Bottom-upguidanceinvisualsearchforconjunctions..pdf

Bottom-Up Guidance in Visual Search for Conjunctions

Michael J. Proulx
Johns Hopkins University

Understanding the relative role of top-down and bottom-up guidance is crucial for models of visual search.
Previous studies have addressed the role of top-down and bottom-up processes in search for a conjunction of
features but with inconsistent results. Here, the author used an attentional capture method to address the role
of top-down and bottom-up processes in conjunction search. The role of bottom-up processing was assayed
by inclusion of an irrelevant-size singleton in a search for a conjunction of color and orientation. One object
was uniquely larger on each trial, with chance probability of coinciding with the target; thus, the irrelevant
feature of size was not predictive of the target’s location. Participants searched more efficiently for the target
when it was also the size singleton, and they searched less efficiently for the target when a nontarget was the
size singleton. Although a conjunction target cannot be detected on the basis of bottom-up processing alone,
participants used search strategies that relied significantly on bottom-up guidance in finding the target,
resulting in interference from the irrelevant-size singleton.

Keywords: attention, visual search, bottom-up, top-down, conjunction

Computational models of eye movements give bottom-up process-
ing, the relative salience of objects in the stimulus, a prominent role
in the guidance of attention (Koch & Ullman, 1985). Many models of
visual search, however, suggest that when the target is defined as a
conjunction of features, search is based primarily on top-down pro-
cessing (e.g., Guided Search [GS]; Wolfe, 1994). The logic of the GS
model is that if the target is defined by one color and one orientation
(e.g., red and vertical) and the distractors are defined by a combination
of one of these target features and another feature (e.g., red and
horizontal or green and vertical), then the output of the bottom-up
feature maps cannot be relied on for guiding attention to the target. A
top-down search strategy is required presumably because each loca-
tion contains at least one of the target’s features and because no
location is featurally unique.

In the present study, bottom-up and top-down processes are defined
in terms of the two different types of input to the activation map in the
GS architecture (Wolfe, 1994). An important additional concept is the
idea of a participant’s search strategy. A participant could strategi-
cally rely on bottom-up processing, top-down processing, or a com-
bination of the two to perform a search task. Thus, a participant could

search for a unique feature singleton (such as a vertical line among
horizontal lines) by strategically relying on bottom-up processing to
search for the unique object in the display or by strategically relying
on top-down processing to search for the target-defining feature of
“vertical” in the display (cf. singleton-detection mode versus feature-
search mode, respectively; Bacon & Egeth, 1994). Note that the use
of bottom-up processing versus top-down processing can be contin-
gent on the strategy used by the participant, consistent with the
contingent-capture hypothesis (Folk, Remington, & Johnston, 1992).
Of course, the use of bottom-up processing might be involuntary (cf.
Theeuwes, 2004); however, the main concern here is not whether
bottom-up processing is automatic or contingent but, rather, whether
bottom-up processing plays a key role in conjunction search.

Egeth, Virzi, & Garbart (1984) were the first to provide evi-
dence for the top-down guidance of attention in conjunction
search. A conjunction search task generally has two distractor
types, with each type sharing one feature with the target. Egeth et
al. kept the number of one distractor type constant and varied only
the number of the other distractor items. They found that partici-
pants could restrict attention to one target feature and search for
the target among those items that shared that feature. Zohary and
Hochstein (1989) also reported evidence in favor of the top-down
guidance of attention by modifying Egeth et al.’s procedure. They
kept the total number of distractors constant and varied which
distractor dominated the display. Response times were fastest
when one distractor dominated and were slowest when equal
numbers of each distractor were present. They concluded that
participants selected the smaller subset in a top-down fashion and
searched for the target among that grouping.

Bacon and Egeth (1997) noted, however, that a strategy of relying
on the bottom-up guidance of attention could also explain Zohary and
Hochstein’s (1989) results. The smaller subset would be more salient
because each item would more likely be surrounded by the other
distractor type that outnumbers it. Therefore, Bacon and Egeth sought
support for a top-down account by instructing participants to search
one subset while also varying which subset dominated the display.

Michael J. Proulx, Department of Psychological and Brain Sciences,
Johns Hopkins University.

