Pre-exposure of repeated search configurations facilitates subsequent contextual cuing of visual search

Abstract

Contextual cuing is the enhancement of visual search when the configuration of distractors has been experienced previously. It has been suggested that contextual cuing relies on associative learning between the distractor locations and the target position. Four experiments examined the effect of pre-exposing configurations of consistent distractors on subsequent contextual cuing. The findings demonstrate a facilitation of subsequent cuing for pre-exposed configurations compared to novel configurations that have not been pre-exposed. This facilitation suggests that learning of repeated visual search patterns involves acquisition of not just distractor-target associations but also associations between distractors within the search context, an effect that is not captured by the Brady and Chun (2007) connectionist model of contextual cuing. We propose a new connectionist model of contextual cuing that learns associations between repeated distractor stimuli, enabling it to predict an effect of pre-exposure on contextual cuing. (PsycINFO Database Record (c) 2014 APA, all rights reserved).

Full text

Pre-Exposure of Repeated Search Configurations Facilitates Subsequent
Contextual Cuing of Visual Search
Tom Beesley
University of New South Wales Miguel A. Vadillo
University College London
Daniel Pearson
University of New South Wales David R. Shanks
University College London
Contextual cuing is the enhancement of visual search when the configuration of distractors has been
experienced previously. It has been suggested that contextual cuing relies on associative learning between the
distractor locations and the target position. Four experiments examined the effect of pre-exposing configu-
rations of consistent distractors on subsequent contextual cuing. The findings demonstrate a facilitation of
subsequent cuing for pre-exposed configurations compared to novel configurations that have not been
pre-exposed. This facilitation suggests that learning of repeated visual search patterns involves acquisition of
not just distractor–target associations but also associations between distractors within the search context, an
effect that is not captured by the Brady and Chun (2007) connectionist model of contextual cuing. We propose
a new connectionist model of contextual cuing that learns associations between repeated distractor stimuli,
enabling it to predict an effect of pre-exposure on contextual cuing.
Keywords: contextual cuing, visual search, pre-exposure, associative learning, auto-association
The visual world we face on a daily basis is not a random collection
of objects. Rather, our environment is highly structured, and the tasks
we conduct within it are undoubtedly facilitated by learning processes
that encode the multitude of contingencies available to us. Perhaps
one of the most commonly performed cognitive tasks is the process of
visual search: detecting a target object within an array of distracting
objects that are not the focus of our goals. For example, we may need
to identify a traffic signal among an array of information in a busy
street, and this search might be facilitated by learning that the location
of the traffic signal is to the right of a particular street sign and above
a bus stop shelter. Learning these contingencies allows for a rapid
direction of our attention toward the possible target location based on
our past experiences.
The effect of experience on repeated visual search is most
clearly demonstrated in laboratory tasks examining contextual
cuing (Chun & Jiang, 1998). In a typical experiment participants
are given instructions that the task measures visual search perfor-
mance, but surreptitiously a set of search configurations (the
organization of the distractor stimuli and the target position) are
repeated across the course of the experiment. Intermixed with
these trials, random search configurations (hereafter random con-
figurations) that do not repeat again during the experiment are
presented to participants. Reaction times (hereafter RTs) on re-
peated search trials are typically considerably faster than those on
random configurations, indicating that participants learn the con-
tingent relationships between distractor locations and target posi-
tion; these effects are robust and have now been shown in numer-
ous research laboratories across many different types of stimuli
and conditions (e.g., Beesley & Shanks, 2012;Jiang & Wagner,
2004;Jiménez & Vázquez, 2011;Kunar, Flusberg, Horowitz, &
Wolfe, 2007;Olson & Chun, 2002).
There is still a great deal of debate as to the mechanisms
responsible for even the most fundamental aspects of performance
in the contextual cuing task. Current discussion has focused on the
question of whether learning reflects a facilitation of attention to
the target location (Chun & Jiang, 1998) or a facilitation in target
detection (e.g., Kunar et al., 2007). What is clear, however, is that
participants are learning to associate the spatial arrangement of
distractor locations with the spatial location of the target, and while
the debate continues as to how this learning manifests itself in
search performance (enhanced localization or enhanced detection),
the present article examines the form that these associative repre-
sentations take in memory. Returning to our initial example, the
current article asks, what is learned about the position of the traffic
signal in relation to the street sign and the bus stop that leads to
facilitated cuing of the location of the traffic signal in the future?
This article was published Online First July 7, 2014.
Tom Beesley, School of Psychology, University of New South Wales;
Miguel A. Vadillo, Division of Psychology and Language Sciences, Uni-
versity College London; Daniel Pearson, School of Psychology, University
of New South Wales; David R. Shanks, Division of Psychology and
Language Sciences, University College London.
This work was supported by Grant ES/J007196/1 from the Economic
and Social Research Council, awarded to David R. Shanks and Tom
Beesley. We are grateful to Brooke Hahn for her assistance in data
collection.
Correspondence concerning this article should be addressed to Tom
Beesley, School of Psychology, University of New South Wales, Sydney,
New South Wales 2052, Australia. E-mail: t.beesley@unsw.edu.au
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
Journal of Experimental Psychology:
Learning, Memory, and Cognition © 2014 American Psychological Association
2015, Vol. 41, No. 2, 348–362 0278-7393/15/$12.00 http://dx.doi.org/10.1037/xlm0000033
348

One possibility is that, during contextual cuing, associations
form only between the individual distractor objects and the target
object of the search. The most formal description of this account
was provided by Brady and Chun (2007), who proposed an asso-
ciative model of contextual cuing that forms direct associations
from distractor locations to target locations. In this model, asso-
ciative learning takes place on a distractor-by-distractor basis, such
that co-occurring distractors acquire independent connections to
the target and can prime the target location independently of other
associations that have been formed. The model suggests that
contextual cuing is competitive, such that existing associations will
modulate the learning of novel but redundant associations (al-
though see Beesley & Shanks, 2012, for a demonstration that such
competition may not occur in contextual cuing).
There is a great deal of supporting evidence that the learning of
distractor–target associations is a key component of contextual
cuing. For example, the contextual cuing effect is reduced or even
abolished when the target position is relocated in a second phase of
the experiment (Makovski & Jiang, 2010;Manginelli & Pollmann,
2009). Potential target stimuli also appear to be afforded special
encoding in the memory representation of stored contexts. Conci,
Sun, and Müller (2011; see also Conci & Müller, 2012) used a task
in which contexts were trained with two target stimuli present on
a given trial, one consistent with the designated task response and
a second that was initially inconsistent. When the second target
became relevant, contextual cuing was still observed (albeit some-
what reduced). These data clearly support the learning of direct
associations between distractor stimuli and target stimuli in the
task.
The current article examines whether learning over and above
that of the individual distractor–target associations takes place
when learning about repeated search contexts. In our example, we
might ask whether searching for a shop sign is facilitated when
carried out in the same traffic scene that we are already familiar
with as a result of searching for the traffic signal. This seems
plausible: When we search for different objects within a scene it
would seem inefficient to encode the search context anew for
every combination of search context and target. Perhaps more
likely is that the process of visual search for one object may lead
(at least in part) to a representation forming for the search config-
uration as a whole. Importantly, such an effect is not anticipated by
any account that describes contextual cuing as resulting only from
the formation of associations between distractor and target stimuli.
For example, according to Brady and Chun’s (2007) model, asso-
ciations that form as a result of consistent searches in fixed
contexts (contextual cuing) would lead to an impairment in con-
textual cuing for subsequent searches at a different location. The
model would effectively need to extinguish the established
distractor–target associations and relearn new distractor–target as-
sociations for these new contingencies. Indeed, there are several
demonstrations of the severe inflexibility of the learning system to
such changes in distractor–target contingencies. Zellin, Conci, von
Mühlenen, and Müller (2013) extended the design of Manginelli
and Pollmann (2009) in which targets were remapped to new
positions, replicating the abolition of contextual cuing but also
demonstrating that a return of the target to its originally trained
position sparked an immediate recovery of contextual cuing. Fur-
thermore, the resilience of this proactive interference on the orig-
inal distractor–target associations was observed even over ex-
tended periods of training. Again, these results suggest that
contextual cuing is driven by the initial encoding of associations
between distractor locations and the initial target position.
In this article we explore whether contextual cuing may involve
the learning of distractor–distractor associations in addition to that
of distractor–target associations. Because contextual cuing exper-
iments have tended to train consistent distractor–target associa-
tions, in many cases it has not been possible to conclude whether
such learning occurs in contextual cuing. To our knowledge only
a few studies have directly sought evidence for distractor–
distractor (or configural) learning in contextual cuing. Of particu-
lar interest is a study by Jiang and Wagner (2004, Experiment 1;
see also Olson & Chun, 2002;Song & Jiang, 2005), who trained
participants in a first stage with repeated patterns of context in a
standard design, before, in a transfer test, combining subsets of
distractors from two contexts that had been paired with the same
target. If learning of distractor–distractor associations is crucial for
contextual cuing, we would expect such recombination of contexts
to disrupt the contextual cuing effect. However, Jiang and Wagner
observed contextual cuing effects of a similar magnitude for the
trained contexts and the recombined contexts. We note, however,
that recombined displays were constructed from two separate
half-patterns in the transfer phase, and so a significant proportion
of the configural information was retained in the transfer displays.
In a second experiment, Jiang and Wagner examined the contex-
tual cuing elicited by transformations on the originally trained
contexts that preserved the original configurations of distractors
(distractor positions were rescaled and uniformly displaced in one
direction). In these transformed displays, the locations of individ-
ual distractors were all changed; hence, if contextual cuing de-
pends solely on relationships between individual distractors and
the target, cuing should be disrupted. Jiang and Wagner found that
contextual cuing was unimpaired by such transformations, provid-
ing some support for the role of configural or distractor–distractor
representations in contextual cuing (although see Brady & Chun,
2007, for a discussion of how their elemental theory may account
for this performance).
In the present experiments we provide a systematic exami-
nation of whether contextual cuing results in the learning of
distractor configurations independently of distractor–target as-
sociations. In all experiments we train a set of configurations
with a pseudo-randomly placed target on each trial. We antic-
ipated that this pre-exposure period would allow for the learning of
consistent configurations of distractors in the absence of a consistent
distractor–target association. In a subsequent phase we train a new
configuration–target association (the configurations become predic-
tive of one particular target location) for these previously exposed
configurations and compare the contextual cuing that occurs to that
for novel repeated configurations. It is important to note that we
do not expect that the repetition of nonpredictive configurations
will produce a contextual cuing effect in and of itself. In fact,
Chun and Jiang (1998, Experiment 3) showed that such repe-
tition of contexts does not produce a contextual cuing effect.
The suggestion is rather that this repetition may result in the
encoding of a representation of the search context, leading to
enhanced learning of subsequent consistent target placements at
a later stage.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
349
PRE-EXPOSURE FACILITATES CONTEXTUAL CUING

