USE OF THESES - Open Research: Home...SOLMAN, R.T. Influence of similarity between target and...
Transcript of USE OF THESES - Open Research: Home...SOLMAN, R.T. Influence of similarity between target and...
-
THESES SIS/LIBRARY TELEPHONE: +61 2 6125 4631 R.G. MENZIES LIBRARY BUILDING NO:2 FACSIMILE: +61 2 6125 4063 THE AUSTRALIAN NATIONAL UNIVERSITY EMAIL: [email protected] CANBERRA ACT 0200 AUSTRALIA
USE OF THESES
This copy is supplied for purposes of private study and research only.
Passages from the thesis may not be copied or closely paraphrased without the
written consent of the author.
-
SELECTION IN VISION
A study of visual search
A thesis submitted in partial
fulfilment of the requirements for
the degree of Doctor of Philosophy
by
Robert T. Solman
Australian National University
January 1977
-
This thesis describes original research
carried out by the author in the Department of
Psychology at the Australian National University;
from March 1973 to January 1976.
Robert T. Solman
-
r i t
PREFACE
am indebted to Dr. Michael Cook of the Australian National
University. Dr. Cook was always willing to discuss my work, his
criticisms were invaluable, and his encouragement during the latter
stages was greatly appreciated. would like to thank Barbara Grosser
who typed this thesis, and Neville Whitworth who bu i It and rna i nta i ned
much of the equipment. would also I i"ke to acknowledge the •
Australian Government for awarding me a post-graduate research
scholarship.
Nine empirical investigati?ns have been reported in this
thesis. The results of five of these nine have been published, and
the results of the remaining four are being considered for publication.
I have listed below the five published papers and two manuscripts which
incorporate the remaining four studies.
SOLMAN, R.T. Influence of similarity between target and i_rrelevant
items on visual information processing. Perceptual and Motor
Skills, 1975, ~. pp.43-48.
Effect of target separation on selective attention.
Peroeptual and Motor SkiUs, 1975, ~. pp. 755-760.
Relationship between selection accuracy and exposure in
visual search. Peraeption, 1975, ~. pp.411-4!8.
Influence of selection difficulty on the time required for
icon formation. Peraeption, 1976, ~. pp.225-232.
Evidence that focal processing involves a build-up of a
visual object. British Journal of Psyahology, in press.
-
lv
Processi~g nominal information during first-stage analysis.
Being considered for publication by the Canadian Journal of
Psyahology.
Relationship beb1een accuracy of search and extended
exposure-time.
Pereeption.
Being considered for publication by
R.T.S.
-
PREFACE
ABSTRACT
TABLE OF CONTENTS
1. INTRODUCTION
1.1
1.2
STIMULUS CONTROL DURING SELECTION
SOME PROPERTIES OF VISUAL SELECTION
VISUAL SEARCH
ATTRIBUTES GOVERNING SELECTION
SERIAL AND PARALLEL PROCESSING
1.3 • THE EMPIRICAL INVESTIGATION
2. EXPERIMENT 1
2.1
2.2
2.3
2.4
2.5
INTRODUCTION
METHOD
DESIGN
MATERIALS AND APPARATUS
PROCEDURE
RESULTS
DISCUSSION
PARALLEL-SERIAL PROCESSING
A TWD-STAGE MODEL
FURTHER INVESTIGATION
Iii
1
3
3
6
8
10
12
13
14
14
15
16
23
24
27
28
v
-
vi
3. EXPERIMENT 2
3.1 . INlRODUCTION 30
3.2 METHOD 31
DESIGN 31
MATERIALS AND APPARATUS 32
PROCEDURE 32
3.3 RESULTS 34
3.4 DISCUSSION 37
4. EXPERIMENT 3
4.1 INlRODUCTION 38
4.2 METHOD 40
DESIGN 40
MATERIALS AND APPARATUS '•1
PROCEOI.RE 41
4.3 RESULTS 43
. 4. 4 DISCUSSION 48
A "TWD-STAGE MODEL 48
PARALLEL-SERIAL PROCESSING 50
RESPONSE SPEEDING 50
MASKING 51
MEMORY SPAN LIMITATIONS 51
4.5 FURTHER INVESTIGATION 52
5. EXPERIMENT 4
5.1 INTRODUCTION 53
5.2 METHOD 54
DESIGN 54
MATERIALS AND APPARATUS 55
PROCEDI.RE 55
-
vil
5.3 RESULTS 55
5.4 DISCUSSION 56
ACCURACY/TIME CURVES 59
6. EXPERIMENT 5
6.1 INTRODUCTION 61
6.2' MEn-taD 61
DESIGN 62
MATERIALS AND APPARATUS 62
PROCEDURE 62
6.3 RESULTS 64
6.4 DISCUSSION 65
IN SUMMARY 67
7. EXPERIMENT 6
7.1 INTRODUCTION 68
7.2 MEJHOD 69
DESIGN 70
MATERIALS AND APPARATUS 70
PROCEDURE 71
7.3 RESULTS 72
7.4 DISCUSSION 75
7.5 FURTHER INVESTIGATION 81
8. EXPERIMENT 7
8.1 INTRODUCTION 82
8.2 ME1HOD 84
'THE PEST PROCEDURE 85
DESIGN 87
MATERIALS AND APPARATUS 88
PROCEDURE 88
-
a.::>
a.4
RESULTS
DISCUSSION
DEVELOPMENT OF STRUCTURE
NAMING
9. EXPERIMENT a
9.1 INTRODUCTION
9.2 METI-IOD
DESIGN
MATERIALS AND APPARATUS
PROCEDURE
9.3 RESULTS
9.4 DISCUSSION
FURTHER I !'NEST I GAT ION
10. EXPERIMENT 9
10.1 INTRODUCTION
10.2 METHOD
DESIGN
MATERIALS AND APPARATUS
PROCEDURE
10.3 RESULTS
10.4 DISCUSSION
FURTHER INVESTIGATION
11. EM"IRICAL FINDINGS AND FINAL DISCUSSION
11.1 THE TIME CoURSE OF VISUAL SEARCH
11 THE STUDIES
DISCUSSION.
11.2 NAME DIRECTED SEARCH
THE STUDIES
DISCUSSION
v ll !
a9
92
92
96
9a
99
99
100
100
102
103
104
105
105
106
106
107
107
109
112
114
114
114
11a
119
119
121
-
11.3 . GEI'ERAL DISCUSSION
NEISSER'S DESCRIPTION OF THE TWD-STAGE MODEL
REFERENCES
APPENDICES
1. DATA RELEVANT TO EXPERIMENT 1
2. DATA RELEVANT TO EXPERIMENT 2
3. DATA RELEVANT TO EXPERIMENT 3
4. DATA RELEVANT TO EXPERIMENT 4
5. DATA RELEVANT TO EXPERIMENT 5
6. DATA RELEVANT TO EXPERIMENT 6
7. DATA RELEVANT TO EXPERIMENT 7
6. DATA RELEVANT TO EXPERIMENT 6
9. DATA RELEVANT TO EXPERIMENT 9
121
122
127
137
138
143
145
154
157
160
164
166
166
-
ABSTRACT
This thesis is concerned with the selective
processing of information which arrives via the visual system. The
introduction discusses the problem of stimulus control during selection
and specifies the aim of the thesis. Nine experiments are reported.
X
The first six examine the relationship between selection accuracy and
stimulus exposure-time. It is suggested that a plausible explanation of
the obtained results can be found in a two-stage .model of information
processing (e.g., Neisser, 1967). The remaining three experiments address
the question of whether a letter-name can direct the selection process,
and the results, while somewhat indecisive, do not indicate that it can.
Specifically,
a) in Experiment 1 accuracy of search for a single target
declined as the number of irrelevant items, and their
physjcal (shape) similarity to the target, increased.