This material is based on work supported by a National Science Foun-
dation Graduate Research Fellowship and an American Psychological
Association Dissertation Award, both awarded to Michael J. Proulx. These
experiments were part of a doctoral dissertation submitted to the Johns
Hopkins University under the advisement of Howard Egeth.

I thank Howard Egeth, Steve Yantis, Amy Shelton, Ed Connor, Steve
Hsiao, Hartmut Leuthold, and Yuhong Jiang for incisive comments and
criticisms. Thanks also to Petar Dimitrov for technical assistance and Chris
Min for help with data collection.

Correspondence should be addressed to Michael J. Proulx, who is now
at the Institute of Experimental II, Bldg 23.03, 00.48, Univer-
siätsstrasse 1, Heinrich Heine University Düsseldorf, Düsseldorf D-40225,
Germany. E-mail: [email protected]

Journal of Experimental : Copyright 2007 by the American Psychological Association
Human Perception and Performance
2007, Vol. 33, No. 1, 48 –56

0096-1523/07/$12.00 DOI: 10.1037/0096-1523.33.1.48

48

Although Bacon and Egeth emphasized the result that participants
could indeed restrict attention to the instructed subset, participants
also abandoned the instructions when the other subset was the smaller
one. Sobel and Cave (2002) noted that this action could be due to a
reliance on either top-down or bottom-up processes, and they pro-
vided evidence from further experiments that was consistent with the
bottom-up account.

Sobel and Cave (2002) concluded that because no smaller subset
exists in standard conjunction search tasks, the “bottom-up system
. . . is of little use in standard conjunction searches” (p. 1067).
However, a participant can implement conjunction search in mul-
tiple ways, and the strategy of the participant must be known for
clarification of the roles of the top-down and bottom-up guidance
of attention for a conjunction target. It cannot be assumed that
participants are necessarily following instructions and searching
for a template of the target’s defining features, as assumed by
recent models such as GS (Wolfe, 1994).

In this study, I adapted an attentional capture paradigm to
examine the reliance on bottom-up and top-down strategies in a
conjunction search with the standard, equal number of each dis-
tractor type. Researchers in two studies used a variation of the
attentional capture method, although with conflicting results. In
one study, Lamy and Tsal (1999) used an additional singleton
paradigm (cf. Theeuwes, 1991) to assess whether bottom-up pro-
cesses are used in conjunction search. None of the additional
singletons they used (defined by color, shape, or both) disrupted
search on the target-present trials (although an effect was seen on
the target-absent trials). Lamy and Tsal concluded that a salient
singleton did not disrupt conjunction search.

In the other study, Friedman-Hill and Wolfe (1995; follow-up to
Experiment 6) examined participants’ ability to ignore an irrele-
vant distractor. They introduced an additional singleton to their
task of having participants restrict search to a color subset and then
search for an odd orientation within that subset. One of the
distractor items (never the target) was oddly textured on each trial.
The conclusion was mixed: Search suffered for some participants
when the irrelevantly textured item was in the target color subset
and for others when it was either color. Some participants’
searches were not affected at all.

One drawback shared by these two studies, however, is that the
tasks required participants to use the additional singleton para-
digm. Because the nontarget singletons could never coincide with
the target, the participants may have been implicitly encouraged to
inhibit the output of a bottom-up process, to some extent; there-

fore, the role of bottom-up processing in standard conjunction
search tasks may be underestimated by the modified tasks in these
studies. To avoid this problem, I introduced an irrelevant feature
that provided no information about the target location but coin-
cided with the target on some trials.

Four Possible Strategies in a Conjunction Search Task

I designed this experiment to use attentional capture to differentiate
four possible strategies that participants could use in a conjunction
search task. A cartoon of the stimuli used and the trial types are shown
in Figure 1. The predicted results, as a function of strategy, are shown
in Figure 2. In the subsequent paragraphs, I provide a brief description
of each strategy and explain how each strategy would be implemented
through use of the bottom-up and top-down systems.

Strategy 1: Efficient Top-Down Selection

The top-down system can be relied on primarily for providing
guidance to the target because no items are unique within either
feature dimension (e.g., half are red, half are green, half are
vertical, and half are horizontal; cf. Wolfe, 1994). There is no
difference in the slope of the functions relating response time to
display numerosity for the three types of target-present trials
(which correspond to Figures 1A, 1B, and 1C).

