entries
were summed over all subjects and correct

Implicit Memory and System Control 31
situations. The proportion of situations expected
to be both correctly recognized and correctly
responded to on the specific situations task based
on assumed independence between the tasks was
separately calculated for each subject. The
difference between the actual proportion and that
expected on the basis of independence was, averaged
over subjects, -.00, for Person U, and +.01 for
Person S, both non-significantly different from
zero, ts < 1, analysing over subjects. That is,
there was no evidence that giving a correct response
to a situation in the specific situations task
depended on recognizing the situation.
The difference between the actual proportion
correct on both tasks and that expected on the basis
of Ostergaard's model of dependence was, averaged
over subjects, -.10, t(20) = 3.48, p < .01, d =
0.76, for Person U, and -.09, t(23) = 4.65, p <
.001, d = 0.95, for Person S. That is, the
dependence between the tasks was less than would be
expected if correct performance on the specific
situations task was based on recognition of the
situations (see Appendix B for further analyses
showing the robustness of this result for Person U).
Discussion
The aims of this analysis of Experiment 2 of
Dienes and Fahey (1995) were, first, to determine

Implicit Memory and System Control 32
the relationship between recognition and performance
on the specific situations task, and, second, to
determine whether the dynamic control tasks are
sensitive to modality shifts. In terms of the first
aim, the results indicated that subjects' ability to
respond to a situation in the specific situations
task did not depend on their ability to recognize
that situation as old, for both Persons S and U. In
the case of Person U, Dienes and Fahey provided
evidence that subjects learned by storing
appropriate responses to specific situations:
subjects' did not perform at above baseline levels
on new situations and further details of subjects'
performance could be fit by a one-parameter look-up
table model.3The amount of dependence was less
than that predicted by Ostergaards' model,
suggesting that performance on the specific
situations task was based on implicit memory.
Person S also showed independence between task
performance and recognition. Dienes and Fahey
(1995) provided evidence that subjects learnt Person
S (unlike Person U) by inducing rules that
generalize beyond specific situations; specifically,
Dienes and Fahey showed that subjects could perform
well on new situations and that subjects'
performance could be fit well by a rule-based model.
Once a subject has learnt a rule, there is no reason

Implicit Memory and System Control 33
to suppose that its successful application depends
on recognition of old situations. Thus, the lack of
dependence in the case of Person S does not indicate
that subjects acquired an implicit look-up table.
The results showed that a modality shift (from
visual to auditory presentation or vice versa) had
no influence on either Person U or Person S4.In
terms of Schacter's (1990) model, this result
suggests that variability in implicit learning on
the person interaction task does not reflect
variability in the perceptual representation systems
underlying performance on standard implicit memory
tasks. Rather learning the person interaction tasks
involves forming associative links between easily
perceived situations and appropriate responses.
Berry (1991) did find a remarkable inflexibility in
transfer of subjects' knowledge of this link from
situations to responses: Subjects could not transfer
their knowledge gained in the context of making a
verbal response on a dynamic control task to make a
typing response. Further, Roediger (1990) and
Tulving and Schacter (1990) argued that implicit
tasks are often sensitive to modality shifts not by
virtue of their implicitness but because they
predominantly involve perceptual processes.
According to Tulving and Schacter, implicit priming
effects can occur on conceptual tasks that do not

Implicit Memory and System Control 34
involve changes in a perceptual representation
system but modifications to semantic memory. For
example, both normals and amnesics show priming in a
task in which subjects are given the name of a
category and are asked to produce the first instance
that comes to mind. Although the knowledge acquired
in learning to control a complex system may be
inflexible over some surface changes (e.g. the
change in cover story used by Berry and Broadbent,
1988; the change in response requirements used by
Berry, 1991), it may be less perceptually bound than
the knowledge acquired in artificial grammar
learning tasks (Dienes & Altmann, 1997).
Discussion
This paper has reported data from two
experiments showing that, when controlling a dynamic
system (the sugar production factory and person U),
being able to respond successfully to specific
situations does not depend on recollective
experience. Rather, the subject appears to know the
correct response because of implicitly formed
associations: implicit learning in these cases is
based on implicit memory.
The conclusion that implicit learning is based
on implicit memory, depends on the claim that
subjects used almost exclusively a look-up table in
controlling the sugar production factory and person