These results could be explained either by sophisticated
parallel-serial models of information processing, or by a
two-stage model; Experiment 2 suggested that selection
was not based on a limited retinal region, thereby
negating the possIbilIty of the dIstance separatIng I terns
influencing the results of Experiment 1; Experiment 3
varied stimulus exposure-time along with similarity and
number of items. The results replicated the effects
obtained in Experiment 1, and the negatively accelerated
accuracy/time curves demonstrated a plateau effect. The
level of the plateau differed for each curve and these
differences were attributed to errors during .initial
-
xi
selection; Experiment 4 found the long term course of the
accuracy/time curves (i.e., exposure-time varied from 100
to 500 msec). This gave subjects the opportunity to begin
a second processing run, but, possibly due to !mage-decay
or to the absence of eye movements, no evidence of a second
run was detected; Experiment 5 supplemented the time
course data obtained in Experiments 3 and 4 by sampling
behaviour over an exposure-time range of 15 to 400 msec.
The results showed that both image-decay and the
opportunity for eye movements had no effect; the results
of Experiment 6 replicated those obtained in Experiment 3,
and indicated that while the relationship between target
and irrelevant items affected accuracy of selection, it did
not influence the time required for this selection.
b) Using the PEST procedure Experiment 7 firstly, demonstrated
that the process of bui !ding up a structural
representation of an alphabetic character requires time,
and secondly, indicated that the prior development of
object structure may be necessary for naming; Experiment 8
showed that the name of a target Jetter need not interfere
with its selection; In Experiment 9 a confounding of shape
and name information made a straight-forward interpretation
of the results difficult. The data, however, did not
suggest that nominal information was used to direct the
selection of the target.
Finally, the discussion speculates further on the nature of the
information processing system underlying the findings.
-
1. INTRODI.X:TION
1.1 Stimulus Control During Selection
One of the most striking and ubiquitous aspects of
visual selection is the way In which characteristics of the stimuli can
and do control the selection process. This control is demonstrated both
1
in time-limited selection tasks and in visual scanning tasks. In both
cases the stimuli are usually constructed from the set of alpha-numeric
characters, and only a limited number of the items presented to the
subject are to be reported. When the time allowed for selection to take
place is limited the usual procedure is to display briefly a number of
Items and to Indicate those to be reported either by the presence of an
adjacent marker (e.g., Averbach and Coriell, 1961), or by specifying some
attribute, such as colour, which distinguishes relevant from irrelevant
Items (e.g.,Eriksen and Collins, 1969; Von Wright, 1968}. The marker or
attribute then controls selection in that it "drives" or directs the
selection mechanism(s) in such a way that the relevant items appear to be
processed, and are selected, to the exclusion of the irrelevant items.
This driving of the mechanisms of selection is even more dramatically
manifested In the scanning task. That Is, subjects are usually required
to scan a list containing upwards of fifty rows of five or six characters
each, with the aim of finding a specific character (termed the target)
(e.g., Neisser, 1963 & 1964). In the course of the search some attnbute(s)
of the target direct selection in such a way that the mass of non-target
characters Is rejected or excluded, and the target is perceived.
-
2
The selection of items defined by attributes such as
location, shape, and colour, implies that the specified attribute needs
to be processed prior to the recognition of the relevant Item, and this,
in its turn, suggests that, in some way or at some level, all items must
be processed before selection takes place. This somewhat paradoxical
suggestion has had ramifications for theories intended to explain
selective phenomena. For example, it was primarily an argument based on
this suggestion, and supported by empirical data (cf., Cherry, 1953;
Cherry and Taylor, 1954; Gray and Wedderburn, 1960; Treisman, 1964a, 1964b,
& 1964c), which led to Treisman's (1964a, 1964b, & l964c) revision of
Broadbent's (1958) "Filter" theory of attention. 1 Alternatively, it gives
support to those psychologists (e.g., Deutsch and Deutsch, 1963), who see
no need for a notion of selection during processing and would therefore
argue, as do Hodge (1959 & 1973) and Lawrence and La Berge (1956), that
selection in vision follows the complete processing of all Items. But, it
is not the intention of this thesis to enter the controversy over the
relative merits of the various theories of attention, 2 nor is it intended
that the perception/response question be considered in any detai1. 3 ·The
general aim of the investigations reported here is to explore some of the
properties of selection during visual search, and to seek an explanation
of search behaviour.
1. Some confusion surrounds the use of this term (see Spearman, 1937 for a detailed account), but for the purposes of this thesis and in line with Broadbent's (1958) description, it will be treated as a p~ceas, which by selectively allocating a scarce resource (capacity, effort, energy) allows the human information processing system to deal with some inputs to the exclusion of others.
2. The reader interested in theories of attention should refer to Broadbent (1958), Broadbent and Gregory (1964), Deutsch and Deutsch (1963), Deutsch, Deutsch, lindsay and Treisman (1967), Trelsman (1964a, 196/.jb, · 1964c, & 1969), as well as to corrprehensive reviews found in Egeth (1967), Kahneman (1973), Moray (1969), Neisser (1967), and Norman (1976).
3. Good reviews of the debate over whether selection in vision can occur during perception or whether this selection must wait until after the perception of all items, can be found in Egeth (1967) and Haber (1966).
-
3
1.2 Some Properties of Visual Selection
As the errp i rica l research reported in the present thesis
is concerned with visual search behaviour and the properties exhibited
therein, the following discussion will be restricted to a description of
visual search behaviour, and to a brief review of studies which examine
directly the attributes governing visual selection, and to a mention of the
debate concerning the mode of processing of elements from brief visual
displays.
1.2.1 Visual Search
The most theoretically oriented treatment of search
behaviour can be found in Neisser (1967), and Is based (in general) on the
results of search studies carried out by him and his colleagues (Neisser,
1963 & 1964); Neisser and Lazar, 1964; Neisser, Novick and Lazar, 1963).
In these experiments subjects scanned from top to bottom through vertical
lists, each consisting of fifty lines (called "items"), to find the target
embedded in each list. In general, the target was a single alpha-numeric
character, and the five or six alpha-numeric characters present in each
line were in some meaningless combination. The consistent finding was a
linear relationship between response latency and the vertical position of
the target in the list. Assuming that the starting time and stopping time
(includes target recognition time and response time) are independent of
the position of the target, their contribution is to the intercept of the
search time function, not to Its slope. The slope is then a measure of
the time required to examine the context items, and it is relatively
uncontaminated by response factors.
An efficient asymptotic search rate was achieved after
several days practice, and the following aspects of search behaviour were
-
recorded:
(a) Subjects reported that they did not "see" the context i terns,
and there was no evidence from post-search tests that these
items were recognized during the search. This result was
supported by Schapiro (1970), using a post-search recognition
task. However, the results of a study specifically designed
to examine the effects of varying degrees of physical
similarity between the target and particular irrelevant items
(Solman, 1975), showed a significantly higher post-search
recognition level for the most "target-like" item. This
suggested that the depth to which field items are likely to
be processed may depend on their physical similarity to the
target;
(b) Search for a given target was faster than search for its
absence;
(c) The physical similarity between the target and field items
had a marked affect on rate of search, with the search rate
declining as the similarity increased. This similarity effect
has been confirmed in a number of studies (Estes, 1972; Gibson
and Yonas, 1966a; Kaplan, Yonas and Shurcliff, 1966; Mcintyre,
Fox and Neale, 1970);
(d) After training, the search rate appeared to be indpendent of
the number of alternative targets, i.e., subjects were as fast
searching for one of ten possible targets as when required to
search for a single target. This result has been the subject
of some controversy, i.e., a similar finding was obtained in
a study by Gibson and Yonas (l966b), where searching for one
of two targets was no harder than searching for one alone.
But, Krlstofferson (1972) and Wattenbarger (1968) have shown
that only when the error rate is high is search equally rapid
-
5
for one and many targets. 4
The experimental results suggested (to Nelsser) that
recognition may best be described as a hierarchical, two-stage system of
decision processes. When searching for a particular target letter, the
first processing stage carries out elementary or low level tests on target
features (it was not possible to specify the exact nature of these features,
but they probably consist of physical or structural attributes, such as
angularity, openness, straight or curved lines, etc.). These tests are
carried out in parallel, and items which fail the test are passed by
without being subjected to second-stage processing. This model implies
that the perception of field items differs from the perception of targets.