Experiment 1
Experiment 1 aimed to examine the effect that pre-exposure to
nonpredictive repeated configurations has on subsequent contex-
tual cuing. This was examined by pairing the pre-exposed config-
urations with consistent target positions in the subsequent training
phase and comparing search efficiency on these configurations to
that on novel repeated configurations. Given the pre-exposure of
nonpredictive configurations in the first phase, the experiment also
provides a conceptual replication of Chun and Jiang’s (1998,
Experiment 3) demonstration of equivalent search times to re-
peated nonpredictive and random configurations.
Method
Participants. Twenty individuals from the University of New
South Wales participated in exchange for course credit. The ex-
periment lasted approximately 25 minutes.
Apparatus. The experiment was conducted on PCs with 19-
in. thin-film-transistor (TFT) monitors set at a resolution of
1,024 ⫻768. Stimulus presentation and response recording was
handled by software programmed using MatLab, Cogent 2000, and
Cogent Graphics (www.vislab.ucl.ac.uk/cogent.php). The experi-
ment was programmed by T. Beesley (whereas subsequent exper-
iments were programmed by M. A. Vadillo). Stimuli were drawn
within the experiment software. Distractor stimuli were letter Ls,
and the target stimulus was a letter T. The letter stimuli were 13
mm square. For the distractors, the vertical line of the letter Lwas
offset slightly (less than 1mm) from the end of the horizontal line
in order to increase the similarity between distractor and target
shapes and therefore increase the difficulty of the visual search
task.
Stimuli were arranged in an invisible square grid of 144 evenly
spaced cells (12 ⫻12), which was positioned centrally on the
screen and was 240 mm square. The fixation cross (displayed
centrally before each trial) was 11 mm square. Responses to the
target stimulus were made with keys Zand Mon a standard PC
keyboard. The background color of the screen was gray. Stimuli
were colored blue, red, green, or yellow. Distractor stimuli were
oriented by rotating the letter Lby 0°, 90°, 180°, or 270°. Target
stimuli were oriented by rotating the letter Tby 90° or 270°. The
color and orientation of the distractors were randomly assigned for
each pattern, with the constraint that there could be no more than
eight distractors of one color and that there were at least two
distractors of each color. Repeated elements of patterns maintained
the same color and orientation for distractor stimuli across repeti-
tions. Target color was also randomly determined and maintained
across presentations of the same pattern, but target orientation was
determined randomly within each block of trials.
Design. Patterns consisted of 17 stimuli (16 distractors plus 1
target) with four distractors placed in each quadrant of the pattern.
Eight target locations were used in total, two in each quadrant of
the screen. Target positions were chosen at random from one of
five locations within each quadrant that were approximately equi-
distant from the center of the screen. Distractors could not appear
in these target locations.
During the first phase of the experiment, participants were
trained with four repeated configurations of distractors. In all
experiments reported here, four configurations were trained for
each trial type. This method has two advantages. First, it has been
suggested that contextual cuing effects are driven by learning one
or two configurations (Smyth & Shanks, 2008) and that training
fewer repeated configurations leads to more robust cuing effects
(Bennett, Barnes, Howard, & Howard, 2009). Thus, training only
a small number of configurations reduces unwanted noise in any
measure of contextual cuing. Second, four configurations allows
for the placement of the target in four quadrants of the screen,
ensuring spatial frequency effects for targets are minimized (see
Jiang, Swallow, & Rosenbaum, 2013). Each of these configura-
tions was presented four times in each block of this phase, and,
importantly, within this block each configuration was presented
with each of four target locations assigned to the repeated config-
urations. Thus, although the configuration of distractors repeated
across trials, it could not be used as a valid cue for target location.
Random configurations were trained alongside repeated configu-
rations during this phase. These two sets of four configurations
(repeated and random) were each assigned four target locations,
one from each quadrant of the display.
During the second phase of the experiment, participants were
presented with the pre-exposed configurations of distractors as
well as a novel set of repeated configurations. Importantly, in this
phase all configurations were predictive of a specific target loca-
tion. This was achieved by pairing one configuration within each
set with one target location in each quadrant of the display. In
order to minimize the effect of any residual associations between
repeated target locations and any of the four target locations they
were presented with in the first phase, we switched the target sets
for pre-exposed repeated configurations. That is, the target loca-
tions used for repeated configurations in the first stage were used
for the novel configurations in the second phase, and the target
locations used for random configurations in the first stage were
used for pre-exposed repeated configurations in the second phase.
Procedure. At the start of the experiment participants re-
ceived instructions detailing the nature of the visual search task.
Example displays were presented, and participants were shown the
correct response for each orientation of the target.
The first phase of the experiment consisted of 5 blocks of 32
trials. Each block contained four presentations of each repeated
configuration with each of the four targets, as well as 16 random
configurations. The second phase of the experiment consisted of
10 blocks of 8 trials, with each of the four pre-exposed repeated
configurations presented once and each of the four novel repeated
configurations presented once.
Within each block, trials were presented in a random order with
the constraint that consecutive trials across adjoining blocks could
not present the same repeated pattern. Any given target position
could not occur on consecutive trials. Target orientation (left or
right) was determined randomly but with an equal number of
presentations of each orientation within a sub-block of 8 trials.
Each trial commenced with a fixation cross presented in the
center of the screen for 1,000 ms, which was then replaced imme-
diately by the pattern of stimuli. RTs were recorded from the onset
of the pattern. Following a valid response (Zor M) the pattern was
removed from the screen. The response–stimulus interval (hereaf-
ter RSI) was 1,000 ms. If participants made an incorrect response
to the target orientation, ERROR! appeared in the center of the
screen for 2,000 ms, prior to the RSI. A rest break of 20 seconds
was given every 108 trials. Trials started automatically after these
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
350 BEESLEY, VADILLO, PEARSON, AND SHANKS