Implicit Memory and System Control 35
U. If subjects used abstract rules then there is no
reason why performance on the specific situations
task should be dependent on performance on the
recognition task. Dienes and Fahey (1995) argued
that control of the sugar production task and of
person U was indeed based almost exclusively on a
look-up table for the amount of training used in the
experiments reported in this paper. Specifically,
Dienes and Fahey found that subjects performed at
baseline levels on new situations; and they also
showed that a one-parameter look-up table model
could simulate levels of performance, subjects
consistency of responding in different situations,
and rate of learning.
Nosofsky (1988, 1991) showed that previous
apparent dissociations between recognition and
classification performance were perfectly consistent
with one memory system underlying both tasks.
Specifically, he showed that if classification is
based on the relative similarity of the test
exemplar to stored exemplars in the different
categories, and recognition is based on the total
similarity of the test exemplar to all stored
exemplars, there can be a zero correlation between
recognition and classification performance. This
dissociation resulted from the different decision
rules for recognition and classification: that is,

Implicit Memory and System Control 36
total compared to relative similarity. Could there
be a similar explanation for the dissociation
between recognition and dynamic control task
performance found in this paper? Dienes and Fahey
(1995) argued that learning to respond appropriately
in a specific situation involves remembering that
specific situation, or a very similar one. Memory
for a specific situation is what the recognition
test was supposed to directly measure. But it may
be that subjects, when making a recognition
judgement, relied on some overall similarity of the
test situation to all training situations. To test
this possibility in Experiment 1, the similarity was
computed between each to-be-recognized situation and
each training situation. Of course, any test
depends on a model of how similarity is computed and
used by subjects. Initially, for simplicity,
similarity was measured on a linear scale that
varied between 1 and 12:
Similarity = 12 - |sugar production in test
situation - sugar production in training situation|.
For each test situation, similarity was summed
over all training situations. For each subject,
test situations were grouped into those that
contributed to hits and those that contributed to
misses on the recognition test; similarity was
averaged separately for hits and misses. If

Implicit Memory and System Control 37
similarity determined recognition judgements, then
it should be higher for hits than for misses.
Indeed, the mean similarity was 675 (SD = 40) for
hits and 623 (SD=79) for misses, t(17) = 3.25, p =
.005, analysing over subjects.
These data are consistent with subjects using a
summed similarity strategy, but also could arise for
other reasons. For example, similarity probably
covaried with how often a test situation occurred in
the training phase (we will call this the frequency
of a situation). On almost any theory, frequency
would affect probability of recollecting a specific
situation5. Performance on the specific situations
task was based not only on memory for the exact
situation, but also on situations numerically
adjacent (Dienes & Fahey, 1995). It may be that
subjects' recollection of situations confused
numerically adjacent situations6. Frequency was
computed by counting the number of times either the
exact test situation occurred in the training phase,
or another situation that was one only one level of
sugar production different. The frequency for hits
(21.1, SD = 6.2) exceeded that for misses (16.6, SD
= 3.4), t(17) = 2.92, p = .010. When similarity
difference was regressed against this frequency
difference, the intercept (21) was not significantly
different from zero, t(16) = 1.34, p = 0.20. That

Implicit Memory and System Control 38
is, as far as we can tell, the similarity of a test
situation to training situations more than one level
of sugar production different did not influence
subjects' tendency to say 'old' on the recognition
test.
In summary, performance on both the specific
situations task and the recognition task were
insensitive to situations that were more than one
level of sugar production different. Thus, our
response to the Nosofsky (1988, 1991) type challenge
is that both the specific situations task and the
recognition tasks seemed to test memory for specific
situations in the same way; there did not appear to
be different decision criteria. The analyses of
dependence between the two tasks showed they were
independent even when recognition of a situation
only one level of sugar production different was
treated as a correct recognition of the original
situation (Appendix B). We suggest that this
independence arose because the two tasks tapped
different memory systems7. If both tasks depended
on familiarity, it is unclear why they should be
stochastically independent. Thus, we suggest that
the recognition task relied on recollective
experience (Gardiner & Parkin, 1990; Tulving, 1985).
Subjects were told on the recognition test to say
'old' if they remembered seeing the situation

Implicit Memory and System Control 39
before; that is, if they had a recollective
experience and not just a sense of familiarity. On
the other hand, performance on the specific
situations task may have relied on implicit
mechanisms that give rise to a sense of familiarity.
Future research could test whether encouraging
subjects to respond on the recognition task
according to familiarity would produce correlations
between the two tasks.
One point of debate in the implicit learning
literature has been whether subjects acquire
unconscious or conscious knowledge (Berry & Dienes,
1993; Hayes & Broadbent, 1988; Reber, 1989; Lewicki,
1986; Shanks & St John, 1994). Our results indicate
that in the case of the dynamic control tasks the
knowledge is conscious in the sense used by Shanks &
St John: Given a specific situation, the subject
can say what the appropriate response should be.
That is, knowledge of the link between situation and
response is conscious. On the other hand, the
knowledge is unconscious in the sense that subjects
do not know why they know the correct response (this
is what Shanks & St John call implicit retrieval):
They do not know that it is because they have been
in that specific situation before. There are also,
of course, other senses in which knowledge can be
unconscious or conscious (see Chan, 1992; Dienes,