The former are rejected without being recognized, and the latter are fully
processed or constructed and thereby recognized.
The suggestion that first-stage processing is carried
out on physical pattern features is consistent with the well known
physiological work of Hubel and Wiesel (1962 & 1965), and although there
Is no guarantee that the! r findings apply to human pattern perception, the
notion of pattern features is now widely accepted (cf., feature lists for
alpha-numeric characters by Geyer and De Wald, 1973; Gibson, 1969;
Sutcliffe, 1971). However, the assumption that these attributes are of a
physical kind, has been challenged by experimental results which suggest
that the names of alpha-numeric characters may also be processed at this
level (see Brand, 1971; Henderson, 1973; Ingling, 1972).
4. Further investigation of this question by Yonas and Pittenger (1973) produced evidence of a decline in rate of search and an increase in the number of errors, when search was for one of four targets as compared to one target alone. However, rather than suggesting that the Neisser et al., (1963) results are difficult to rep! icate, the authors point out that their findings indicate that, when conditions permit, subjects can and do make subtle adjustments in their speed· accuracy trade offs. This point was later confirmed and elaborated by Neisser (1974).
-
6
Neisser's (1967) explanation of search behaviour has not
been universally accepted. But, the important aspect of this behaviour is
that, whatever the pattern attributes are which direct the search, it is
efficient and fast with some forty to sixty items (Neisser, 1967)
processed every second.
1.2.2 Attributes governing selection
Discussion of the attributes which direct or govern
selection is more frequently found in the literature dealing with time-
limited selection than in the literature which deals directly with visual
search. This is perhaps due to the versatility of the partial-report
technique introduced by Sperling (1960). Sperling presented subjects
with a tachistoscopically displayed array of letters, and after
termination of the display they received an auditory tone specifying which
one of the presented rows was to be reported. The superiority of
performance under these conditions when compared with "who 1 e report"
(subjects were asked to report as many items as possible) was taken as
evidence for the existence of a short-term visual memory. 5 However, the
importance of the study in the present context is that, a) location was
used as a basis for selective processing of part of the display, and b)
the technique offers a method of investigating the effectiveness of other
attributes as controllers of selection.
Using the partial-report technique, Von Wright (1968)
compared the effectiveness of a number of stimulus attributes. His
results showed that selection was possible if the attribute specified was
location, chromatic colour, achromatic colour, or size, but when
5. The existence of a rapidly decaying, short-term visual memory (see Averbach and Co riel 1, 1961; Coltheart, Lea and Thompson, 1974; Dick, 1974; Nelsser, 1967) is now generally accepted (Holding, 1970, 1971, 1972 ,& 1973 is a notable exception).
-
7
orientation was the basis for selection performance deteriorated markedly.
'In a similar study, using a circular arrangement of display items, Eriksen
and Collins (1969) showed that efficient selection was possible when a line
appeared beside the i tern to be reported, when the "to-be- reported" i tern
was indicated by an arrow radiating from the centre of the circle, and
when an instruction specified the report of an Item diametrically opposed
to the one indicated by the line. However, subjects were unable to select
efficiently when the cirterion was a number specifying a particular item
position.
While more remains to be done before it wi 11 be possible
to completely specify the attributes which can effectively direct the
selection process, both the Von Wright (1968) study and the Eriksen and
Collins (1969) study suggest (in contrast to the Brand (1971), Ingling
(1972), and Henderson (1973) studies) that they may be of a formal or
physical kind.
The partial-report technique has also proved useful for
investigating the "tuning" limits of selective processing. Eriksen and
his colleagues (Colegate, Hoffman and Eriksen, 1973; Eriksen and Hoffman,
l972a & 1973; Eriksen and Rohrbaugh, 1970). in a series of studies in
which they varied the interval between the onset of an indicator (usually
a line beside the position where a to-be-reported item was to appear) and
the display, showed that even when given sufficient time to fully process
the indicator, a) subjects were incapable of restricting processing to
the indicated item if the separation between adjacent itemswas less than
approximately one degree of visual angle, and b) the number of errors
increased with an increase in the total number of display items. These
results suggest that there are limits to the precision with which
processing capacity can be allocated, and the lower limit may be
represented (with the stimuli used) in spatial terms as an item separation
-
8
of one degree of visual angle. Also, the preliminary processing, which
isolates the controlling attribute in order that selection may occur, is
of sufficient depth for an increase in the total number of items to cause
a decrease in the probability of the to-be-reported item being selected.
1.2.3 Serial and parallel processing
The mode of processing of elements from brief visual
displays has been examined in some detail in recent years. 6 The most
productive method of investigation has been the detection paradigm, where
the subject is asked to make a simple response as soon as he detects some
predefined aspect of the display. The simple response, and prior
specification of what is to be looked for, means that there is no need to
store any but the predefined aspects of the display in short-term memory,
thereby limiting the involvement of memory. The two major detection tasks
are search, and same-different discrimination. In the visual search task
used the subject may be asked to indicate whether or not the target is
present in the array or alternatively, he may be asked to indicate which
one of two (or more} specified targets is contained in ar. array. In the
same-different task subjects are asked to indicate whether the elements of
the array are all identical to one another or whether (at least} one
element differs from the others.
A serial model is one in which a decision is reached by
examining the elements of a display one after the other, and a parallel
model allows for the stimultaneous examination of stimulus elements. Some
of the strongest support for the serial model comes from yes-no search
tasks. Sternberg (1966 & 1967} studied visual search in a situation where
memory search was also investigated. Subjects were given a short list of
6. Although this issue is not germane to the empirical work reported later in the thesis, the nature of some of the data necessitates that serial and parallel interpretations be considered. Consequently, a brief preliminary discussion is in order.
-
9
digits to memorise, and then presented with a visual display of a varying
number of digits. Themsk required that the subject Indicate whether or
not any of the digits in memory matched any of the digits In the display.
The principal finding was that reaction time increased with increases in
the number of display items, and Sternberg concluded that both V'isual and
memory search progress In serial at about the same rate (37 msec/element),
but memory search is exhaustive while visual search is self terminating.
Other investigators have also obtained results showing fairly uniform
increases in reaction time with increases in the number of display items
(e.g., Atkinson, Holmgren and Juola, 1969; Estes and Wessel, 1966;
Nickerson, 1966). Apart from some debate over whether search (visual and
memory) is self terminating or exhaustive, they support the suggestion
that the items are examined one after the other.
There are two main parallel processing models to be
found In the literature. One assumes that processing capacity is limited,
and that as the number of items (or amount of material ) to be processed
is increased, the available capacity must be more thinly spread over the
parallel input. Consequently, errors and reaction time would be expected
to show an increase as a function of the n4mber of display Items (see
discussions by Atkinson, Holmgren and Juola, 1969; Corcoran, 1971;
Townsend, 1971c). The other assumes that the processing capacity is
independent of the number of items to be processed (an Independent-
parallel-channels model), and that each element is allocated to a
separate channel. This model would predict no effect due to increases
In the number of display Items, unless the total number of items exceeded
the available number of channels. Results obtained by Egeth, Jonides, and
Wall 0972), and Gardner (1973) gave some support for a model of
independent parallel channels.
An interesting aspect of this research is that rather
-
than demonstrating that items are processed one at a time in serial or
simultaneously. in parallel, the data have suggested to some
investigators that the mode of processing need not conform exclusively
to either class of model (e.g., Estes, 1972; Rosenbaum, 1974). This
complicated situation has been further exacerbated by Townsend (1972)
pointing out that some serial and parallel models are empirically
indistinguishable.