breaks. The transition between the pre-exposure and training
phases was not signaled to participants.
Results
All participants performed the task with a high degree of accu-
racy: Mean accuracy was 99.4%. Two participants produced mean
RTs two standard deviations (hereafter SDs) greater than the mean
of the sample, and they were removed as outliers. Trials on which
an incorrect response was made, trials immediately after a rest
break, and trials longer than 10 seconds were all discounted from
the analysis of RTs. RTs in this and all subsequent experiments
were natural log transformed, and those RTs longer than 2 SDs
above or below the mean RT were discounted. All analyses were
conducted on natural log transformed data, although nontrans-
formed data are presented in the figures.
The data from the pre-exposure phase are presented in Figure
1A. As expected, search times to (nonpredictive) repeated config-
urations and random configurations were equivalent across the
course of this phase. This was confirmed by a two-way repeated
measures analysis of variance (ANOVA) with factors of configu-
ration (repeated vs. random) and block, which revealed a main
effect of block, F(4, 68) ⫽19.92, p⬍.001. There was no main
effect of configuration, F(1, 17) ⫽1.23, p⫽.28, and no interac-
tion between configuration and block, F(4, 68) ⫽1.04, p⫽.39.
Of particular interest are the data from the training phase, which
are presented in Figure 1B. Search times decreased rapidly across
the block, with a clear tendency for faster search times on the
pre-exposed repeated configurations. This was confirmed by a
two-way ANOVA with factors of configuration (pre-exposed re-
peated vs. repeated) and block, which revealed a main effect of
configuration, F(1, 17) ⫽6.53, p⬍.05, and block, F(9, 153) ⫽
2.79, p⬍.01, but no interaction between these factors (F⬍1).
Discussion
In Experiment 1, participants experienced a pre-exposure phase
in which a set of four configurations not predictive of target
location was presented. The results from this phase represent a
replication of Chun and Jiang’s (1998) Experiment 3: Contextual
cuing did not occur for repeated configurations trained with targets
placed in inconsistent locations. However, Experiment 1 sought to
examine whether this pre-exposure of repeated configurations con-
ferred any benefit to contextual cuing with these patterns in a
subsequent stage. Indeed, the pre-exposure phase had exactly this
consequence: When the pre-exposed configurations were subse-
quently trained with consistent target locations, a contextual cuing
effect stronger than that for a comparable novel set of repeated
configurations was observed.
The results suggest that although the configurations were incon-
sistently paired with target placements in the first phase, represen-
tations of the pre-exposed configurations were formed that later
facilitated target search. One possibility is that the representation
that is stored during pre-exposure acts as a more salient cue for the
contextual cuing that develops in the second phase, leading to
stronger associations with the target location. Our procedure will
presumably have resulted in an absence of any reliable distractor–
target pairings during the pre-exposure phase and will also have
ensured that these associations would be unable to contribute to
A. RTs during the pre-exposure phase of Experiment 1.
B. RTs during the training phase of Experiment 1.
1800
2000
2200
2400
2600
2800
12345
Search efficiency (RT)
Blocks of 32 trials
Repeated
Random
1600
1700
1800
1900
2000
2100
2200
2300
12345678910
Search efficiency (RT)
Blocks of 8 trials
Repeated pre-exposed
Repeated
Figure 1. (A) Reaction times (RTs) during the pre-exposure phase of
Experiment 1. (B) RTs during the training phase of Experiment 1.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
351
PRE-EXPOSURE FACILITATES CONTEXTUAL CUING

cuing in the second phase. As such, the results of Experiment 1 are
difficult to account for by a purely elemental model of contextual
cuing (such as that provided by Brady & Chun, 2007). The results
suggest instead that contextual cuing is driven, at least in part, by
the learning of distractor–distractor associations or the formation
of configural representations during pre-exposure.
Experiment 2
In Experiment 1 repeated configurations were pre-exposed with
inconsistent target locations in a standard visual search task. Our
account of this effect centers on the encoding of associations
within the distractor configuration itself. The placement of the
target position across all four quadrants of the display during
pre-exposure meant it was rather unlikely that the pre-exposure
advantage stemmed from learning distractor–target associations
during the pre-exposure phase. Although the interaction effect was
not significant, the facilitation in cuing for pre-exposed configu-
rations did not appear to transfer directly from pre-exposure,
because search times on pre-exposed and novel repeated configu-
rations were equivalent in the first block of the training phase (see
Figure 1B). Following this initial block, a clear difference was
observed in the response times to pre-exposed and novel repeated
configurations.
If the pre-exposure effect does not result from learning associ-
ations between distractors and the target position, it follows that
the pre-exposure effect should be observed even when the config-
uration of distractors is pre-exposed without a target present on
each trial. Thus, in Experiment 2 participants experienced config-
urations on both target-present and target-absent trials during the
pre-exposure phase before receiving a standard contextual cuing
task consisting of target-present trials only. Importantly, we pre-
exposed two sets of configurations, one set presented only on
target-present trials during the pre-exposure phase and one set
presented only on target-absent trials in the pre-exposure phase. As
in Experiment 1 these configurations were intermixed with random
configurations for both target-present and target-absent trials. In a
subsequent stage, both sets of configurations were trained with
consistent target locations to examine contextual cuing. We again
included a novel set of repeated configurations against which we
could assess the rate of learning. Random configurations were
included in this phase, allowing us to assess the extent of learning
for each configuration type.
Method
Participants. Fifty-two participants from University College
London participated in the experiment. Participants were paid £4
with a penalty imposed of 2 pence for every incorrect response or
for every response slower than 3 seconds. The experiment lasted
for approximately 35 minutes.
Apparatus. The experiment was conducted on PCs with 17-
in. TFT monitors set at a resolution of 1,280 ⫻1,024. The
experiment was programmed by M. A. Vadillo and was similar to
Experiment 1 except for the following minor differences. The letter
stimuli were 14 mm square. Unlike in Experiment 1, the vertical line
of the letter Lwas not offset from the end of the horizontal line. This
made the visual search task marginally easier than Experiment 1
by making distractors less similar to the target. The stimuli formed
a 270-mm invisible square grid. The fixation cross was 8 mm
square. For no-target trials participants issued a “target-absent”
response with the spacebar.
Design. Patterns consisted of 17 stimuli (16 distractors plus 1
target) with four distractors placed in each quadrant of the pattern.
During the first phase of the experiment, participants were trained
with eight repeated configurations of distractors, four in the target-
present set and four in the target-absent set. For the target-present
set, each configuration was presented an equal number of times
with a target in each of the four quadrants of the screen. During
this stage, the target could appear in any randomly chosen location
of the screen, with the only constraint that, for any given repeated
pattern, the target could not appear again in one quadrant until it
had already appeared in all the other quadrants. Additionally, the
target could appear in any color. Therefore, as in Experiment 1,
although the configuration of distractors repeated across trials, it
could not be used as a valid cue for target location. For the
target-absent set, the four configurations were presented an equal
number of times but without target stimuli. Random configurations
of both target-present and target-absent forms were intermixed
with repeated configurations during this phase.
During the second phase of the experiment, participants were
presented with both sets of pre-exposed configurations of distrac-
tors as well as a new set of repeated configurations. All configu-
rations were of the target-present type and were predictive of a
specific target location. During this phase 16 target locations were
used, four in each quadrant of the screen. Target positions were
chosen at random from a circle of locations approximately equi-
distant from the center of the screen. Distractors could not appear
in these target locations. Four different sets of four targets were
selected and used for the four different sets of configurations:
repeated pre-exposed (target-present); repeated pre-exposed (tar-
get-absent); repeated (novel); and random. Each set contained one
target from each quadrant of the display.
Procedure. The procedure was similar to that of Experiment
1 with the exception of the following changes. Participants were
instructed to use their index fingers for keys Zand Mand to signal
a target-absent response by pressing the spacebar with the thumb
of their dominant hand.
The first phase of the experiment consisted of 24 blocks of 16
trials. Each block contained one presentation of each of the four
repeated configurations from both the target-present and target
absent sets, as well as four random target-present configurations
and four random target-absent configurations. The second phase of
the experiment consisted of 12 blocks of 16 trials, with configu-
rations from both sets of pre-exposed repeated configurations
presented once and each of the four novel repeated configurations
presented once, along with four random configurations. All con-
figurations were of the target-present type during the second phase.
Unlike in Experiment 1, in the present experiment the target
orientation (left or right) was not constrained to an equal number
of left/right presentations within sub-blocks; the target orientation
was decided randomly on a trial-by-trial basis. Participants
searched for the target and responded with keys Zor Mfor
detection of the target (according to its orientation) or with the
spacebar if the target was absent. A rest break of 20 seconds was
given every 100 trials.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
352 BEESLEY, VADILLO, PEARSON, AND SHANKS