Implicit Memory and System Control 40
Altmann, Kwan, & Goode, 1995; Dienes & Berry, 1997;
Dienes & Perner, 1996) which these results do not
address: future research could usefully do so.
Finally, we note that the results reported in
this paper may be bounded by the amount of learning
subjects receive. With more extended learning
periods, subjects acquire greater amounts of
explicit knowledge (Stanley et. al., 1989; Squire &
Frombach, 1990). Subjects learn general rules, but
they may also come to explicitly remember specific
situations. Further, the implicit learning itself
may become less well approximated by a look-up table
as learning proceeds, and subjects may progressively
extrapolate and interpolate to greater degrees
around the situations trained on. This is also a
matter for future research to address.

Implicit Memory and System Control 41
Appendix A
Ostergaard's (1992) model of dependence
Ostergaard argued that in the implicit memory
literature previous studies had not considered a
model of what the greatest degree of dependence
could be between two tasks. He provided a model of
dependence between memory tests that takes into
account the variance not related to the study
episode, as a way of assessing the greatest degree
of dependence that could be expected between two
tasks. Showing that the dependence between two
tasks is nonsignificant is only informative if the
degree of dependence is less than the amount that
could be expected. We can use the Ostergaard model
to determine the degree of dependence that would be
expected if correct responding on the specific
situations task was based on explicit recognition.
The effect of study on each task is calculated as
(proportion of positive responses to studied items)
- (proportion of positive responses to nonstudied
items); or, in symbols, P(S+) - P(N+). 'Positive
response' refers to a correct response on the
specific situations and to a 'yes' response on the
recognition task. Call the task with the smallest
study effect task 1; the study effect for this task
can be represented as P(S1+ from study) = P(S1+) -
P(N1+). If all the traces of items accessible on

Implicit Memory and System Control 42
task 1 are also accessible on task 2, then the P(S1+
from study) of the study items given a positive
response on task 1 will also be given a positive
response on task 2. That is, in our case, it is
assumed that the subject can recognize the
situations that contributed to the study effect of
the specific situations task; this is the assumption
of an explicit look up table. There will be other
situations that subjects can give a correct response
to in the specific situations task, however, because
of chance, or the use of a pre-existing explicit
strategy that is independent of learning. The
proportion of such situations can be estimated by
P(N1+). Some of these situations will be recognized
but could not form the basis of an explicit look-up
table; for example, they may be recognized as old,
but the subject could not remember the response to
them. The proportion of situations jointly correct
on both tasks for independent reasons can be
estimated by P(N1+) X P(S2+). In total, the
proportion of positive responses given to both tasks
is P(S1+) - P(N1+) + P(N1+) X P(S2+). Thus, the
proportion of situations responded to correctly on
both tasks can be predicted by the model to be the
smaller study effect plus an additional chance
component.

Implicit Memory and System Control 43
This calculation assumes that the study effect
for task 1 is constant across the possibly different
contexts and instructions of the two tasks. This is
a reasonable assumption for our application:
Situations were presented to subjects in exactly the
same detail on the specific situations and
recognition tasks. Thus, we could reasonably expect
the degree of dependence between the specific
situations and recognition tasks predicted by the
Ostergaard model if performance on the specific
situations task relied on explicit recognition.
In this paper, degree of dependence was
calculated separately for each subject and analyzed
over subjects. This is so that a Simpson's paradox
could not be present in the data due to collapsing
over subjects (but it could of course arise due to
other reasons). We cannot discount all possible
artifactual reasons for why two tasks may fail to
correlate - there can always be unknown covariates,
but we can discount various plausible reasons, and
therefore make a plausible case that the
demonstrated independence is theoretically
informative. Making plausible cases is all we can
do as scientists in any situation.
The Ostergaard model takes into account the
fact that the effect of study could be smaller for
one task rather than the other, and the level of