1.3 The Empirical Investigations
The reported empirical studies take a look at
10
accuracy of search in brief visual displays. They were designed to
examine variables affecting the accuracy with which target item(s) are
selected in a tachistoscopically displayed circular arrangement of
alphabetic characters. The majority of the studies examined the effects
of changes in the number of irrelevant stimulus items, and changes in the
degree of visual similarity between target and irrelevant items. In
particular, the first study manipulated these variables at a masked
stimulus exposure-time of 100 msec. The tachistoscopic exposure was
intended to maximize any influence that variations in stimulus
characteristics might have on the accuracy of selection, and the results
did in fact show an interaction between number and similarity. Subsequent
discussion of this finding (see Experiment 1 for details) specified the
need for a concurrent variation of exposure-time, and the empirical
investigations which followed were designed to explore the relationship
between selection accuracy and masked exposure-time.
Although inquiry into the form of the relationship
between selection accuracy and exposure-time constitutes the main thrust
of the empirical investigations, the final three studies (i.e., Experiments
7, 8, and 9) have as their aim the investigation of a more specific issue.
-
That is, they were designed to examine the possibility that the name of
letters of different case can direct the selection process.
11
-
12
2. EXPERIMENT 1
2.1 Introduction
As referred to in the introduction to this thesis, Neisser
and his associates (Neisser, 1963; Neisser, Novick and Lazar, 1963;
Neisser, 1967) noted that an increase in the physical similarity between
the target item(s), and the irrelevant items in a visual search task
caused the search rate to decline. More recently, similarity between the
target and irrelevant items has been varied in studies by Estes (1972),
and Mcintyre, Fox, and Neale (1970) using brief visual displays. In
accordance with the earlier observations they found that increases in
similarity caused reaction time to increase and accuracy to decrease, and
thus confirmed the importance of similarity as a variable influencing the
selection process. This suggests that manipulation of the physical
similarity between the target and the irrelevant items (if accompanied by
variation in the number of irrelevant items), could aid in explaining
selection during visual search.
In this study the physical similarity between the target and
the irrelevant items, was manipulated in a situation using a
presentation sufficiently brief to force subjects to respond before they
had processed sufficient information to be confident of being correct.
This procedure was intended to maximise the influence of information
processing carried out early in the process of selecting and recognising
the target. In many studies evidence of errors made early in the
selection process is difficult to detect, as subjects usually have
sufficient time to compensate for incorrectly selected items by further
-
13
processing. On the other hand, if the input is avai !able (in the form of
the physical stimulus and/or in the form of a short-term visual store)
for only a brief period, accuracy of response should depend directly on
the precision of the early processing. Therefore, the aim of this study
was to examine the effect of variations in the similarity between target
and irrelevant items, on the accuracy of early processing during
selection of the target in a visual search.
2.2 Method
The task was visual search. Throughout the experiment a
trial consisted of the following steps. After placing a stimulus card in
the tachistoscope E gave a verbal ready signal upon which S fixated the
cross, and pressed a hand switch •1hich removed the cross and initiated
the display. The display remained for 100 msec, and was replaced by a 50
msec mask. 7 At the completion of each trial S recorded his response by
marking one of 18 possible target positions on a prepared sheet, and by
placing a confidence rating (4 ~ sure, 3 = almost sure, 2 = might have been, and 1 = guessing) alongside. A response was recorded as correct if
the target position or one either side was marked.
7. It is assumed that the backward masking procedure prevent further processing of the stimulus input. Whether this is due to the perceptual integration of the contours of the test stimulus with those of the mask (Eriksen, 1966; Kinsbourne and Warrington, 1962), or to the interruption of information processing caused by the arrival of the mask (Dilello, Lowe and Scott, 1974; Kahneman, 1968; Kohlers, 1968; Liss, 1968; Turvey, 1973), is unimportant here.
-
14
2. 2. 1 Des i g n
A three-way design was used, with the similarity between
the target and the irrelevant items (high-similarity, and low-similarity)
and the number of display items (2, 6, 12, and 18) varied within Ss, and
the order of presenting the high-similarity and the low-similarity
displays varied between Ss.
Subjects were 6 undergraduates (4 males, and 2 females)
with normal 6/6 vision, and each received $1 per hour for taking part in
the experiment. Three Ss searched high-similarity displays before low-
similarity displays, and vice versa for the other 3. All Ss searched
for 7, 1-hour sessions (3 practice, and 4 experimental).
2.2.2 Materials and apparatus
The .upper case letters of the alphabet provided an over-
learned and well-defined population of nameable visual forms, and a
subset of 13 items was selected for use in the experiment. To assess
the inter-item similarity the confusion matrices derived by Townsend
(1971a, and 1971b), and the feature analyses specified by Geyer (1970),
Gibson, Osser, Schiff, and Smith (1963), Laughery (1969), Rochester,
Johnson, Amdahl, and Mutter (1959), and Sutcliffe (1971) were examined.
The letter F provided a reasonable range of inter-item similarity when
compared with all other letters, and was selected as the target item.
After selecting the target, the non-target items were ranked in ascending
order on the basis of their physical similarity to the target, and the
top and bottom six letters were selected to constitute the high-similarity
(E,T,H,J,P,L) and low-similarity (C,G,O,Q,U,W) sets respectively.
-
15
Stimulus displays were constructed by mounting black
Letraset letters (18 pt Grotesque 216) on white cards. The letters
could be positioned at any of the 18 equally spaced loci on the
circumference of an Imaginary circle. When positioned, the letters were
0.27° of visual angle high with a stroke width of 0.06°, and a minimum
distance of 0.43° separating adjacent Items. In total 128 display cards
were constructed, I.e., there were 12 present (with an F) and 4 absent
(without an F) cards constructed for each level of number combined with
the two levels of similarity (see Figures 1, 2, 3, and 4 for examples).
The fixation card had a black cross mounted at Its centre
and the mask consisted of a jumble of broken, distorted, and whole
Letraset letters. All cards were displayed In a Model GB Scientific
Prototype Three-Channel Tachistoscope with channel luminance for fixation,
stimulus, and mask of 10.8, 17.2 and 38.7 mlllllamberts respectively.
Ss recorded their response on sheets with the 18 loci clearly shown.
2.2.3 Procedure
The nature of the experimental task was explained to S
upon his arrival at the first practice session. He was told that it was
a visual search task, that when present the target appeared only once,
that the fixation cross should be In clear focus before the initiation
of a display, that the confidence rating given was to refer to the
presence of the target and not to certainty concerning Its location, and
he was discouraged from responding negatively. He was then given 64
trials with one similarity level followed by 64 trials with the other, at
a stimulus exposure of 400 msec. The order of presenting the high-
similarity and low-similarity displays, with two consecutive experimental
sessions devoted to each, was maintained for any particularS throughout
-
16
the study. The same procedure was followed for practice sessions 2 and
3 using stimulus exposures of 200 msec and 100 msec respectively.
During each experimental session Ss searched on 16
practice (12 present and 4 absent), and 132 experimental (100 present,
and 32 absent) trials, at an arbitrarily selected brief stimulus exposure
of 100 msec. This provided a total of 50 responses for each
experimental condition. The 2, 6, 12, and 18 item displays were
randomly presented with 4 displays 3 present and 1 absent) at each
level. The modification of a simple random presentation, was used to
ensure that an appropriate absent card appeared with each 3 present cards.
2.3 Results
The experimental procedure raised the question of what ~1as
to constitute a correct response. During the practice sessions it was
noted that even when confident that they had found the target, subjects
were frequently unable to place it without making position errors, and
since these errors were concentrated one position either side of the
target position, the target position and one either side was scored as
correct.
The mean number of correct responses, for the high-
simi 1 ari ty and low-simi 1 ari ty conditions in combination with the number
of display items, are shown in Table 1 (individual data is shown in
Appendix 1, Table 1), and are illustrated as the percentage of correct
responses in Figure 5. Subjects were able to search the 2-item, low-
similarity displays without error, and the 2-item, high-similarity
displays with an average error of only 1%, and as a consequence, data for
these displays were not included in the statistical analysis.
-
p l T H F J
T H J
J T l H E
E
p
F
p
p
l T
p
H T
E
H
F
17
l T F P
H H
p L
T J
T
FIG. 1: Examples of high-similarity, present displays containing 2, 6,
12, and 18 items.