Results
Two participants produced error rates on target-present trials
that were greater than two standard deviations above the mean
error rate of the sample, and three additional participants produced
mean RTs that were more than two standard deviations above the
mean of the sample. These five participants were not included in
the final analysis. The criteria used in Experiment 1 were applied
for the exclusion of long RTs and erroneous trials, with outliers
identified separately for target-present and target-absent trials
(given the clear difference in RTs on these trials; see below). Data
were combined into two-block averages for purposes of data
presentation and analysis.
Figure 2A shows the data for the pre-exposure phase. Search
times were clearly much faster on target-present trials than on
target-absent trials. This is likely to be due to a target-absent
trial requiring an exhaustive search, whereas a target-present
search can be terminated, on average, after inspection of half
the items. A three-way repeated measures ANOVA with factors
of configuration (repeated vs. random), target (present vs. ab-
sent), and block revealed a main effect of configuration, a main
effect of target, and a main effect of block (all Fs⬎28.3, all
ps⬍.001). Of interest, there was an interaction between
configuration and target, F(1, 46) ⫽6.41, p⬍.05, indicating
that the difference in response times between repeated and
random configurations was greater in the case of the target-
absent trials. There were significant interaction effects between
configuration and block, and target and block (Fs⬎2.09, ps⬍
.05); the three-way interaction effect was not significant, F(11,
506) ⫽1.22, p⫽.27. In order to understand the significant
interaction between target and configuration, we conducted
separate ttests on target-present and target-absent data to
explore the difference between responses to pre-exposed and
random configurations. They revealed that RTs to pre-exposed
configurations were faster than those to random configurations
on the target-absent trials, t(46) ⫽7.11, p⬍.001, but not on the
target-present trials, t(46) ⫽1.50, p⫽.14.
Figure 2B shows the data for the training phase. As in
Experiment 1, it appears that contextual cuing was stronger for
pre-exposed contexts but that this effect is limited to those
contexts trained with targets in the first phase. These data were
subjected to a repeated measures ANOVA with factors config-
uration (pre-exposed target-present repeated, pre-exposed
target-absent repeated, novel repeated, and random) and block.
It revealed a main effect of configuration, F(3, 138) ⫽6.40,
p⬍.001; a main effect of block, F(5, 230) ⫽11.25, p⬍.001;
but no interaction between these factors, F(15, 690) ⫽1.42,
p⫽.13. Follow-up planned comparisons revealed that respond-
ing on the novel repeated configurations was faster than re-
sponding on random configurations, F(1, 46) ⫽24.73, p⬍
.001, demonstrating a contextual cuing effect in the training
phase. Responses to configurations pre-exposed on target-
present trials were faster than to novel repeated configurations,
F(1, 46) ⫽5.19, p⬍.05, replicating the results of Experiment
1, but there was no difference between responses to configura-
tions pre-exposed on target-absent trials and novel repeated
configurations (F⬍1). Finally, responses to configurations
pre-exposed on target-present trials were faster than responses
to configurations pre-exposed on target-absent trials, F(1,
44) ⫽4.88, p⬍.05.
Although the interaction between configuration and block
was not significant, it is of particular interest to examine
whether the contextual cuing effect for pre-exposed target pres-
ent configurations developed during the training phase or was
present immediately on transfer to the training phase. An anal-
ysis on the first block of trials found that over the first 16 trials
within that block there was no contextual cuing effect (RTs to
repeated configurations ⫽1,022 ms; RTs to random configu-
rations ⫽1,033 ms; t⬍1), but a contextual cuing effect was
present over the second set of 16 trials in Block 1 (RTs to
repeated configurations ⫽953; RTs to random configura-
tions ⫽1,054 ms). t(46) ⫽2.84, p⬍.01. Thus, although there
was a rapid development of contextual cuing for the pre-
exposed target-present configurations at the start of the training
phase, it was not present at the immediate outset of this training
phase.
A
B
800
1000
1200
1400
1600
1800
2000
2200
2400
123456789101112
Search efficiency (RT)
Blocks of 32 trials
Repeated target-present Repeated target-absent
Random target-present Random target-absent
900
950
1000
1050
1100
123456
Search efficiency (RT)
Blocks of 32 trials
Repeated pre-exposed (target-present)
Repeated pre-exposed (target-absent)
Repeated
Random
Figure 2. (A) Reaction times (RTs) during the pre-exposure phase of
Experiment 2. (B) RTs during the training phase of Experiment 2.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
353
PRE-EXPOSURE FACILITATES CONTEXTUAL CUING

Discussion
In Experiment 2, participants were trained initially with repeated
but nonpredictive configurations across two different types of
search trial: those with targets present and those without targets.
There was a clear contextual cuing effect on the target-absent
trials, but as in Experiment 1 this was not observed in the target-
present trials. We note that this particular finding for the target-
absent trials is inconsistent with published data from Kunar and
Wolfe (2011); we discuss this discrepancy, along with the similar
finding from Experiment 4, in the General Discussion.
Following pre-exposure, participants received a standard con-
textual cuing task (target-present trials throughout) in which pre-
exposed configurations were trained as predictive of specific target
locations. Contextual cuing for these configurations was compared
to that for a set of novel repeated configurations that had not been
pre-exposed. The results showed clear contextual cuing effects,
with faster responses made to repeated configurations than random
configurations. Furthermore, responses to those configurations that
had been pre-exposed on target-present trials were faster than
responses to novel repeated configurations. This result replicates
the finding in Experiment 1 that pre-exposing a configuration
facilitates subsequent contextual cuing for that configuration.
However, a pre-exposure effect was not observed for those con-
figurations pre-exposed on target-absent trials.
It is surprising that target-absent configurations did not confer a
pre-exposure benefit on subsequent cuing, given that these con-
texts were experienced for a substantially greater duration on each
trial during the pre-exposure phase. Indeed, the exhaustive search
that participants must perform on these configurations is likely to
increase exposure to a greater spatial proportion of each search
configuration across the many presentations during this phase. It
might therefore be the case that the lack of a pre-exposure benefit
from target-absent trials is due not to a deficit in the encoding of
that configuration per se but instead to the development of a strong
association between the configuration and the terminating (space-
bar) response as a result of the target-absent procedure. Indeed, the
significant cuing effect for pre-exposed target-absent configura-
tions in the first phase suggests that participants had learned to
consistently press the spacebar in response to this set of repeated
configurations. It is possible that this learned response may have
interfered with responding to the new target locations in the
subsequent training phase. This account therefore suggests that the
facilitation resulting from pre-exposure may have been in direct
competition with the initial terminating response and as a result
may have masked any effect of pre-exposure on learning in the
second phase. Experiments 3 and 4 explore possible accounts of
why a pre-exposure effect was not observed for target-absent trials
in Experiment 2.
Experiment 3
Experiment 3 used a time-out procedure for target-absent pre-
exposure: Participants had to wait for the trial to terminate after a
fixed duration on target-absent trials. During the pre-exposure
phase, participants were presented with repeated configurations on
target-absent trials and experienced random configurations on both
target-present trials and target-absent trials. The use of target-
present trials during pre-exposure is important for maintaining
vigilance in the task to ensure encoding of the repeated configu-
rations. During the subsequent training phase, participants were
trained in a standard contextual cuing task with only target-present
trials. They experienced pre-exposed and novel repeated configu-
rations intermixed with random configurations.
Method
Participants. Forty participants from University College Lon-
don participated in exchange for £3–£4 depending on their perfor-
mance, with the same criteria used as in Experiment 2. The
experiment lasted for approximately 30 minutes.
Apparatus. The materials and apparatus were identical to
those in Experiment 2.
Design. Experiment 3 examined only pre-exposure on target-
absent trials. Configurations were created in the same manner as
for Experiment 2. Four repeated configurations were pre-exposed
in the first phase. A set of four targets (one in each quadrant of the
display) were used for these configurations in the second phase of
the experiment. Two more sets of four targets (one in each quad-
rant) were used for novel repeated configurations and random
configurations.
Procedure. The procedure was similar to that of Experiment
2 with the following changes. The pre-exposure phase consisted of
24 blocks of 20 trials. Within each block, each of the four target-
absent repeated configurations was presented once. These were
intermixed with 4 random target-absent configurations and 12
random target-present configurations. During this phase, trials
timed out after 2,000 ms for target-absent trials. The subsequent
training phase consisted of 12 blocks of 12 trials. Each of the four
configurations from each set of repeated configurations (pre-
exposed and novel) was presented once and was intermixed with
four random configurations. The training phase consisted of target-
present trials only.
Results
Three participants produced error rates on target-present trials
that were greater than two SDs above the mean error rate of the
sample. One additional participant produced a mean RT that was
more than two SDs below the mean of the sample. These four
participants were not included in the final analysis.
The criteria used in Experiment 1 were applied for data exclu-
sions. Data were combined into two-block averages for purposes
of data presentation and analysis. Figure 3A shows the accuracy
across the pre-exposure phase. Accuracy was very high for target-
absent trials, but there were no differences in the accuracy for
repeated and random configurations, t(35) ⫽1.00, p⫽.32. Ac-
curacy on target-present trials was initially worse than that on
target-absent trials but reached comparable levels by the end of the
phase. A one-way ANOVA found that RTs on target-present trials
decreased steadily across the course of the pre-exposure phase,
F(11, 385) ⫽28.81, p⬍.001.
Figure 3B shows RT data during the training phase. It appears
that a contextual cuing effect emerged during training, with a
numerically larger effect occurring for the novel repeated config-
urations. These data were subjected to a repeated measures
ANOVA with factors of configuration (pre-exposed repeated,
novel repeated, and random) and block. It revealed a marginally
significant effect of configuration, F(2, 70) ⫽2.69, p⫽.08;
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
354 BEESLEY, VADILLO, PEARSON, AND SHANKS