Implicit Memory and System Control 44
dependence is calculated accordingly. For example,
If subjects use an explicit look up table, they
would have to (1) recognize the current situation as
an old one or not; (2) if it is old, recall whether
the target followed the situation or not; and (3),
if it did, recall the response given to the
situation. The recognition test used in this paper
only assessed the first component of this process,
and thus only a subset of the situations recognized
as old would be given a correct response in the
specific situations task. This just means that the
effect size on the recognition task should be larger
than for the specific situations task. Consistent
with the assumption of an explicit look-up table, we
assumed that the population effect of study would be
greater for the recognition task than for the
specific situations task for each subject. The
calculations were conducted accordingly: In the
expression used to predict the proportion of
positive responses to both tasks [i.e. P(S1+) -
P(N1+) + P(N1+) X P(S2+)], task 1 (i.e. the task
meant to have the smallest effect of study) was
always the specific situations task.
Because the Ostergaard analysis was calculated
separately for each subject, the expression was
often determined for cells where there were small
amounts of data. The effect of small numbers of

Implicit Memory and System Control 45
observations is always to reduce power. In terms of
Type I errors, note that the expected value of the
Ostergaard estimate of proportion of positive
responses on both tasks can be obtained by
substituting the expected value of each of its terms
in the expression. (This statement assumes that the
proportion of positive responses to nonstudied items
on task 1 is largely independent of the proportion
of positive responses to studied items on task 2 for
each subject, because these two proportions are
involved in a product term. This is a reasonable
assumption and in fact is an assumption made by the
Ostergaard analysis in general.) The expected value
of each term is not changed by small numbers. Thus,
small numbers of observations do not distort the
expected value of the Ostergaard prediction, and an
analysis based on small numbers will not involve any
inflation of Type I error rate. That is, finding a
significant difference between the Ostergaard
predicted proportion and the obtained proportion
could not be explained by the small numbers of
observations.
Note that independence between the recognition
and specific situation tasks could arise simply by
responding "old" to everything. This type of
problem is addressed by Ostergaard's analysis. If
subjects responded "old" to everything, the

Implicit Memory and System Control 46
difference between independence and the degree of
dependence predicted by Ostergaard's model would be
zero. If data are significantly less than the
prediction of Ostergaard's model, then trivial
explanations of this sort are ruled out.

Implicit Memory and System Control 47
Appendix B
Effect of generalization between situations on the
Ostergaard model
The analyses of the degree of dependence
between the tasks reported in the results sections
of Experiments 1 and 2 assumed that there was no
generalization between situations. In fact, Dienes
and Fahey (1995) showed that there was
generalization between situations. For the sugar
production task, Dienes and Fahey showed that
subjects treated situations that differed by only
one level of sugar production as if they were the
same (there was little generalization between
situations that differed by two or more levels of
sugar production). This generalization could
increase the apparent amount of independence between
performance on the recognition and specific
situations tasks. For example, subjects may fail to
recognize a situation as old, but still produce the
correct response on the specific situations task
because they remember a similar situation. To deal
with this potential problem, the analyses on the
data from Experiment 1 were repeated but subjects
were regarded as having correctly recognized an old
loosely correct situation either if (1) they
correctly recognized that particular situation; or
(2) they correctly recognized another loosely

Implicit Memory and System Control 48
correct situation that was only one level of sugar
production different and to which they gave a
correct response in the specific situations task.
The difference between the actual proportion of
situations in which subjects performed correctly on
both tasks and the proportion expected on the basis
of independence was, averaged over subjects, .004
(SD = .043), which is non-significantly different
from zero, t(19) < 1. That is, there was no
evidence that giving a correct response to a
situation in the specific situations task depended
on recognizing that situation or a similar
situation. The difference between the actual
proportion correct on both tasks and that expected
on the basis of Ostergaard's model of dependence
was, averaged over subjects, -.05 (SD = .08), t(13)
= 2.16, p < .05, d = 0.63. That is, the dependence
between the tasks was less than would be expected if
correct performance on the specific situations task
was based on recognition of the situation or a
similar one1.
In terms of the person interaction task, Dienes
and Fahey (1995) showed that when interacting with
Person U, subjects treated situations as similar if
they differed by no more than four levels of Ellis'
behaviour or the subjects' behaviour. As for
Experiment 1, this generalization could increase the

Implicit Memory and System Control 49
apparent amount of independence between performance
on the recognition and specific situations tasks.
To deal with this problem, the analyses conducted on
the data from Experiment 2 were repeated but the
subject was regarded as having correctly recognized
an old loosely correct situation either if (1) they
correctly recognized that particular situation; or
(2) they correctly recognized another loosely
correct situation that was no more than four levels
of Ellis' behaviour or the subjects' behaviour
different, and to which they gave a correct response
in the specific situations task. The difference
between the actual proportion and that expected on
the basis of independence was, averaged over
subjects, -.01 for Person U, and .00 for Person S,
both non-significantly different from zero, ts<1.
That is, there was no evidence that giving a correct
response to a situation in the specific situations
task depended on recognizing the situation or a
similar situation. The difference between the
actual proportion correct on both tasks and that
expected on the basis of Ostergaard's model of
dependence was, averaged over subjects, -.06, t(20)
= 2.70, p < .02, d = 0.59, for Person U, and .00, t
<1,d< 0.20, for Person S. That is, for Person U,
the dependence between the tasks was less than would
be expected if correct performance on the specific

Implicit Memory and System Control 50
situations task was based on recognition of the
situation or a similar one. For Person S, the data
did not distinguish between Ostergaard's model and
the model of independence.