Strategy 2: Inefficient Top-Down Selection

The top-down system can be relied on primarily for randomly
selecting locations and for conjoining the features at each location
until the target is found (cf. Treisman & Gelade, 1980). Strategy 2
might be differentiated from Strategy 1 by an examination of the
target-present slopes. If this second strategy were operative, one
would predict that the slopes would be steeper than those reported
by Wolfe, Cave, and Franzel (1989), perhaps more like those
reported by Treisman and Gelade. Most important, in either case
the data would indicate that bottom-up processes have little or no
role in standard conjunction search.

Strategy 3: Bottom-Up Selection Within a Subset

The observer can restrict search to one feature (red) and then search
for the feature singleton in the other dimension among just that subset
(the uniquely oriented bar; cf. Zohary & Hochstein, 1989). Note that
for Strategy 3 (two panels), the slowest target-present response times

Figure 1. Examples of a conjunction search task with an irrelevant feature added to the display. The target is
the right-tilted black bar. The irrelevant feature can coincide with the target or either nontarget type, as shown.
TSing � target singleton; NS-O � nontarget singleton-orientation; NS-C � nontarget singleton-color.

49CONJUNCTION SEARCH

(RTs) are on those target-present trials in which the size singleton
coincides with a distractor that is within the subset being searched
through. For example, in Strategy 3 (orientation subset) the slopes
indicate that responses to the target are slowed by the presence of the
size singleton when it coincides with a distractor that shares the
target’s orientation (and are perhaps even speeded to the target when
the target coincides with the size singleton).

Strategy 4: Bottom-Up Selection on All Features

Observers can search for the unique object in the display, as the
target is the only item that is both red and vertical; therefore, if

objects can be compared in their entirety as conjunctions of fea-
tures, the distractors can be excluded in groups (cf. Duncan &
Humphreys, 1989; Found, 1998). For Strategy 4, a prediction
could be made that the irrelevant feature would influence visual
search no matter with which object it coincides.

Method

Participants

Participants were 40 undergraduates reporting normal or corrected-to-
normal vision. All gave informed consent and took part either for payment
or for a course requirement.

Figure 2. Predicted detection response times for the new conjunction experiments.

50 PROULX

Apparatus and Stimuli

Participants were 55 cm from the screen and used a chin rest in a dimly lit
room. Stimuli were presented by a C�� and OpenGL program on an
IBM-compatible computer. In the actual experiment, bars were either blue or
green and either right tilted (45°) or left tilted (– 45°). They were dispersed
randomly in the cells of an invisible grid subtending 6°, 7°, or 8° of visual
angle (with 7 � 7, 8 � 8, and 9 � 9 grid sizes, respectively) for a corre-
sponding display numerosity (three, five, or nine bars, respectively). The bars
were arranged within a subset of the cells of the grid, which were 1° apart,
center to center, and the bar positions were each displaced by a random vertical
and horizontal factor of � 0.2°. The nonsingleton bar size subtended 0.6° of
visual angle in length and 0.15° in width. The size singleton bar subtended 0.9°
in length and 0.15° in width. There was no fixation point, and the background
was black. A size singleton was present on every trial.

Participants were randomly assigned to one of four feature-assignment
groups (10 per condition), each of which had a different set of features
assigned to the target or the nontargets: (a) Group A target was blue and
right tilted (and nontargets were either green and right tilted or blue and left
tilted); (b) Group B target was blue and left tilted; (c) Group C target was
green and right tilted; and (d) Group D target was green and left tilted.

The size singleton appeared on each trial and coincided with the target
on 1/d of the trials, where d is the number of elements in the display. The
size singleton coincided equally often with each nontarget type on the
remainder of the trials. In a pilot experiment with 14 participants, I
determined that the size singleton could capture attention in an orientation
singleton detection task, thus demonstrating that the size singleton was
sufficiently salient and conceptually replicated previous studies in this
regard (Bacon & Egeth, 1994; Theeuwes, 1991).

Procedure

Participants were instructed to look for the particular features that
defined the target for their condition and were informed of the 1/d rela-
tionship between the size singleton and the target. A display of bars
appeared on each trial, and the participant responded “present” or “absent”
with a keypress. Errors were signaled with auditory feedback. Each trial
began after a 2-s intertrial interval. Each participant participated in two
blocks of 270 trials per block. Each block included an equal number of
target-absent and target-present trials and an equal number of trials for each
display numerosity. Order of trial types was randomized. Participants
began with a practice block of 20 trials.