a significant effect of block, F(5, 175) ⫽18.95, p⬍.001; and a
significant interaction between configuration and block, F(10,
350) ⫽1.96, p⬍.05. Separate two-way ANOVAs provided
pairwise comparisons for the levels of the configuration variable.
Contrasting the two repeated configurations revealed no effect of
configuration, F(1, 35) ⫽1.30, p⫽.26. There was a main effect
of block, F(5, 175) ⫽16.43, p⬍.001, but no interaction between
these factors, F(5, 175) ⫽1.82, p⫽.11, suggesting that contextual
cuing did not occur more readily for pre-exposed than novel
repeated configurations; indeed, the means are in the direction of
less cuing for pre-exposed patterns. Contrasting the pre-exposed
repeated configurations with random configurations revealed no
main effect of configuration, F(1, 35) ⫽1.41, p⫽.24, and no
significant interaction with block (F⬍1). The main effect of block
was significant, F(5, 175) ⫽12.09, p⬍.001. Contrasting the
novel repeated configurations with random configurations re-
vealed a significant effect of configuration, F(1, 35) ⫽5.30, p⬍
.05; a significant effect of block, F(5, 175) ⫽15.38, p⬍.001; and
a significant interaction between these factors, F(5, 175) ⫽3.32,
p⬍.01.
Discussion
In Experiment 3 participants were pre-exposed to repeated con-
figurations only on target-absent trials. On these trials the correct
response was to allow the trial to time out after 2 seconds. In a
subsequent training phase, pre-exposed configurations were
trained as predictive of target locations. However, learning about
these pre-exposed configurations occurred at a similar rate as the
learning that was observed for novel repeated configurations.
Thus, the pattern of data was similar to that seen for the pre-
exposed target-absent configurations in Experiment 2.
Although the aim of the pre-exposure procedure in Experiment
3 was to limit the response competition occurring on target-absent
trials, it is possible that the time-out procedure led to the devel-
opment of strong inhibitory response associations to pre-exposed
configurations. That is, the correct response on target-absent trials
was to refrain from responding, and it is possible that this inhibi-
tion may have interfered with a subsequently learned response in
the training phase.
Experiment 4
In Experiment 4, participants experienced a pre-exposure phase
in which they counted the number of stimuli within the configu-
ration. Configurations contained either eight or nine distractors,
and participants were asked to respond with two keys on the
keyboard (odd/even). Importantly, pre-exposed configurations
were presented an equal number of times with eight and nine
distractors. This procedure therefore allows for pre-exposure of the
repeated configurations without the development of a consistent
interfering response.
Method
Participants. Thirty-four participants from the University of
New South Wales participated in exchange for payment of
AUS$12. The experiment lasted for approximately 30 minutes.
Apparatus. The materials and apparatus were identical to
those in Experiments 2 and 3. Participants used keys Zand Mto
respond even or odd to trials in the pre-exposure phase, and the
same keys were used to respond to target orientation in the training
phase.
Design. Configurations were created in the same manner as
for Experiments 2 and 3, except for the fact that search displays
now consisted of 8 distractors. Four repeated configurations were
pre-exposed in the first phase. Each configuration was presented
24 times during pre-exposure and was paired with the odd and
even responses the same number of times (12 each). This was
A
B
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1.00
123456789101112
Accuracy
Blocks of 40 trials
Repeated target-absent
Random target-absent
Random target-present
900
950
1000
1050
1100
1150
1200
123456
Search efficiency (RT)
Blocks of 24 trials
Repeated pre-exposed
Repeated
Random
Figure 3. (A) Accuracy of response during the pre-exposure phase of
Experiment 3. (B) Reaction times (RTs) during the training phase of
Experiment 3.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
355
PRE-EXPOSURE FACILITATES CONTEXTUAL CUING

achieved by adding an additional random element to the configu-
ration on half of the trials.
Procedure. The procedure was similar to that used in Exper-
iments 2 and 3 with the following changes. The pre-exposure
phase consisted of 24 blocks of 8 trials. Within each block, each of
the four repeated configurations was presented once, and these
presentations were intermixed with four random configurations.
During this phase, participants had to determine whether there was
an odd or even number of stimuli in the display. The subsequent
training phase consisted of 12 blocks of 12 trials. Each of the four
configurations from each set of repeated configurations (pre-
exposed and novel) was presented once and intermixed with four
random configurations.
Results
Two participants produced error rates on trials in the training
phase that were greater than two SDs above the mean error rate of
the sample, and one additional participant produced a mean RT
that was more than two SDs above the mean of the sample. These
three participants were not included in the final analysis. The
criteria used in Experiments 1–3 were applied for data exclusions.
Data were combined into two-block averages for purposes of
presentation and analysis.
Figure 4A shows RTs for the counting task across the 12 blocks
of the pre-exposure phase. A clear difference was observed, with
faster decision times to repeated configurations over random con-
figurations. This was confirmed by a repeated measures ANOVA
with factors configuration (repeated vs. random) and block, which
revealed a main effect of configuration, F(1, 30) ⫽38.06, p⬍
.001, and a main effect of block, F(11, 330) ⫽8.95, p⬍.001. The
interaction between these factors was not significant, F(11, 330) ⫽
1.49, p⫽.13.
Figure 4B shows RT data during the training phase. A contex-
tual cuing effect emerged during training, with responses to re-
peated configurations (pre-exposed and novel) faster than those to
random configurations. These data were subjected to a repeated
measures ANOVA with factors configuration (pre-exposed re-
peated, novel repeated, and random) and block, which revealed a
significant effect of configuration, F(2, 60) ⫽10.07, p⬍.001;
a significant effect of block, F(5, 150) ⫽9.55, p⬍.001; and a
significant interaction between configuration and block, F(10,
300) ⫽3.82, p⬍.001. Pairwise comparisons of the different
configurations revealed that responses to novel repeated configu-
rations were faster than those to random configurations, F(1, 30) ⫽
11.25, p⬍.01, a standard contextual cuing effect, and that
responses to the pre-exposed configurations were also faster than
those to random configurations, F(1, 30) ⫽19.43, p⬍.001. There
was no difference between the responses to pre-exposed and novel
repeated configurations (F⬍1), although this comparison did
reveal a marginally significant interaction with the factor of block,
F(1, 30) ⫽3.13, p⫽.09. Exploratory analysis revealed that there
was a significant facilitation in responding to pre-exposed config-
urations (over novel repeated configurations) on the very first
block of the training phase, t(30) ⫽3.02, p⫽.005, but not in any
other (when corrected for multiple comparisons). As in Experi-
ments 1 and 2, the contextual cuing effect for pre-exposed repeated
configurations appears to emerge rapidly in the training phase: An
analysis on the first block of trials found that over the first 12 trials
within that block there was no contextual cuing effect (RTs to
repeated configurations ⫽863; RTs to random configurations ⫽
871), t⬍1, but a contextual cuing effect was present over the
second set of 12 trials in Block 1 (RTs to repeated configura-
A
B
1500
1600
1700
1800
1900
2000
2100
2200
2300
2400
2500
123456789101112
Search efficiency (RT)
Blocks of 16 trials
Repeated Random
750
800
850
900
950
123456
Search efficiency (RT)
Blocks of 24 trials
Repeated pre-exposed
Repeated
Random
Figure 4. (A) Reaction times (RTs) during the pre-exposure phase of
Experiment 4. (B) RTs during the training phase of Experiment 4.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
356 BEESLEY, VADILLO, PEARSON, AND SHANKS