Implicit Memory and System Control 51
References
Bassili, J. N., Smith, M. C., & MacLeod, C. M.
(1989). Auditory and visual word stem completion:
Separating data-driven and conceptually driven
processes. Quarterly Journal of Experimental
Psychology,41, 439-453.
Berry, D. C. (1991). The role of action in implicit
learning. Quarterly Journal of Experimental
Psychology, 43A, 881-906.
Berry, D. C., Banbury, S., and Henry, L. (1997).
Transfer across form and modality in implicit and
explicit memory. The Quarterly Journal of
Experimental Psychology, 50A, 1-24.
Berry, D. C., & Dienes, Z. (1993). Implicit
learning: theoretical and empirical issues. Hove:
Erlbaum.
Berry, D.C. & Broadbent, D.E. (1984). On the
relationship between task performance and
associated verbalizable knowledge. Quarterly Journal
of Experimental Psychology,36, 209-231.

Implicit Memory and System Control 52
Berry, D. C., & Broadbent, D. E. (1987). The
combination of explicit and implicit learning
processes. Quarterly Journal of Experimental
Psychology,39, 585-609.
Berry, D.C. & Broadbent, D.E. (1988). Interactive
tasks and the implicit-explicit distinction.
British Journal of Psychology,79, 251-272.
Berry, D. C. and Dienes, Z. (1993). Implicit
learning: Theoretical and empirical issues. Hove:
Lawrence Erlbaum.
Broadbent, D. E., Fitzgerald, P., Broadbent, M. H.
P. (1986). Implicit and explicit knowledge in the
control of complex systems. British Journal of
Psychology,77, 33-50.
Brooks, L. (1978). Nonanalytic concept formation and
memory for instances. In E. Rosch & B.B. Lloyd
(Eds.), Cognition and Categorization (pp.169-211).
Hillsdale, N.J.: Erlbaum.
Brown, A. S., Neblett, D. R., Jones, T. C., &
Mitchell, D. B. (1991). Transfer of processing in
repetition priming: Some inappropriate findings.

Implicit Memory and System Control 53
Journal of Experimental Psychology: Learning,
Memory, and Cognition,17, 514-525.
Challis, B. H., Chiu, C.-Y., Kerr, S. A., Law, J.,
Schneider, L., Yonelinas, A., & Tulving, E. (1993).
Perceptual and conceptual cueing in implicit and
explicit retrieval. Memory,7, 127-151.
Chan, C. (1992). Implicit cognitive processes:
theoretical issues and applications in computer
systems design. Unpublished D.Phil thesis,
university of Oxford.
Cho, J.-R., & Mathews, R. C. (1996). Interactions
between mental models used in categorization and
experiential knowledge of specific cases. Quarterly
Journal of Experimental Psychology, 49A, 572-595.
Cleeremans, A. (1993). Mechanisms of implicit
learning: Connectionist models of sequence
processing. Cambridge, MA: MIT Press.
Cohen, J. (1988). Statistical power analysis for
the behavioural sciences (2nd ed.). Hillsdale, NJ:
Lawrence Erlbaum Associates.

Implicit Memory and System Control 54
Dienes, Z., & Altmann, G. T. M. (1997). Transfer of
implicit knowledge across domains: How implicit and
how abstract? In D. C. Berry (Ed.), How implicit is
implicit learning? (pp. 107-123). Oxford: Oxford
University Press.
Dienes, Z., Altmann, G., Kwan, L, Goode, A. (1995)
Unconscious knowledge of artificial grammars is
applied strategically. Journal of Experimental
Psychology: Learning, Memory, & Cognition,21, 1322-
1338.
Dienes, Z., & Berry, D. C. (1997). Implicit
learning: Below the subjective threshold.
Psychonomic Bulletin and Review,4, 3-23.
Dienes, Z., & Fahey, R. (1995). The role of
specific instances in controlling a dynamic system.
Journal of Experimental Psychology: Learning,
Memory, & Cognition,21, 848-862.
Dienes, Z., & Perner, J. (1996). Implicit knowledge
in people and connectionist networks. In G.
Underwood (Ed), Implicit cognition (pp 227-256).
Oxford: Oxford University Press.
Flexser, A. J. (1991). The implications of item
differences: Commentary on Hintman and Hartry.