-
H H p p
l L
H p
T L L
E T T
T
' p
L l
l
J
J
J
H J
l
E
p T J
TTT l
H
T
fiG. 2: Examples of high-similarity, absent displays containing 2, 6,
12, and 18 Items.
18
H H
p
-
19
u Q u u Q u Q w
w 0 c w w 0 0 F c Q G u F
OGo wo GQu w
1 ..
Q F Q
c u u
G
FIG. 3: Examples of low-similarity, present displays containing 2, 6,
12, and 18 items.
F
-
uuw G 0
0 0 w
u WGu uw
c
0
u Q
0 Q
u G
Q
w
G
0 Q
Q
20
c WQ
w
u 0 OGG
0
FIG. 4: Examples of 1()'(1-simi larity, absent displays, containing 2, 6, 12, and 18 items.
-
21
TABLE
Mean number of correct responses (maximum 50) for the
combination of high-similarity and low-similarity conditions with the
number of display items.
Number of Disp Jay I terns High-Similarity Low-Simi Jarity
2 49.50 50.00
6 35.50 47.33
12 15.67 41.50
18 11 . 17 34.00
Observations of Figure 5 showed that accuracy declined as
the number of items contained in the displays increased, and that while
this occurred in both high-similarity and low··similarity conditions, it
was substantially greater in the former. These observations were
supported by a three-way analysis of variance carried out on the data.
The analysis (see Appendix 1, Tab Je 2) showed that both the rna in effects
of similarity between the target and the irrelevant items 8 and number of
displ9y items were significant (F(1,20) = 266.9, P < 0.01; and F(2,20) =
81.02, P < 0.01 respectively), and that the two-way interaction between
similarity and number "'as significant (F(2,20) = 11.89, P < 0.01). The
three-way interaction between similarity, number, and the order of
B. The high-similarity set of irrelevant items contained the letter E, and since F and E share common features (Solman, 1975), it was possible that the significant effect of similarity was caused by confusion between these two letters. To examine this possibility the contribution of E was removed (this was accomplished by considering as correct all responses whereForE were recorded (Appendix 1, Table 3) and the data reanalysed (Appendix 1, Table 4). The results confirmed the effect of similarity on accuracy (F(1,4) = 16.09, P < ~.05), and its interaction with the number of display items (F(2,8) = 6.48, P < 0.05). Therefore, the similarity effects in the main data (Table 1, and Figure 5) cannot be attributed solely to the subjects confusing F and E.
-
1.1.1
-
23
presentation of the high-similarity and low-similarity conditions also
reached significance (F(2,20) = 4.63, P < 0.05).
2.4 Discussion
The significant decrease In accuracy caused by the Increase
in physical similarity between the target and the irrelevant items, is
consistent with the data obtained by Estes (1972), and Mcintyre, et aZ.
(1970). The decline in accuracy with increasing numbers of display items
is consistent with the results of a number of reaction time studies (e.g.,
Atkinson, Holmgren, and Juola, 1969: Estes and Wessel, 1966; Holmgren,
1970), but it contradicts the findings of Egeth, Jonides, and Viall (1972).
Using circular stimulus displays containing 1 to 6 items, Egeth, et al.
found that the number of items did not influence reaction time in a number
of search tasks, and they considered their resu 1 ts to be l nd i cat l ve of
l·nformation processing along independent parallel channels. Ho.vever, when
the number of display items··was varied beb1een subject groups, reaction
times were significantly shorter for l-item displays than for 6-item
displays. This latter finding is not consistent with the independent-
parallel-channels interpretation, and suggests that subjects may have
adopted different strategies9 in the two situations and/or reaction time
measures, by allowing optional and unnecessary extra processing, may have
been ureliable as estimates of essential processing time. The substantially
greater performance decrement caused by increasing numbers of items in the
present high-similarity condition when compared with the low-similarity
condition (supported by the significant two-way interaction between
similarity and number), is consistent "lith the results obtained by Estes,
and its implications for models of information processing will be discussed
later.
9. The results of investigations by Neisser (1974) and Yonas and Pittenger (1973), suggest that the selection mechanism is sufficiently flexible for this to be a likely explanation.
-
24
The significant three-way interaction between similarity,
number and the order of presentation of the two similarity conditions,
reflected an Inconsistency in the performance of one group of subjects,
i.e., one subject group performed at a consistently higher level than the
other when searching low-similarity displays, but this superiority was not
maintained when searching 12-item and 18-item, high-similarity displays.
Therefore, it is of no obvious theoretical importance.
2.4.1 Parallel-serial processing
Any study of selection in vision needs to include some
mention of the mode of processing issue. But, as is forcefully pointed
out by Townsend (1972), the parallel-serial question is a complicated
one, and the sophistication of many of the theoretical models has
temporarily outstripped the ability of investigators to distinguish
between them. (" •.• despite its theoretical importance ( .•.• ) the
parallel-serial question should be ruled as unresolved in most of the
contexts in which it has been studied." Townsend, 1972, p.198). Under
these circumstances, there can be no suggestion that the present data are
unable to be explained by some parallel or serial information processing
models. However, they do present problems for relat·ively simple,
strictly serial and strictly parallel types.
For example:
(a) A serial mode of processing can account for the high-
similarity curve (see Figure 5) if the consequences of a brief
(100 msec) masked exposure are considered. Under these
conditions the mater.ial in the display and its "image", is not
av.ailable for sufficient time for subjects to process all items
from any but the 2-item displays. That is, in general the
the number of display items will exceed the average number
~rocessed. The most obvious prediction from a serial model,
-
25
is that once the number of display items exceed the
average number which can be processed in the time
available, any further increase should either have no
influence on this average value or slightly depress it
(if larger numbers introduce some early processing
difficulty). Estimates of the average number of items
processed from displays containing 6, 12, and 18 items
(see Appendix 1, Table 5 for individual data), show that
the numbers of items processed are 4.26, 3.76 and 3.69
respectively, and the differences between them are not
significant. On the other hand, the corresponding
estimates for the low-similarity curve show that the
average number of items processed is dependent on the
number of items present in the display, and not on the
processing time available. That is, the estimated
numbers are 5.68, 9.96, and 12.24 respectively and the
differences between them are significant. This result
would not be predicted by a serial model where item
processing times are equal, independent, and not
influenced by display density.
(b) A limited capacity parallel processing model with a
roughly equal distribution of capacity per item, can
account for the low-similarity curve but not for the high-
similarity curve. If there is a limit to the amount of
processing capacity available, then an increase in the
number of items to be processed requires a reduction in
the capacity allotted to each item. Since there is only
limited processing time, a reduction in capacity per
item should cause a roughly proportional decrease in
performance. A decrease of this kind can be seen in the
-
26
low-similarity condition, where the addition of 6 items
from 6 to 12 gave rise to a 12% drop in performance and
the addition of a further 6 items from 12 to 18 dropped
performance by 15%. The similar decrease in the high-
similarity case is not proportional (40% and 9%
respectively). In a simi Jar manner a model of
independent parallel channels with approximately six
avai !able channels (Egeth, et al. 1972) can only account
for the ]o>;;-similarity data. These results are explained
by assuming that the processing time is sufficient to
partially process 18 items,i .e., the six channels are
twice "refi !led", and a progressive decay of information
causes an increase in the proportion of errors with each
refill. To account for the high-similarity data, it is
necessary to assume that the increase in the d iff i cu 1 ty
of the discriminations increases processing time
sufficiently to prevent the channels being refilled.
While this is possible, the necessary threefold increase
l s not consistent l•li th a notion of independent channe 1 s,
The present data allows us to reject these simplified
models. There are, however, a variety of more sophisticated models based
on this parallel-serial distinction which might account for the results.
For example, a combination of parallel and serial processes would account
for the data by assuming that the mode of processing depends on the
difficulty of the selection task (i.e., a parallel mode when the target
and the irrelevant items are of different shape, and a serial mode when
they are similar). We cannot reject this on the basis of the present data.