tions ⫽821; RTs to random configurations ⫽889), t(30) ⫽2.41,
p⬍.05.
Discussion
In Experiment 4, participants were pre-exposed to repeated
configurations in a novel counting procedure. Participants were
clearly able to perform this counting task more readily for
repeated configurations than for novel configurations. This is
similar to the contextual cuing effect observed on target-absent
trials in the pre-exposure phase of Experiment 2; possible
accounts of this effect are provided in the General Discussion.
During the subsequent training phase, participants experienced a
standard contextual cuing procedure in which they were trained
with pre-exposed configurations and novel repeated configurations
intermixed with random configurations. These data revealed a
transient but robust effect of pre-exposure at the outset of this
phase, representing an effect of pre-exposure on subsequent con-
textual cuing in configurations pre-exposed on target-absent trials.
Given the absence of a target search task in this pre-exposure
phase, these data provide additional support to the conclusion that
the pre-exposure effect reflects learning at the level of the distrac-
tor configuration.
General Discussion
Four experiments examined the extent to which pre-exposure of
a configuration of stimuli during visual search led to the subse-
quent facilitation of contextual cuing within that configuration. In
Experiment 1, participants were trained with repeated configura-
tions of distractors that did not reliably signal the location of the
target. During this pre-exposure phase no contextual cuing effect
was observed, replicating similar findings from Chun and Jiang
(1998, Experiment 3). However, it appears that the pre-exposure
phase did result in these configurations being encoded into mem-
ory, because subsequent learning about the pre-exposed configu-
rations was greater than that for novel repeated configurations. In
Experiment 2, nonpredictive configurations were pre-exposed ei-
ther on trials that featured a target stimulus or on trials where the
target was absent. The findings revealed a pre-exposure effect
replicating that of Experiment 1 but only for those configurations
pre-exposed on target-present trials. Experiment 3 examined
whether the removal of a learned target-absent response from the
task (by using a time-out procedure, as opposed to a self-
terminating procedure) would reveal a pre-exposure effect. How-
ever, there was again no evidence for a pre-exposure effect on
target-absent trials. In Experiment 4, participants counted the
number of distractors on the screen during the pre-exposure phase,
thus removing any form of consistent motor response from the
pre-exposure phase. A transient but significant facilitation in cuing
was observed for the pre-exposed configurations during the sub-
sequent training phase.
It should be noted that the major results (i.e., the pre-exposure
effects observed in Experiments 1 and 2 with target-present con-
figurations) were sustained across the entire training phase,
whereas that observed in Experiment 4 with target-absent config-
urations was reflected only by a difference in performance in the
first block of the training phase. Although further experimentation
will be needed to explore these different patterns of data, we
suggest at least two plausible hypotheses. First, we note that the
counting task in Experiment 4 led to substantially longer RTs in
the pre-exposure phase (as a result of the more exhaustive search
necessary for this task) and also resulted in a contextual cuing
effect during this phase (which was not present with target-present
configurations in Experiments 1 and 2). This aspect of the data
suggests that participants may have become aware of the repetition
of the configurations in this phase, perhaps at a much earlier stage
than in the standard visual search task. This difference in explicit
encoding and the possibility of participants approaching the task in
a more strategic manner may well have resulted in rapid acquisi-
tion of the novel repeated configurations in the second phase,
limiting pre-exposure to a transient effect in Experiment 4.
Pre-Exposure Results in Acquired Representations
for Search Contexts
The results of the current experiments suggest that when par-
ticipants are exposed to repeated configurations of stimuli during
visual search, learning occurs not only between individual distrac-
tor elements and the target position but also within that repeated
set of distractor elements. We have tended to refer to these patterns
of distractors as “configurations,” and one possible means by
which this learning could take place is in the formation of a distinct
configural representational unit for each search context encoun-
tered in the task. Configural models of associative learning have
played a dominant role as explanatory tools across a range of
human and animal learning studies. According to Pearce’s (1987,
1994,2002) model of configural learning, each presentation of a
unique pattern of stimuli will result in a configural representation
forming for that pattern of input to the model. This unique repre-
sentational unit then acquires associative strength directly with
significant events (in the current case, the spatial position of the
target within the display).
The results of the current experiments can certainly be accom-
modated within the framework of a configural model, with some
assumptions. According to this account, over a series of presenta-
tions, the pre-exposure of a search context will result in the
formation of a configural representation for that spatial arrange-
ment of distractors. One could assume that this process is gradual
and involves the stochastic sampling of the complete veridical
search context. Over time, therefore, a strong representation of the
complete search context will form in memory and will act as a
closer match to any future samples of the search context on
subsequent occasions. Thus, in the training phase, when contexts
are paired consistently with target locations, each new sampling of
the search context will result in a strong activation of the pre-
exposed configural representation, leading to a consistent strength-
ening of a single configuration–target association. In contrast, for
the novel repeated configurations, these configural representations
must be established during the training phase; hence, weaker
contextual cuing would be expected during the initial stages of the
training phase.
This adapted configural account is certainly not the only candi-
date explanation for the current results. One could imagine, in-
stead, that the representational form of the search context is en-
tirely elemental but involves a large network of associations
forming between the distractor elements. We envisage this as a
process of auto-association (e.g., McClelland & Rumelhart, 1985)
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
357
PRE-EXPOSURE FACILITATES CONTEXTUAL CUING

operating across repeated search contexts, such that during the
pre-exposure phase associations are formed not just between indi-
vidual distractor stimuli and the target location but also between all
the distractor stimuli within the configuration. Thus, over the
course of pre-exposure, this model would learn the co-occurrences
of distractors within repeated configurations. According to auto-
associative models, when a pattern of input is presented to the
model, the activation on the input layer is a product of this external
input and the internally generated input from the model. Thus,
pretraining allows the model to produce a more stable pattern of
activation on the input layer (relative to a novel pattern), which
subsequently permits stronger associations to form between dis-
tractor locations and the target position.
We present an implementation of the auto-associative model
and contrast its pattern of simulated results with those provided by
the Brady and Chun (2007) model. The aim of these simulations is
to account for three key aspects of the data observed with regard
to the robust pre-exposure effects obtained with target-present
configurations in Experiments 1 and 2: (a) an absence of contex-
tual cuing during the pre-exposure phase itself; (b) no difference in
responding on repeated and random configurations at the outset of
the training phase; (c) the development and sustained benefit to
contextual cuing for pre-exposed over novel configurations.
Simulations With an Auto-Associative Model of
Contextual Cuing
The auto-associative model takes as its starting point the Brady
and Chun (2007) model of contextual cuing.
1
It uses the same
structure to encode direct associations between the distractor po-
sitions and the target position, but it includes an additional layer of
associations that acquires associations between the individual dis-
tractor positions. Figure 5 provides an illustration of the weight
structures within the model. Formally, the model sums the inter-
nally generated input for each unit, in
i
, and uses the sigmoid
activation function to establish a final pattern of activation across
the input units, a
i
,
ini⫽
共
兺
jExtj·AAWij
兲
, (1)
ai⫽1⁄(1⫹e⫺ini), (2)
where Ext
j
is the external input to the model on unit jand AAW
ij
is the auto-associative weight between input unit jand input unit i.
The model determines the activation on the output units by a
similar process,
ino⫽
共
兺
iai·Woi
兲
, (3)
ao⫽1⁄(1⫹e⫺ino), (4)
where W
oi
is the weight between input unit iand output unit o.
From the activation across the network we take two measures of
performance: the activation rank of the target unit (how active is
the target relative to other units) and the mean square error of the
activation across the input units. The latter measure reflects how
accurate the model is in re-creating the pattern of external input
from its self-generated input via the auto-associative network and
therefore reflects the development of the associations between
distractors.
Following the model’s prediction on each trial, both sets of
weights are adjusted. Weights within the auto-associative network
are adjusted by
(⌬AAWij)t⫽␸·(Exti⫺ai)·Extj⫹␪·(⌬AAWij)t⫺1, (5)
where ␸is a parameter determining the rate of associative learning
and ␪is a parameter determining the contribution of momentum to
weight changes. Weights for the input–output associations are
adjusted by
(⌬Woi)t⫽␸·(T⫺ao)·ai⫹␪·(⌬Woi)t⫺1, (6)
where Tis a teaching signal, taking a value of 1 if the target is
presented on that output unit and 0 otherwise.
The model was trained with patterns created in an identical
manner as those in Experiment 1 and with the same procedure
(number of blocks, trial randomization, etc.). Results were calcu-
lated as the average performance from 100 simulated subjects,
with a learning rate (␸) of .1 and a momentum term (␪)of.1.
Figure 6A shows the target rank data across the 5 blocks of
pre-exposure. The model predicts a decline in the number of
targets checked as a result of its learning of the most frequent
target positions in the display. However, it does not predict a
contextual cuing effect and therefore accurately simulates the
pattern of data observed during the pre-exposure phase of Exper-
iment 1. Figure 6B shows the target rank data across the training
phase of the experiment. Although the model does not anticipate a
difference in performance on pre-exposed and novel repeated
1
Note that Brady and Chun’s (2007) model uses a fixed weight structure
to modify the associative learning that takes place between target and
distractor stimuli as a function of the distance between these stimuli. For
simplicity, we do not employ these fixed weights in simulations with the
auto-associative model, although a comprehensive model of contextual
cuing would combine these. For our simulations of the Brady and Chun
model we used a scaling factor of 10—the largest parameter suggested by
Brady and Chun—which allows the model to learn more about the sur-
rounding distractor configuration.
i1 i2 i3 iN
o1 o2 o3 oN
L
L
L
T
Scene memor
y
Object-Target
associations
L
L
L
T
Figure 5. An auto-associative model of contextual cuing. The model
features two sets of weights. Associations forming between input units (i1
...iN) reflect a representation of the co-occurrences of distractor stimuli
within repeated patterns of context (scene memory). Associations forming
between input units and output units (o1...oN) reflect representations
between distractor locations and the spatial location of paired target stim-
uli.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
358 BEESLEY, VADILLO, PEARSON, AND SHANKS