Implicit Memory and System Control 55
Journal of Experimental Psychology: Learning,
Memory, and Cognition,17, 338-340.
Gardiner, J. M. (1991). Contingency relations in
successive tests: Accidents do not happen. Journal
of Experimental Psychology: Learning, Memory, and
Cognition,17, 334-337.
Gardiner, J. M., & Parkin, A. J. (1990). Attention
and recollective experience in recognition memory.
Memory and Cognition,18, 23-30.
Hayes, N.A., & Broadbent, D. E. (1988). Two modes of
learning for interactive tasks. Cognition,28, 249-
276.
Hayman, C., & Tulving, E. (1989). Contingent
dissociation between recognition and fragment
completion: The method of triangulation. Journal
of Experimental Psychology: Learning, Memory, and
Cognition,15, 228-240.
Hintzman, D. L. (1991). Contingency analyses,
hypotheses, and artifacts: Reply to Flexser and to
Gardiner. Journal of Experimental Psychology:
Learning, Memory, and Cognition,17, 341-345.

Implicit Memory and System Control 56
Hintzman, D. L., & Hartry, A. L. (1990). Item
effects in recognition and fragment completion:
Contingency relations vary for different subsets of
words. Journal of Experimental Psychology:
Learning, Memory, and Cognition,16, 955-969.
Lewicki, P. (1986). Nonconscious social information
processing. New York: Academic Press.
Logan, G. (1988). Towards an instance theory of
automatization. Psychological Review,95, 492-527.
Marescaux, P-J., Luc, F., & Karnas, G. (1989). Modes
d'apprentissage selectif et nonselectif et
connaissances acquies au control d'un processes:
Evaluation d'un modele simule. Cahiers de
Psychologie Cognitive,9, 239-264.
Mathews, R. C., Buss, R. R., Chinn, R., & Stanley,
W. B. (1988). The role of explicit and implicit
learning processes in concept discovery. Quarterly
Journal of Experimental Psychology,40, 135-165.
Nosofsky, R. M. (1988). Exemplar-based accounts of
relations between classification, recognition, and
typicality. Journal of Experimental Psychology:
Learning, Memory, and Cognition,14, 700-708.

Implicit Memory and System Control 57
Nosofsky, R. M. (1991). Tests of an examplar model
for relating perceptual classification and
recognition memory. Journal of Experimental
Psychology: Human Perception and Performance,17,3-
27.
Ostergaard, A. L. (1992). A method for judging
measures of stochastic dependence: Further
comments on the current controversy. Journal of
Experimental Psychology: Learning, Memory, and
Cognition,18, 413-420.
Ostergaard, A. L. (1994). Who is mistaken about
priming in “recognition/identification” experiments?
A reply to Tulving and Hayman. European Journal of
Cognitive Psychology,7, 1-11.
Perruchet, P. (1994). Learning from complex rule-
governed environments: On the proper function of
conscious and unconscious processes. In C. Umilta &
M. Moscovitch (Eds), Attention and Performance XV:
conscious and nonconscious information processing pp
(811-835). Cambridge, MA: MIT Press.

Implicit Memory and System Control 58
Poldrack, R. A. (1996). On testing for stoachastic
dissociations. Psychonomic Bulletin & Review,3,
434-448.
Reber, A. S. (1967). Implicit learning of
artificial grammars. Journal of Verbal Learning
and Verbal Behaviour,6, 855-863.
Reber, A.S. (1989). Implicit learning and tacit
knowledge. Journal of Experimental Psychology:
General, 118, 219-235.
Roediger, H. L., III (1990). Implicit memory:
Retention without remembering. American
Psychologist,45, 1043-1056.
Roediger, H. L., III, & Blaxton, T. A. (1987).
Effects of varying modality, surface features, and
retention interval on priming in word fragment
completion. Memory & Cognition,15, 379-388.
Roediger, H. L., III, & McDermott, K. B. (1993).
Implicit memory in normal human subjects. In F.
Boller J. Grafman (Eds.), Handbook of
Neuropsychology, Vol. 8. Elsevier Science
Publishers.