-
27
2 .IJ. 2 A two-stage mode 1
As discussed in the introduction to the experiments,
Neisser considers that human information is characterised by a two-stage
hierarchy. The first stage or the period of general diffuse analysis
processes the entire parallel input in a fast, crude, and wholistic
fashion, and, as might be expected, is prone to error. In contrast, the
second stage or the period of detailed concentrated analysis processes
input "items" in a sequential, deliberate, attentive, and detailed manner.
And, this two-stage model of analysis provides a plausible explanation of
the interaction between the variable of similarity and the variable of
number.
Thus, the increase in the similarity between the target
and the irrelevant items, and the Increase in the number of irrelevant
items can be seen as changing the complexity of the task. That Is, the
feature processing carried out during first-stage analysis is more likely
to result in an error (i.e., result in the selection of an irrelevant
item) when, in the high-similarity condition, the irrelevant Items share
many features with the target. Also, as the number of Items is increased '
(independent of their feature relationship with the target), the mass of
feature material to be processed increases, and this will increase the
chances of an error during selection simply because there are likely to
be more items identified as having target-like features. These errors,
made during the early stage of processing, result in the detailed analysis
of an irrelevant item, and under conditions of lengthy or non-masked
exposures, they can be compensated for by the processing of more items.
However, in the case of the 100 msec masked exposure-time repeated
detailed processing is unlikely, and an increase in errors during the
first stage will cause a decline in performance. Hence, the interaction
-
28
between similarity and number, is due to the addition of the same numbers
of items in the high-similarity and low-similarity conditions, causing a
much greater decline in accuracy in the former. This is not surprising
since the addition of physically similar items introduces many target
features, and makes the task of selecting the target item, on the basis of
its features, very difficult.
2.5 Further investigation
At least two aspects of this study suggest avenues for
further investigation. That is:
(a) When constructing the stimulus displays, the dlstaoce between
adjacent items increased (see Figures 1, 2, 3, & 4). Since there is
reason to suspect that "traversing" empty space may require time (see the
torch-beam analogy detailed by Eriksen and Hoffman 1972b), in
experimental situations where time is limited, performance may be
influenced by the distance separating Items. While it is unlikely that
an effect of distance would alter the obtained pattern of results (see
Table I and Figure 5) in a manner invalidating the interpretations
detailed above, an examination of the issue is appropriate.
(b) The data obtained were sampled at a constant, arbitrarily selected,
stimulus exposure-time of 100 msec. A more comprehensive sampling
technique, exploring relationships between accuracy and stimulus exposure-
time, might throw more light on the nature of that early processing which
enables the selection of the target. Specifically, the experimental
findings of this study constitute a cross-section through a number of
accuracy/time curves (i.e., variation in exposure-time generates a curve
for each stimulus type). The obtained differences are, therefore,
differences in height at 100 msec, and they could represent many different
things, e.g., they may represent differences in slope or differences in . .
-
29
intercept.
-
30
3. E:
-
items share few physical features with the targets, then it should be a
relatively sim~le task to separate the ta~gets from the irrelevant items.
Once the targets are isolated, and if we accept that further processing
is carried out in a serial fashion (Eriksen and Colegate, 1971; Neisser,
1967), one target Y~i 11 be processed befor-e the other. If we consider
processino capacity to be analagous to a beam of light from a torch
(Eriksen and Hoffwan, 1972b), then it must be "moved" separately from
one target to the other, and if a brief masked stimulus exposure limits
the availability of the material, the total number of targets processed
should decli~e with increasing distance between them. The s;:>ecific aim
of the study to be reported was to examine this prediction.
3.2 Method
The task vms to search for both of two instances of a
single target letter, and an experiment
-
32
and the distance separating the targets (0.67, 1.27, 1.80, 2.20, 2.43,
and 2.53 degrees of visual angle) were varied within subjects.
Subjects were 4 undergraduates (2 males, and 2 females)
with normal 6/6 vision, who received $1 per hour as payment for taking
part in the experiment. Two Ss were given the exposures in ascending
order and the other 2 were given them in descending order. All Ss
completed 2 practice and 9 experimental sessions each of duration
approximately 1 hour. At the completion of each trial they recorded the
positions of the targets, indicated which was the first to be located or
if they were both located at the same time, and rated each on the four
point confidence scale (4 = sure, 3; almost sure, 2 =might have been,
1 =guessing). A target was c::>rrectly located if the correct position or
one either side was marked.
3.2.2 Materials and appar~tus
The apparatus was that used in Experiment 1. The letter
F was selected as the target and G, 0, Q, U and W were selected as the
irrelevant item set. The display items appeared at 12 clock positions
on th~ imaginary circle used in Experiment 1, and there were 36, 12-item
stimulus displays constructed each containing two F's. That is, given
the restriction that over the 36 displays each of the 12 positions was
equally likely to contain an F, a random selection of 6 of the 12
possible target arrangemen~s was made for each separation (see Figure 6).
3.2.3 Procedure
Upon arrival at the first practice session Ss were told
that the task v1as a visual search, that the target \·lotJld appear twice in
all displays, that the cross ¥!as to be fixated before the initiation cf
-
33
G Q
"" Q 0 w
0 0 F Q
u u w u
G f G u w
Q f F G
0
w Q G 0 u F F Q w u
w f G
G u 0 Q u 0 0 f w Q
FIG. 6: Examples of stimulus displays containing tvlo icer.tical targets.
-
the display, that the confidence rat! ngs were to refer to target
presence, and that gu"'sses were to be spread over the avai !able
positions. They we,re also informed that a response, including a
confidence rating, should be made to all targets. They were then
familiarised with the stimulus displays and given 54 practice trL;ls (18
at each exposure). Stimulus exposures of 40, 80 and 160 msec were given
in ascending or descending order during practice, and the order presented
to each subject was maintained throughout the study. The same procedure
11as followed for the second practice session using the experimental
stimulus exposures (20, 40 and 80 msec).
After completion of practice Ss attended 9 experimental
sessions. During ti1e first 8 they completed 6 practice (2 at each
exposure) and 108 experimental trials. The experimental trials were
administered in 3 blocks of 36 tria is, i.e., une block of trials for
each of the 3 exposure-times. During sesolon 9 only 36 experimental
trials were completed (12 at each exposunc.), and this made up a total o"
50 responses for each experimental condition.
3.3 Results
The results in Table 2 and Fi'gure 7 (see Appendix 2,
Table 6 for individual data) showed that the distance separating the
targets did not influence the number ocrrectly reported. A three-Nay
analysis of variance (see Appenrlix 2, Table 7), with the order of
presenting the !itimulus exposures as a between subjects effect, and
stimulus exposure and distance separating the t1·10 identical targets as
within subjects effects, supported this observation. That is,the distance
separating the targets was not a significant variable (F(5, 10) = 2.35,
P > 0.05). On the other hand, the stimulus exposure-time and interaction
-
35
of this variable with the distance separating the targets, both reached
significance (F(2,4) = 40.72, P < 0.01; and.F(10,20) = 6.49, P < 0.001
resrect i ve ly).
TABLE 2
Hean·number of correctly located targets for the three
experinental stimulus exposure-times and the six distances separating
the two targets.
Stimulus Exposure in msec.
Separation (Visual 20 40 80 Angle Deg reEos)
0.67 62.00 70.25 76.50
1.27 54.75 75.00 80.50
1. 80 56.50 73.00 74.25
2.20 54.25 64.25 75.50
2.43 51.00 74.50 74.50
2.53 56.75 72.25 76.00
The significant improvement in performance with
increasing stimulus exposure-ti!Tle (F(2,4) = 40.72, P < 0.01) was of no
direct concern. That is, it was not surprising that increasing the time
12 avai ]able for processing the input improved performance. However, an
explanation of the significant interaction between target separation and
stimulus exposure (F(10,20) = 6.49, P < 0.001) is not obvious from the
data. Observation of Figure 7 shows that performance at a seraration of
0.67° for 20 msec and at 2.20° for 40 msec, does not conform to the
general pattern of re3ults. The data obtained by Eriksen and Hoffman
12. l t should be noted that later studies (in particular Experiment 3) show that increases in stimulus exposure-time are not necessarily related to improvements in performance.