configurations initially, it clearly learns more rapidly about the
pre-exposed configurations across this training phase. Thus, the
model accurately models the pre-exposure effect observed in our
experiments. Figures 6C and 6D show the mean-square error on
the input units during the pre-exposure and training phases, re-
spectively. These data show that the model rapidly acquires asso-
ciations on the auto-associative layer that allow it to produce stable
input representations for repeated configurations during pre-
exposure. From the outset of the training phase, the model can use
this stability to facilitate the development of associations between
the input and output layers of the network, leading to enhanced
target detection for pre-exposed configurations.
Simulations With the Brady and Chun (2007) Model
Simulations were conducted with an implementation of the Brady
and Chun (2007) model and the procedure from Experiment 1. First,
we trained the model using the parameters specified in Brady and
Chun (2007): learning rate (␸) of .001; momentum term (␪) of .95;
bottom-up activation of .1. The bottom-up activation parameter pro-
vides a boost in the activation for output units representing stimuli that
are present in a given pattern, effectively suppressing locations in
which stimuli are not present. Figures 7A and 7B present the simu-
lation results across the pre-exposure and training phases, respec-
tively. Figure 7A shows a small but highly consistent contextual cuing
effect for pre-exposed configurations during the pre-exposure phase.
In addition, Figure 7B shows a substantial impairment in the devel-
opment of contextual cuing in the pre-exposed configurations relative
to that for the novel repeated configurations. The pattern of results
from this simulation does not provide a match to the ordinal pattern of
data observed in Experiment 1.
Further exploration of the parameter space revealed that increasing
the bottom-up parameter to a very high value (.99) reversed the
A B
C D
0
5
10
15
20
25
12345
Search efficiency (target rank)
Blocks of 32 trials
Repeated Random
0
1
2
3
4
5
6
7
8
12345678910
Search efficiency (target rank)
Blocks of 8 trials
Repeated pre-exposed
Repeated
0
0.1
0.2
0.3
0.4
0.5
12345
MSE on input units
Blocks of 32 trials
Repeated Random
0
0.1
0.2
0.3
0.4
0.5
12345678910
MSE on input units
Blocks of 8 trials
Repeated pre-exposed
Repeated
Figure 6. Simulations of the pre-exposure effect (Experiment 1) with the auto-associative network. Panels A
and B show the ranked position of the activation of the output unit for the target position (among all output units)
during the pre-exposure and training phases, respectively. Panels C and D show the mean square error (MSE)
of the activation on the input units during the pre-exposure and training phases, respectively.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
359
PRE-EXPOSURE FACILITATES CONTEXTUAL CUING

pattern of results in the training phase. These simulation results are
shown in Figures 7C and 7D (note that the scale is reduced in Figures
7C and 7D compared to Figures 7A and 7B to highlight key differ-
ences). Figure 7C shows the data from the pre-exposure phase. Again,
a contextual cuing effect is shown for pre-exposed configurations, a
pattern that was not observed in Experiments 1 or 2 for target-present
trials (nor in Experiment 3 of Chun & Jiang, 1998). In this simulation,
pre-exposure appears to confer a short-lived facilitation on contextual
cuing in the training phase, at least over the first 2–3 blocks. There are
two patterns within the data, however, that suggest the effect is not
driven by a facilitation in learning about pre-exposed configurations
per se. First, the difference between performance on pre-exposed and
novel repeated configurations appears largest at the start of the train-
ing phase; in Experiment 1 and in the simulation results with the
auto-associative model, however, the pattern reflected a growth in
contextual cuing across the training phase, at least until asymptotic
performance was reached. In addition, it appears that the rank position
of the target for pre-exposed configurations does not change between
the pre-exposure and training phases. Thus, the high bottom-up com-
ponent leads the model to simply identify the set number of target
locations, which appears to restrict the development of contextual
cuing.
Summary of the Simulation Results
In summary, a model that includes an auto-associative layer—
which allows it to learn distractor–distractor associations within
the pre-exposed configurations—provides a more accurate fit to
the data from Experiment 1 (and the hence the data from the
target-present condition in Experiment 2) than does a model that
learns distractor–target associations only. The auto-associative
model achieves this facilitation in contextual cuing for pre-
exposed configurations by learning to produce more stable patterns
of activation across the input units, which accurately reflect the
A B
C D
0
5
10
15
20
12345
Search efficiency (target rank)
Blocks of 32 trials
Repeated Random
0
5
10
15
20
12345678910
Search efficiency (target rank)
Blocks of 8 trials
Repeated pre-exposed
Repeated
0
1
2
3
4
5
6
7
12345
Search efficiency (target rank)
Blocks of 32 trials
Repeated Random
0
1
2
3
4
5
6
7
12345678910
Search efficiency (target rank)
Blocks of 8 trials
Repeated pre-exposed
Repeated
Figure 7. Simulations of the pre-exposure effect (Experiment 1) with the Brady and Chun (2007) network for
two different sets of parameters. Data show the ranked position of the activation of the output unit for the target
position (among all output units) during the pre-exposure (Panel A: Simulation 1; Panel C: Simulation 2) and
training phases (Panel B: Simulation 1; Panel D: Simulation 2).
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
360 BEESLEY, VADILLO, PEARSON, AND SHANKS

external input. This increased accuracy leads to the formation of
stronger distractor–target associations once these patterns become
predictive of specific target locations in the training phase.
Although this new connectionist model highlights the importance
of distractor–distractor learning in the contextual cuing task, it should
be noted that such learning is clearly insufficient to drive contextual
cuing alone. Indeed, the model extends rather than abandons the
component processes assumed by Brady and Chun (2007). The pres-
ence of distractor–target associations is crucial for the contextual
cuing effect in the model, and it is clear that distractor–target learning
plays an important role in the contextual cuing effect (e.g., Conci &
Müller, 2012;Makovski & Jiang, 2010;Manginelli & Pollmann,
2009;Zellin, Conci, von Mühlenen, & Müller, 2011). Furthermore, it
should be noted that these associations form only part of the array of
information acquired during repeated visual search. When the oppor-
tunity arises, people, for example, exploit high-frequency target po-
sitions at the expense of low-frequency positions (e.g., Jiang et al.,
2013), anticipate an upcoming required response (e.g., Jiménez &
Vázquez, 2011), and learn relationships between object identity and
target location (e.g., Endo & Takeda, 2004). Future developments of
the model will seek to accommodate such findings and could begin to
translate these findings to the more complex process of real-world
scene learning, which is thought to include the encoding of semantic
and global scene information (e.g., Brockmole, Castelhano, & Hen-
derson, 2006;Brockmole & Henderson, 2006a,2006b;Brooks, Ras-
mussen, & Hollingworth, 2010).
We have provided simulations of the data from the target-
present condition but not the target-absent condition. Although
there was evidence for a pre-exposure effect in Experiment 4 for a
target-absent condition, this experiment used a unique counting
procedure to reduce the impact of response competition from
target-absent responses on subsequent training (see above). We
attempted to simulate the data only from conditions in which the
standard contextual cuing procedure involving visual search for
target stimuli was used throughout. It is important to note that
the auto-associative model is not restricted to simulating only the
pre-exposure effect in the target-present conditions, as it is the
learning of co-occurring distractors in the model (the low MSE for
pre-exposed patterns) that provides the simulated result; the pres-
ence of a target element is irrelevant to the development of these
associations. However, any effective simulation of Experiment 4
(with the auto-associator or the Brady and Chun model) would
need to make substantial changes to the way in which outcomes
are encoded (e.g., to allow for even and odd responses to be made,
as opposed to a prediction of the target position). Similarly, a
simulation of the data from the target-absent conditions from
Experiments 2 and 3 could be achieved by allowing the model to
learn a no-target response in the pre-exposure phase. Given these
additional complications and the clear evidence for a robust effect
in the target-present conditions, it is perhaps prudent to collect
more data on the pre-exposure effect in the target-absent condition
before a comprehensive simulation that includes these conditions
is conducted.
Contextual Cuing on Target-Absent Trials
In Experiments 2 and 4, we found that contextual cuing was
observed during the pre-exposure phase for the configurations
pre-exposed on target-absent trials (Experiment 3 did not permit
assessment of this); contextual cuing did not occur for configura-
tions pre-exposed on target-present trials in Experiments 1 and 2.
We have suggested that contextual cuing arose on target-absent
trials in Experiment 2 as a result of participants learning a
no-target response (the spacebar) to these repeated configurations.
This termination of the search process can be viewed as a condi-
tioned response elicited by the repeated configuration, much in the
same way that the localization of attention to the target may be
elicited by the repeated configuration as an explanation of the
standard contextual cuing effect. In Experiment 4, however, no
consistent response was issued by participants to repeated config-
urations in the pre-exposure phase. Instead, participants were
asked to count the number of distractors in the configuration and
report whether there was an even or odd number. Given that the
configuration contained an even number of distractors on half of
the trials and an odd number on the remaining trials, it seems
unlikely that the contextual cuing effect occurred as a result of the
learning of a consistent conditioned motor response. What seems
more likely is that pre-exposure allowed participants to rapidly
identify the configuration of distractors (or parts of the configura-
tion), perhaps avoiding the need to attend to each distractor via a
serial search process. This is a somewhat speculative account of
the effect in the pre-exposure phase of Experiment 4, and further
experimental work will be needed to examine this account in more
detail. However, it is worth noting that this rapid detection of the
pre-exposed configuration is in keeping with the auto-associative
model of contextual cuing. That is, by learning the distractor–
distractor associations within a repeated configuration, the model
is able to retrieve from auto-associative memory the whole con-
figuration on the basis of a partial sampling of the pattern. Thus,
the model would anticipate enhanced pattern completion, which
may assist in the counting task used in the pre-exposure phase of
Experiment 4.
Finally, we note that although we have provided an account of
the contextual cuing effect observed during pre-exposure in Ex-
periments 2 and 4, this pattern of data is at odds with recent studies
conducted by Kunar and Wolfe (2011). Of their five reported
experiments, their Experiment 1 is the closest match to Experiment
2 of the current article. Twelve participants received a mixture of
target-present and target-absent trials. Participants had to simply
indicate whether a target was present or absent. Twelve repeated
patterns were used (6 paired with a target in a particular location;
6 presented consistently with no target), and targets could appear
in one of 12 locations. They also presented their repeated contexts
a similar number of times (30) as we did in the pre-exposure
phase of Experiment 2 (24). Although the procedure of this
experiment is remarkably similar to our pre-exposure procedure
in Experiment 2, the results are not: Kunar and Wolfe observed
a contextual cuing effect on target-present but not on target-
absent trials. This latter finding is at odds with Experiment 2, in
which we observed a robust contextual cuing effect during the
pre-exposure phase with target-absent trials. It is possible that
the source of this discrepancy is the combined training of
target-predictive configurations: Target-present repeated con-
figurations were nonpredictive of target location in our Exper-
iment 2 but were predictive of it in Kunar and Wolfe’s exper-
iment. If we imagine there is a limit to the number of context–
target associations that can be acquired during contextual cuing,
then perhaps in Kunar and Wolfe’s experiment the simultaneous
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
361
PRE-EXPOSURE FACILITATES CONTEXTUAL CUING