Implicit Memory and System Control 59
Sanderson, P. M. (1989). Verbalizable knowledge and
skilled task performance: Association,
dissociation, and mental models. Journal of
Experimental Psychology: Learning, Memory, and
Cognition,15, 729-747.
Schacter, D. L. (1987). Implicit memory: History
and current status. Journal of Experimental
Psychology: Learning, Memory, and Cognition,13,
501-518.
Schacter, D. L. (1990). Perceptual representation
systems and implicit memory: Towards a resolution
of the multiple memory systems debate. Annals of
the New York Academy of Sciences, 608, 543-571.
Schacter, D. L., & Graf, P. (1989). Modality
specificity of implicit memory for new
associations. Journal of Experimental Psychology:
Learning, Memory, & Cognition,15, 3-12.
Schwartz, S. H. (1966). Trial-by-trial analysis of
processes in simple and disjunctive concept
attainment tasks. Journal of Experimental
Psychology,72, 456-465.

Implicit Memory and System Control 60
Servan-Schreiber, E., & Anderson, J. R. (1990).
Learning artificial grammars with competitive
chunking. Journal of Experimental Psychology:
Learning, Memory, and Cognition,16, 592-608.
Shanks, D. R. (1997). Dissociating long-term memory
systems: Comment on Nyberg and Tulving (1996).
European Journal of Cognitive Psychology,9, 11-120.
Shanks, D. R., & Gluck, M. A. (1994). Tests of an
adaptive network model for the identification and
categorization of continuous dimension stimuli.
Connection Science,6,59-89.
Shanks, D. R., & St. John, M. F. (1994).
Characteristics of dissociable human learning
systems. Behavioural and Brain Sciences,17, 367-
448.
Squire, L. R., & Frombach, M. (1990). Cognitive
skill learning in amnesia. Psychobiology,18,
109-117.
Stanley, W.B., Mathews, R.C., Buss, R.R. & Kotler-
Cope, S. (1989). Insight without awareness: On the
interaction of verbalization, instruction, and
practice in a simulated process control task.

Implicit Memory and System Control 61
Quarterly Journal of Experimental Psychology, 41,
553-577.
Tulving, E. (1985). How many memory systems are
there? American Psychologist,40, 385-398.
Tulving, E., & Hayman, C. G. (1993). Stochastic
independence in the recognition/identification
paradigm. European Journal of Cognitive Psychology,
5, 353-373.
Tulving, E., & Schacter, D. L. (1990). Priming and
human memory systems. Science, 247, 301-306.
Tulving, E., Schacter, D. L., & Stark, H. A. (1982).
Priming effects in word fragment completion are
independent of recognition memory. Journal of
Experimental Psychology: Learning, Memory, and
Cognition, 8, 336-342.
Vokey, J. R., & Brooks, L. R. (1992). Salience of
item knowledge in learning artificial grammars.
Journal of Experimental Psychology: Learning,
Memory, & Cognition,18, 328-344.

Implicit Memory and System Control 62
Warrington, E. K., & Weiskrantz, L. (1978). Further
analysis of the prior learning effect in amnesic
patients. Neuropsychologia,16, 169-176.
Weldon, M. S., & Roediger, H. L. (1987). Altering
retrieval demands reverses the picture superiority
effect. Memory & Cognition,15, 269-280.
Whittlesea, B. W. A., & Dorken, M. D. (1993).
Incidentally, things in general are particularly
determined: An episodic-processing account of
implicit learning. Journal of Experimental
Psychology: General, 122, 227-248.
Willingham, D.B., Nissen, M.J. & Bullemer, P.
(1989). On the development of procedural knowledge.
Journal of Experimental Psychology: Learning,
Memory, and Cognition, 15, 1047-1060.


Implicit Memory and System Control 63
Footnotes
1Generalization between instances may increase
performance on new situations in the specific
situations task. However, performance on new
situations is used in the Ostergaard analysis as a
measure of baseline performance in the absence of
learning. Performance on new situations which are
more than one level of sugar production different
from any old correct situation was .10 (SD = .32),
no less than the performance on all new situations
(.07, SD = .27). That is, generalization did not
detectably increase performance on new situations,
and so the Ostergaard analysis would not be
affected.
2An analysis of the data from Experiment 1 of Dienes
and Fahey (1995) also revealed that the relation
between performance on the specific situations task
and recognition was nonsignificantly different from
independence and significantly lower than that
expected by Ostergaard's model. In the Dienes and
Fahey experiment, unlike the experiment reported in
the current paper, situations were presented in the
specific situations and recognition tasks as
combinations of sugar production and work force on
the previous trial. This creates a potential
problem with drawing strong conclusions from the