-
100
80
L!J
~ 40 '~ .... z L!J 0 a:: w £).. 20
stimVJius exposure (rn sec:)
80 O.,,.,..D
1·00 2·00 3·00 TARGET SEPARATION (deg, of visual angle)
FIG. 7: The percentage of correctly located targets for the three exposL!re-times and the six distances separating the targets.
-
(1972b) suggest that it is not possible for the focal mechanism to
selectively allocate capacity, if items zre sepa1·ated by less than 1n of
visual
-
4. EXPERIMENT 3
4.1 Introduction
In Experiment 1 the a':curacy of target selection, under
varying conditions of context similarity and number of context items,
was examined at a single arbitrarily selected stimulus exposure-time of
100 msec. It was pointed out that these effects can only be fully
interpreted if we kno\1 the form of the accuracy/time functions under each
condition, and as e consequence a sampling of behaviour over a range of
exposure-time is needed. It was the aim of this study to carry out just
such a samp 1 i ng, but first the .use of accuracy measures rather tha~
reaction times requires some justification.
Previous investigators of selection in vision have
usual!)' manipulated the number of display iter;-,s, 1·1ithout considering
relationships such as the physical similarity betv1een the target and the
irrelevant items, and they have relied to a large extent upon reaction
time as a measure of processing time (e.g., Atkinson, Holmgren, and
Juola, 1969: Estes and Wessel, 1966; Sternberg, 1967). Reaction time
has proved a powerful tool fer investigating information processing
mechanisms, but it suffers from a major disadvantage. When reaction
time is used subjects are given unlimited time and ambisuous instructions
{Edwards, 1961), being requested to perform the task both as quickly and
as accurately as possible. As a consequence they have considerable
freedom to choose a processing strategy and in particular might be
expected to carry out optional processing (usually in the interest of
accuracy) beyond the minimum required for successful completion of the
task. It is obvious that these are not optimal conditions for
-
demonstrating the essential stages in processing. For example, the
model of human information processing 11hich suggests that early
39
processing is carried out on the entire input in parallel (independent
of the nature of the further processing undergone by selected portions
of this input), might imply a high probability of error during the
early parallel stage. In particular, as we have seen, Neisser (1967)
suggested that first-stage processing is prone to error. Yet observed
error rates in reaction time experiments are low. However, given
unlimited processing time the absence of errors is not fatal to the
theory since subjects are given the opportunity to make further
selections when later processing reveals that the Initial selection was
incorrect.
The considerations above suggest that in general,
reaction time studies should be supplemented by accuracy studies, and
that a specific comparison of the two measures would be of interest.
Such a comparison was made in the foll~•ing study. Accuracy and reaction
time measures were compared for a visual search task using stirrulus
displays containing twelve items or six items, and two levels of
physical similarity between the target and the irrelevant item;'.
Subjects were required to call out the position of the target as quickly
as possible, and by exposing the stimuli for pre-set, brief, masked
periods, it was possible to obtain relationships betv1een accurcy (at
positioning the target) and exposure, and reaction time and exposure.
It was prod! cted that at the short exposures subjects accuracy would be
limited by the accuracy of the early stages of processing, and, in
particular their performance would reflect the accuracy of target
selection by these stages.
-
4.2 Method
The task was visual search, and an experimental trial was
essentially the same as described for Experiment 1. When S pressed the
switch a display and a timer were initiated. The display remained
visible for a prespeclfied duration, and its termination was followed by
a 50 msec mask. S was asked to call out the position of the target as
quickly as possible, and his voice stopped the timer allm·ling for the
collection of both accuracy and reaction time measures. Confidence
ratings ware also collected, and a response was considered correct if
the target position or one either side was reported.
4.2. 1 Design
Accuracy and reaction tlme measures were collected for
displays containing 12 items or 6 items, at exposures of 180, 100, 75,
55, 45, 35, and 20 msec when target and irrelevant items were simi Jar in
shape, and 100, 75, 50, 35, 25, 20, and 15 msec when they were
dissimi Jar. Taking the common exposures of 100, 75, .35, and 20 ~sec
enabled the treatment of the data as a three-way design, with similarity
varied between subjects, and number of items and stimulus exposure-time
varied vlithin subjects.
Subjects were 8 undergraduates (6 males and 2 females)
with normal 6/6 vision, and each received $1 per hour for taking part
in the study. All Ss searched for 9, 1-hour sessions (3 practice and 6
experimental), with 4 searching high-similarity and 4 searching 10\v-
similarity dispi
-
41
4.2.2 Materials and apparatus
The materials and apparatus were as for Experiment 1
with the following exceptions: a) the target was again F, but E was
excluded from the high-similarity set of irrelevant items (leaving T, L,
J, P, and H), and C was excluded from the low-similarity set (leaving G,
D, Q, U, and W); b) vocal reaction times were collected by means of a
voice key, connected via a modular digital system to a Fluke Counter-
Timer; and c) stimulus items were placed at 12 equally spaced loci
corresponding to the 12 clock positions. When positioned the minimum
distance separating any two letters was 0.67° and 1.27° on 12-item and
6-item displays respectively.
When constructing stimulus displays the target was
positioned equally often at each clod< position, and 24 cards v1core
constructed for each combination of simililrity and number giving a total
of 96 cards (see Figure B).
4.2.3 Procedure
The nature of the experimental task was explained to S
upon his arrival at the first practice session. During this explanation
it was emphasised that the position of the target should be called out
as quickly as possible, that the confidence ratings were to refer to
target presence not to the accuracy with which it could be positioned,
and that guessing responses should be evenly distributed. He was then
given 112 practice trials with the similarity level relevant to his
group, at exposures of 400, 300, 200, 150, 125. 100 and 50 msec. The
presentation order for the 12-item and 6-item displays was 12, 6, 12, 6,
or 6, 12, 6, 12 (2 Ss in each grcup received one of these orderings),
-
F p
L p
p H T
l T
p J H
J H T
J
Q u G u F G Q
w 0 G
F 0
w G w Q u
FIG. 8: Examples of high-similarity and ]a,,-sir,>ilarity stimulus displays containing 12 items or 6 items.
l
F
0
-
4J
and was held constant throughout the study. The cards were presented in
I; batches of 28 with 4 trials at each of the 7 stimulus exposure-times.
The second practice session followed the same procedure, and the third
'las used to obtain an estimate of the exposure required to achieve an
accuracy of 50%. These estimates served as guidelines for deciding on
the experimental exposure-times.
During the experimental sessions 1, 2, 4, and 5 Ss
completed 14 practice and 112 experimental trials, and during sessiors 3,
and 6 they completed 14 practice and 126 experimental trials. These
trials were administered in the manner of practice sessions a11d 2.
The experimental stimulus exposure-times were used and they provided a
total of 50 responses for each relevant experimental condition.
4. 3 Results
As mentioned in Experiment 1, the experimental procedure
raised the question of what was to constitute a correct response, and to
make a decision the distributions of responses around the target
position were examined. The response frequencies at distances of 2, 3,
13 4, s,· and 6 positions away from the target appeared to be constant (see Appendix 3 Table 8), and smaller than those at a distance of
position away. Consequently responses which indicated the target
position or one position either side were treated as correct, and the
13 To decide if the frequencies at distances not less than 2 posi t.ions away from the target were reflecting a fairly stable chance level of performance, linear regressions were calculated for each of the 28 conditions. Seventeen of the estimates of slope obtained from these regressions failed to reach significance (Appendix 3, Table 8). This supported the observation that errors in positioning the target were (in the maio) confined to one position either side of the correct one, and suggested that guessing responses had been evenly distributed.
-
scores obtained in this manner were corrected for chance 14 (uncorrected
data is shown in Appendix 3, Table 9).
The mean number of correct responses, and the mean vocal
reaction times for non-guessing correct responses, are shown in Table 3
(see Appendix 3, Tables 10, and 11 for individual data). In order to
assess the effects of the experimental variables, three-way analyses of
variance v1ere ccrried out on these data (the levels of exposure-time were
those common to both similarity levels, i.e., 100, 75, 35 and 20 msec).