training of context–target pairings in one set restricted the
development of the context–no target associations in the other.
In contrast, the nonpredictive contexts in the current Experi-
ment 2 would not have exhausted such a limit. Of course, this
account is purely speculative; understanding why we obtained a
cuing effect with target-absent trials but Kunar and Wolfe
obtained a null result in their experiment will certainly be a
theoretically interesting avenue for future experimental work.
Conclusion
In conclusion, we have demonstrated a novel pre-exposure
effect in contextual cuing. The implication of this effect is that
contextual cuing does not merely reflect associative learning be-
tween individual distractor locations and the target position.
Rather, it suggests that the process of repeated visual search leads
to a plethora of associative connections between elements of the
search context. We have suggested one mechanism by which these
associations can be acquired that we hope will facilitate future
examinations of the representational content acquired during con-
textual cuing of visual search.
References
Beesley, T., & Shanks, D. R. (2012). Investigating cue competition in
contextual cuing of visual search. Journal of Experimental Psychology:
Learning, Memory, and Cognition, 38, 709–725. doi:10.1037/a0024885
Bennett, I. J., Barnes, K. A., Howard, J. H., & Howard, D. V. (2009). An
abbreviated implicit spatial context learning task that yields greater
learning. Behavior Research Methods, 41, 391–395. doi:10.3758/BRM
.41.2.391
Brady, T. F., & Chun, M. M. (2007). Spatial constraints on learning in
visual search: Modeling contextual cuing. Journal of Experimental Psy-
chology: Human Perception and Performance, 33, 798–815. doi:
10.1037/0096-1523.33.4.798
Brockmole, J. R., Castelhano, M. S., & Henderson, J. M. (2006). Contex-
tual cueing in naturalistic scenes: Global and local contexts. Journal of
Experimental Psychology: Learning, Memory, and Cognition, 32, 699–
706. doi:10.1037/0278-7393.32.4.699
Brockmole, J. R., & Henderson, J. M. (2006a). Recognition and attention
guidance during contextual cueing in real-world scenes: Evidence from
eye movements. Quarterly Journal of Experimental Psychology, 59,
1177–1187. doi:10.1080/17470210600665996
Brockmole, J. R., & Henderson, J. M. (2006b). Using real-world scenes as
contextual cues for search. Visual Cognition, 13, 99–108. doi:10.1080/
13506280500165188
Brooks, D. I., Rasmussen, I. P., & Hollingworth, A. (2010). The nesting of
search contexts within natural scenes: Evidence from contextual cuing.
Journal of Experimental Psychology: Human Perception and Perfor-
mance, 36, 1406–1418. doi:10.1037/a0019257
Chun, M. M., & Jiang, Y. (1998). Contextual cueing: Implicit learning and
memory of visual context guides spatial attention. Cognitive Psychology,
36, 28–71. doi:10.1006/cogp.1998.0681
Conci, M., & Müller, H. J. (2012). Contextual learning of multiple target
locations in visual search. Visual Cognition, 20, 746–770. doi:10.1080/
13506285.2012.694376
Conci, M., Sun, L., & Müller, H. J. (2011). Contextual remapping in visual
search after predictable target location changes. Psychological Research,
75, 279–289. doi:10.1007/s00426-010-0306-3
Endo, N., & Takeda, Y. (2004). Selective learning of spatial configuration
and object identity in visual search. Perception & Psychophysics, 66,
293–302. doi:10.3758/BF03194880
Jiang, Y. V., Swallow, K. M., & Rosenbaum, G. M. (2013). Guidance of
spatial attention by incidental learning and endogenous cuing. Journal of
Experimental Psychology: Human Perception and Performance, 39,
285–297. doi:10.1037/a0028022
Jiang, Y., & Wagner, L. C. (2004). What is learned in spatial contextual
cuing—configuration or individual locations? Perception & Psycho-
physics, 66, 454–463. doi:10.3758/BF03194893
Jiménez, L., & Vázquez, G. A. (2011). Implicit sequence learning and
contextual cueing do not compete for central cognitive resources. Jour-
nal of Experimental Psychology: Human Perception and Performance,
37, 222–235. doi:10.1037/a0020378
Kunar, M. A., Flusberg, S. J., Horowitz, T. S., & Wolfe, J. M. (2007). Does
contextual cuing guide the deployment of attention? Journal of Exper-
imental Psychology: Human Perception and Performance, 33, 816–
828. doi:10.1037/0096-1523.33.4.816
Kunar, M. A., & Wolfe, J. M. (2011). Target absent trials in configural
contextual cuing. Attention, Perception, & Psychophysics, 73, 2077–
2091. doi:10.3758/s13414-011-0164-0
Makovski, T., & Jiang, Y. V. (2010). Contextual cost: When a visual-
search target is not where it should be. Quarterly Journal of Experimen-
tal Psychology, 63, 216–225. doi:10.1080/17470210903281590
Manginelli, A. A., & Pollmann, S. (2009). Misleading contextual cues:
How do they affect visual search? Psychological Research, 73, 212–221.
doi:10.1007/s00426-008-0211-1
McClelland, J. L., & Rumelhart, D. E. (1985). Distributed memory and the
representation of general and specific information. Journal of Experi-
mental Psychology: General, 114, 159–188. doi:10.1037/0096-3445
.114.2.159
Olson, I. R., & Chun, M. M. (2002). Perceptual constraints on implicit
learning of spatial context. Visual Cognition, 9, 273–302. doi:10.1080/
13506280042000162
Pearce, J. M. (1987). A model for stimulus generalization in Pavlovian
conditioning. Psychological Review, 94, 61–73. doi:10.1037/0033-295X
.94.1.61
Pearce, J. M. (1994). Similarity and discrimination: A selective review and
a connectionist model. Psychological Review, 101, 587–607. doi:
10.1037/0033-295X.101.4.587
Pearce, J. M. (2002). Evaluation and development of a connectionist theory
of configural learning. Animal Learning & Behavior, 30, 73–95. doi:
10.3758/BF03192911
Smyth, A. C., & Shanks, D. R. (2008). Awareness in contextual cuing with
extended and concurrent explicit tests. Memory & Cognition, 36, 403–
415. doi:10.3758/MC.36.2.403
Song, J.-H., & Jiang, Y. V. (2005). Connecting the past with the present:
How do humans match an incoming visual display with visual memory?
Journal of Vision, 5, 322–330. doi:10.1167/5.4.4
Zellin, M., Conci, M., von Mühlenen, A., & Müller, H. J. (2011). Two (or
three) is one too many: Testing the flexibility of contextual cueing with
multiple target locations. Attention, Perception, & Psychophysics, 73,
2065–2076. doi:10.3758/s13414-011-0175-x
Zellin, M., Conci, M., von Mühlenen, A., & Müller, H. J. (2013). Here
today, gone tomorrow—Adaptation to change in memory-guided visual
search. PLoS ONE, 8(3), e59466. doi:10.1371/journal.pone.0059466
Received October 22, 2013
Revision received April 29, 2014
Accepted May 2, 2014 䡲
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
362 BEESLEY, VADILLO, PEARSON, AND SHANKS