Implicit Memory and System Control 64
apparent stochastic independence in the Dienes and
Fahey experiment because they provided evidence that
subjects were responding to the sugar production
task on the basis of single features. That is,
subjects responded to a given level of sugar
production in a similar way regardless of the level
of work force with which it was combined, and they
also responded to a given level of work force in a
similar way regardless of the level of sugar
production. Thus, even if subjects did not
recognize a particular combination of sugar
production and work force as being old (and
incorrectly responded 'new') in the recognition
task, they may still have recognized the level of
sugar production by itself as being old and
responded appropriately in the specific situations
task. This would reduce the expected degree of
statistical dependence between the tasks below that
predicted by the Ostergard model.
3Further, Dienes and Fahey (1995) presented evidence
that subjects responded to the specific situations
task on the basis of the combination of Ellis'
behaviour and the subjects' behaviour on the last
trial. Because in this case, unike Experiment 1 of
Dienes and Fahey (see footnote 2), subjects were
responding to both the specific situations task and

Implicit Memory and System Control 65
the recognition task on the basis of the same
features, Ostergaard's model provides a plausible
estimate of the amount of dependence expected if
subjects' responses on the specific situations task
were based on explicit recollection.
4Brown et al. (1991) argued that modality shifts in
implicit memory tasks may be more likely to appear
when study modalities are manipulated within rather
than between subjects so as to focus subjects'
attention on perceptual features of the stimuli.
However, there have been a number of studies finding
substantial modality shifts on implicit memory tasks
using between subjects designs (Bassili et al.,
1989; Weldon & Roediger, 1987).
5Considering new test situations, the similarity for
false alarms (493, SD = 128) was higher than for
correct rejections (470, SD = 92), though not
significantly so, t(13) = 0.4 (95% CI on the
difference went from -151 to 105, indicating that
the data were not sensitive enough to distinguish
the interesting alternatives).
6Or that familiarity depended only on exact matches
and immediately adjacent situations. Similarity is

Implicit Memory and System Control 66
known to drop off at a faster than linear rate with
distance in other paradigms (Shanks & Gluck, 1994).
7A similar analyses was conducted on the data from
Experiment 2. Similarity was defined by a city block
metric (as the dimensions were presumably
perceptually separable; Shanks & Gluck, 1994):
Similarity = 23 - |difference in Ellis' behaviour|
- |difference in subject's behaviour|. Thus, the
minimum similarity is 1 and the maximum is 23. The
difference between hits (961, SD = 56) and misses
(885, SD = 76) was significant, t(19) = 3.97, p =
.001. For frequency, computed as the number of
times that exact situation occurred in training, the
difference between hits (2.6, SD = 0.9) and misses
(1.3, SD = 0.3), was also significant, t(19) = 5.55,
p< .001. When similarity difference was regressed
against frequency difference, the intercept (20) was
not significantly different from zero, t(18) = 0.76
(the intercept lay within the body of the data).
That is, on this model of similarity, the data
provided no evidence that subjects' recognition
decisions were determined by any situation in the
training phase other than the exact situation tested
for.


Implicit Memory and System Control 67
Table 1
Contingency table for situations to which subjects
gave the correct response in the training phase of
Experiment 1
Situations Task
Correct Incorrect
Recognition Task
Subjects response:
'Old' 29 53
'New' 12 16
Note. On the recognition task, 'old' is the correct
response, 'new' is the incorrect response. On the
situations task, 'correct' refers to the subject
giving the correct response when tested on this
task, and 'incorrect' refers to the subject giving
the incorrect response when tested on this task:
All the situations had been given the correct
response in the training phase.


Implicit Memory and System Control 68
Table 2
Specific situations task/proportion correct
Person: U S
Modality: Same Diff. Same Diff.
Study priming:
New .30 (.13) .33 (.18) .47 (.19) .54
(.15)
Old .53 (.25) .54 (.23) .70 (.20) .74
(.23)
Note. Standard deviations appear in parentheses.


Implicit Memory and System Control 69
Table 3
Contingency table for Person S
Situations Task
Correct Incorrect
Recognition Task
Subjects response
'Old' 105 24
'New' 39 10
Contingency table for Person U
Situations Task
Correct Incorrect
Recognition Task
Subjects response
'Old' 56 21
'New' 50 18
Note. On the recognition task, 'old' is the correct
response, 'new' is the incorrect response. On the
situations task, 'correct' refers to the subject
giving the correct response when tested on this
task, and 'incorrect' refers to the subject giving
the incorrect response when tested on this task:
All the situations had been given the correct
response in the training phase (i.e. a response that
led to target or just one level off).


Implicit Memory and System Control 70
Figure 1.
The displays seen by subjects during the learning
phases of Experiment 1 (top panel) and Experiment 2
(bottom panel).