Results

The error rates are shown in Table 1; they followed the same
general pattern as the RT data (see Figure 3), indicating that a
speed-accuracy tradeoff likely did not confound the comparisons
of interest. RTs greater than 3.5 standard deviations from the mean
for each participant were counted as errors, resulting in a loss of
0.88% of the trials. I conducted initial analyses of variance

(ANOVAs) to find whether feature-assignment group (a between-
subjects factor) had any interaction effects with the within-subjects
variables. All interactions with feature-assignment group failed to
reach significance. Thus, the following analyses collapse across
group assignment.

Analysis of Search Efficiency by Trial Type

A repeated-measures ANOVA that analyzed the target-present
data revealed significant effects of trial type, F(2, 78) � 59.0, p �
.01; display numerosity, F(2, 78) � 223.8, p � .01; and an
interaction between these two factors, F(4, 156) � 15.5, p � .01.
A separate ANOVA found a significant interaction between dis-
play numerosity (3, 5, 9) and nontarget-singleton trial type (non-
target singleton-color [NS-C] vs. nontarget singleton-orientation
[NS-O]), F(2, 78) � 4.2, p � .05. The mean RTs are shown in
Figure 3. As seen in Figure 3 and in Table 2, the search slope for
the target-singleton trials (18 ms/item) was shallower than the
search slopes for the nontarget-singleton trials (NS-C: 33 ms/item;
NS-O: 29 ms/item); this observation is supported by the significant
interaction between trial type and display numerosity. Although an
obviously large difference was not seen between the slopes of the
nontarget-singleton trial types, a significant interaction between
nontarget-singleton trial type and display numerosity was found.

Figure 3. The response times for the main experiment are plotted as a
function of display numerosity and trial type: target absent (Abs), nontarget
singleton-color (NS-C), nontarget singleton-orientation (NS-O), and target
singleton (TSing).

Table 1
Error Rates for Each Trial Type

Trial type

Display numerosity

3 5 9

Target present
Target singleton 3.2 4.2 3.8
Nontarget singleton-color 3.6 6.5 10.9
Nontarget singleton-orientation 4.3 5.9 8.8

Target absent 3.0 3.1 3.8

Table 2
Search Slopes (in ms per Item) and Intercepts (in ms) by Trial
Type

Trial type Slope Intercept

Target present
Target singleton 18 624
Nontarget singleton-color 33 589
Nontarget singleton-orientation 29 613

Target absent 35 545

51CONJUNCTION SEARCH

Intertrial Effects

In recent studies, researchers have noted that the introduction of a
singleton could have its apparent prioritization effect as a result of
intertrial priming rather than of reliance on bottom-up processing
(Olivers & Humphreys, 2003; see also Maljkovic & Nakayama,
1994). Generally, this speeded response occurs because the target’s
defining features are the same from one trial to the other; RTs are
slower if the target’s features change. To ensure that priming does not
account for the effect of the irrelevant-size singleton in the main
experiment, the effects of the previous trial also were considered for
this main experiment. If the size singleton was prioritized even on
target-singleton trials that followed nontarget-singleton trials, which
would not be predicted by the priming account, then the mean RT on
those trials also should be significantly lower than the mean RT of the
nontarget-singleton trials (which are shown in Figure 3). A t test
revealed that the mean RT for the target-singleton trials that followed
a nontarget-singleton trial (710 ms for NS-C as trial n – 1, and 705 ms
for NS-O as trial n – 1) were faster than the mean RTs for the
nontarget-singleton trials (783 ms for NS-C; 786 ms for NS-O) and all
four t tests were significant, p � .0125 (Bonferroni-corrected for
multiple comparisons). The mean RT effects of the singleton seen in
Figure 3 were driven just as much by trials preceded by a different
trial type, and it is hard to conceive how this could be attributed to
priming.

Further Analysis and Discussion

The main result was that the irrelevant-size singleton influenced the
prioritization of the objects in the display. The size singleton was

more likely to be attended first. As evidenced by the significant
interaction between display numerosity and trial type, the target was
processed more efficiently if the size singleton coincided with it, but
the target was processed less efficiently if the size singleton coincided
with a nontarget. This result implies that bottom-up processing is used
in search for a conjunction of features. The target-present data pre-
sented in Figure 3 are most similar to the predicted detection response
times for Strategy 4 in Figure 2, in which participants rely on
bottom-up processing of the display to find the target object. Further-
more, it appears that the participants in this main experiment were
more likely to respond “absent” when a nontarget (vs. a target)
coincided with the singleton (see Table 1). This also suggests that
participants were more likely to attend to the size singleton because it
resulted in their missing the target more often when the size singleton
did not coincide with the target.

Strategic control: Bottom-up processing of the display or
subset-selective search? The significant interaction between dis-
play numerosity and the nontarget-singleton trial types leads to an
interesting question: Were the participants in this main experiment
attending to the entire display and using the irrelevant-size singleton
to prioritize items for further processing, or were they attending to a
subset of the items (e.g., the target-defining color) and searching for
a singleton within that subset (e.g., the unique orientation within the
target-defining color subset)? The data were examined for split-half
reliability because if the participants are engaging in a strategy in
which they attend to the color subset and search for a singleton within
that subset, then this strategy should be consistent between partici-
pants and within an individual’s data set.

The scores are displayed in a scatterplot in Figure 4, in which

Figure 4. Scatterplot of the difference of the nontarget-color slope minus the nontarget-orientation slope for
each participant.

52 PROULX

the even trials are on the x-axis and the odd trials are on the y-axis.
The slopes for the nontarget-trial types were compared to see
whether the participants maintained the same strategy, which
would be indicated by having one slope (e.g., the nontarget-color
slope) greater than the other slope (e.g., the nontarget-orientation
slope). The difference between the slopes (nontarget-color minus
nontarget-orientation) resulted in two scores for each participant,
one for the odd trials and one for the even trials. The dotted line is
the linear best fit for the data. Participants used no reliable strat-
egy; this observation was supported by a Pearson correlational
analysis that did not reach significance, r � –.18, p � .25. As
Figure 4 also makes clear, however, some participants were using
a consistent strategy. This suggests that the apparent support for
Strategy 4 (Figure 2) might actually arise from a mixture of both
types of Strategy 3 and Strategy 4. In either case, the data support
the notion that the participants relied on bottom-up guidance; the
following experiments may provide some insight into which strat-
egy is more prominent.

Explicit instructions for subset search. As a follow-up, I per-
formed three experiments with a new set of participants for each.
In the first experiment, I attempted to have all participants engage
in a strategy of searching within the orientation subset (n � 10) or
the color subset (n � 10) by giving them instructions to do so. All
participants searched for a target that was green and right tilted
(45°). Nontargets were green and left tilted (– 45°) or blue and
right tilted. Participants were randomly assigned to one of two
conditions (10 per condition), with a different set of instructions
for each condition: (a) Participants were instructed to search for
the target among the color (green) subset; (b) participants were
instructed to search for the target among the orientation (right-
tilted) subset. Similar to the main experiment, separate ANOVAs
for the data in Figures 5A and 5B exhibited a significant interac-
tion of display numerosity by trial type: F(6, 54) � 3.9, p � .01 for
the color-instructed group and F(6, 54) � 3.1, p � .05 for the
orientation-instructed group. The main consideration here, how-
ever, was the effect of instructional condition. Taking the data in
Figures 5A and 5B as a whole, there were main effects of display
numerosity, F(2, 36) � 155.6, p � .01, and trial type, F(3, 54) �
6.9, p � .01, but the main effect of instructional condition failed to
reach significance, F(1, 18) � 2.4, p � .10. The interactions of
display numerosity by instructions, F � 1; trial type by instruc-
tions, F � 1.3, p � .10; and the three-way interaction, F � 1.5,
p � .10, all failed to reach significance (keep in mind that fewer
participants were used in the follow-up studies; the statistical
power is reduced here compared with the main experiment). Be-
cause there was no reliable effect of the instructional manipulation,
the conditions were combined for further analyses and are dis-
played in Panel 5C. There were significant effects of trial type,
F(2, 38) � 11.3, p � .01; display numerosity, F(2, 38) � 144.4,
p � .01; and Trial Type � Display Numerosity interaction, F(4,
76) � 3.9, p � .01. The results are shown in Figure 5. The data are
surprisingly similar to the main experiment and, thus, indicate that
participants did not adhere to one search strategy, as instructed.

Increased orientation heterogeneity: Implicit motivation for
subset search? In the second experiment, I made an implicit
attempt to affect the search strategies used by the participants (n �
12) by introducing additional orientation heterogeneity to the
display. The stimuli and results are shown in Figure 6. All target-
colored (green) nontargets were oriented 45° to the left, but the

F
ig

u
re

5
.

R
es

po
ns

e
ti

m
es

pl
ot

te
d

fo
r

th
e

fi
rs

t
fo

ll
ow

-u
p

ex
pe

ri
m

en
t
by

in
st

ru
ct

io
na

l
co

nd
it

io
n

an
d

w
it

h
th

e
da

ta
co

ll
ap

se
d

ac
ro

ss
in

st
ru

ct
io

na
l
co

nd
it

io
ns

.
A

:
C

ol
or

–s
ub

se
t

in
st

ru
ct

io
na

l
co

nd
it

io
n.

B
:

O
ri

en
ta

ti
on

–s
ub

se
t

in
st

ru
ct

io
na

l
co

nd
it

io
n.

C
:

C
ol

la
ps

ed
ac

ro
ss

in
st

ru
ct

io
na

l
co

nd
it

io
ns

.
T

he
ap

pa
ra

tu
s

an
d

st
im

ul
i

w
er

e
th

e
sa

m
e

as
th

os
e

of
th

e
m

ai
n

ex
pe

ri
m

en
t,

ex
ce

pt
on

ly
on

e
co

nd
it

io
n

w
as

us
ed

.
A

bs
�

ta
rg

et
ab

se
nt

;
N

S
-C

�
no

nt
ar

ge
t

si
ng

le
to

n-
co

lo
r;

N
S

-O
�

no
nt

ar
ge

t
si

ng
le

to
n-

or
ie

nt
at

io
n;

T
S

in
g

�
ta

rg
et

si
ng

le
to

n.

53CONJUNCTION SEARCH

right-tilted (blue) nontargets could be either 25°, 35°, 55°, or 65°
to the right. There were significant effects of trial type, F(2, 22) �
14.4, p � .01; display numerosity, F(2, 22) � 46.4, p � .01; and
Trial Type � Display Numerosity interaction, F(4, 44) � 7.5, p �
.01. Search was most efficient on the target-singleton trials (19
ms/item) and less efficient for the nontarget-singleton trials (NS-C:
35 ms/item; NS-O: 31 ms/item). The additional clutter would make
selecting the subgroup defined by the target’s orientation ex-
tremely unlikely; however, selectively searching the items that
matched the target’s color would be simple. However, the
results indicated, that, similar to the main experiment and the
follow-up experiment just described, participants relied on
bottom-up processing across all items in the display and appar-
ently did not restrict processing to just a subset of the items (see
Figure 6).

Importance of target versus distractor salience. I conducted a
third and final follow-up experiment to examine the generality of
the findings of the previous experiments to a new combination of

features that define the target, the nontargets, and the irrelevant
feature singleton. The target was now defined as a conjunction of
size and orientation, and an irrelevant color singleton was pre-
sented at each trial (see Figure 7A). All participants (n � 11)
searched for a target that was large and right tilted at 45°. Non-
targets were large and left tilted at – 45° or small and right-tilted.
For this experiment, the irrelevant singleton was defined by color.
Although most of the bars were blue, one bar was red on each trial.
The small bars were the same size as those used in the prior
experiments, subtending 0.6° of visual angle in length and 0.15° in
width. In contrast, the large bars subtended 1.2° of visual angle in
length and 0.3° in width. The results are shown in Figure 7D. The
target-present data revealed significant effects of trial type, F(2,
20) � 5.7, p � .05, and display numerosity, F(2, 20) � 58.2, p �
.01; however, the Trial Type � Display Numerosity interaction
failed to reach significance, F(4, 40) � 1.54, p � .20. The results
suggest that the relatively high salience of the large target (cf.
Braun, 1994) was strong enough to overcome the relatively low

Figure 6. A: Example of the target-present trials in the second follow-up experiment. The target is shown as
the right-tilted black bar. A target is present and also is the size singleton. …