The accuracy data revealed significant effects for similarity, number,
and exposure-time (F(1,6) = 11.71, P < 0.05; F(1,42) = 91.81, P < 0.01;
and F(3,ll2) = 176.99, P < 0.01 respectively), and significant interactions
of similarity by exposure-time,number by exposure-time, and similarity
by number by exposure-time (F(3,42) ~ 4.36, r < 0.01; F(3,42) = 4.32,
P < 0.01, and F(3,42) • 3.45, P < 0.05 respectively) 15 (see Appendix 3,
Table 12 for summary of analysis of variance). The reaction time data
revealed significant effects for number, and for exposure-time (F(1,6) =
7-57, P < 0.05; and F(,3,18 = 7.14, P < 0.01 respectively) (see Appendix
3, Table 13 for summary of analysis of variance).
14 Contributing to both correct and incorrect responses
-
45
TABLE 3
The average number of correct responses and the average
reaction time for correct non-guessing responses, for the high-similarity
and low-similarity conditions and the number of display items, as a
function of stimulus exposure-time.
12 ITEMS 6 I TEf1S Stimulus Reaction Reaction Exposure Number Time Number Time
(msec) Correct (msec) Correct (msec)
180 27.7 1039 38.3 1045 HIGH-
SIMILARITY 100 25.1 11 71 37.2 1010
75 22.0 991 34.0 866
55 20.3 12.46 30.7 963
45 21.3 112 7 26.3 1080
35 9.7 116 7 25.7 998
20 4.3 1456 3.7 1165
100 43.3 861 47.3 697
75 37.0 888 46.7 663 LOW-
SIMILARITY 50 39.7 942 45.0 663
35 27.9 960 112.0 791
25 21.0 1049 30.7 830
20 5.7 1012 16.0 937
15 0.0 1282 3.0 1043
Since the significant interactions revealed by the acct1racy
data merely reflected the general convergence of the four accuracy/
exposure-time curves (see Figure 9) as performance approached chance
level for brief exposures, they were of relatively minor theoretical
-
interest. Consequently, the main distinction between accuracy and
reaction ti~e measures revealed by the analyses, was the failure of
similarity to have a significant influence on the latter (F(1,6) = 2.43;
P > 0.10). The failure of this difference to reach significance was
particularly surprising in the light of previous findings (Estes, 1972;
Mcintyre, Fox, and Neale, 1970; tleisser, 1967), and hence to draw any
conclusions based on it would be hazardous. However, consideration of
the complete accuracy/exposure and reaction timefexposure relationships,
suggested that accuracy provided a better measure of the time required
for processing, particularly at very brief exposures. That is, fO!-
decreasing exposures of less than 50 msec accuracy decreased very
rapidly, whereas reaction time increased. There is nothing particularly
unusual about increasing reactior. times under the5e experimental
conditions (Teichner and Krebs, l97lJ), but it is difficult to interprzt
them as demonstrating a steady increase in processing time accompanying .a
steady decrease in stimulus exposure-time. More specifically, &nd
assuming that the mask terminated processing, the increasing reaction
times suggest that these measures can contain components thct do not
reflect processing time.
-
100
80
60
1-(.) w 0.:: 0.:: 0 40 (J
w C)
-
48
4.4 Discussion
The discussion will be concerned with interpreting the
experimental results demonstrated by the accuracy/exposure curves in
Figure 9. In particular, why does each of the curves display the marked
deceleration so that In each case improvement in performance beyond the
50 • • d • ff • 1 d 16 d I msec po1nt IS 1 -leU t to etect, an , now may differences in the
level of these "plateaus" be explained. However, it should first be
noted that the data of Experiment i sectioned these curves at the 100
msec point, and therefore, represent differences in the plateau levels.
4. 4.1 A two-stage model
The four accuracy/exposure curves (Figure 9) can be
explained by Neisser's (1967) tvJO-stage model. We need only suppose
that within the experimental range of exposure-time there is on the
average insufficient time for further selective processing If the first
item constructed Is not the target. Then If there is only sufficient
time for the construction of one item, the accuracy/time curves represent
fairly directly, the relationships between the.probablllty of a correct
selection on the first item processed and stimulus exposure-time. This
allows a straight-forward explanation (along the same 1 ines as for
Experiment 1) of the experimental findings. The significant effects of
increasing, a) the number of irrelevant items, and b) the similarity
between the target and these items, can be easily explained in terms of
16 Observations of accuracy/time curves in the present study (Figure 9) suggests a very sla• Improvement in performance for ex~osure-tlmes greater than 50 msec. However, separate analyses (see Appendix 3, Table 15) carried on the hlgh··simllarlty data at 55, 75, 100, and 180 msec, and on the la1-s lmi lari ty data at 50, 75, and 100 msec failed.to reveal any significant effect of exposure (F(3,6) = 4.07, P > 0.05 and F(2,4) = 4.72, P > 0.05 respectively).
-
increased difficulty at the feature processing stage .of selection. When
the number of stimulus Items Is irrcreased the task of selecting the
correct item (on the basis of its features) becomes more difficult, i.e.,
the probability of the selected item being the target decreases, or
conversely, the chances of an irrelevant item being incorrectly selected
are increased. This is because early processing is elementary and prone
to error, and as a consequence an incrense in the number of items
processed increases the amount of feature material and the 1 ikelihood that
an incorrect item will be selected. Changes in the similarity between the
target and the irrelevant items also influences the accuracy of the
feature processing stage of selection, with errors increasing as irrelevant
items become more like the target since it becomes more,difficult to select
the correct i tern on the basis of its features.
The shapes cf the curves, with the rapid improvement in
performance prior to 50 msec and post-50 msec stability, can be accounted
for by considering the processing tires required by the two stages of
analysis. The period up to ~nd including the rapid initial improvements
covers the times necessary to carry out an initial selection, i.e., ir-. the
main the processing required is coro;lleted in less than 50 msec. Selected
' items are constructed from the icon one after the other (Eriksen and
Co legate, 1971; Neisser, 1967), and since it might be expected that the
times necessary for this detal led analysis are at least as long as those
required for icon information, very few I terns can be constructed in less
than 200 rnsec. Therefore, the level of performance reached when
information is only available for a brief period, is largely dependent on
the probability of correct initial selection from the icon, and will shO'.v
very little improve.ment after the icon is complete •.
-
50
!;,4.2 Parallel-serial ~rocessing
A mode of processing interpretation of the accuracy/time
curves faces the problem discussed in Experiment 1. Namely, it was not
the intention of the study to distinguish between parallel and serial
models, and the data certainly are not capable of doing so. But, even
in the light of this difficulty strictly parallel and strictly serial
models do not account easily for the data. That is, as available
processing time increases, in the absence, of memory overload, the most
obvious prediction from these models would be an improvement in
performance, and yet observed performance fails to show significant
improvement after 50 msec. It may be possible to develop il parallel-
serial account of the data. In the present context, h01•ever, such an
account is relevant only if it ,can explain (along with an account of the
plateau effect) the differences between curves 1-1i thout asst•>ni ng that
they reflect different rates of error during selection. The author is
not familiar with a model of this kind, but it is possible th~t the
differences between levels may be due to mlsperceptions of irrelevant
items as targets or viee ver-sa. (The question of incorract selection or
mlsperception is discussed further in Experiment 6.)
4.4.3 Response speeding
When instructing the experimental subjects a speeded
response was requested. It is therefore possible that the fai iure of
performance to Improve sign! ficantly beyond 50 msec, IYas due to subjects
adopting a strategy which traded accuracy for· speed, i.e., as the
exposure-time Increased, instead of \'Jaitin~ for the additional lnform'!ltion
they hurried their response and accuracy suffered accordingly. However,
this does not affc:ct the previously detailed interpretation of the
differences in plateau level as reflecting the probability of correct
-
51
initial selection.
4.4.4 Masking
As previously mentio~ed (Footnote 7), whether a backward
masking procedure prevents further processing of the stimulus input due
to a perceptual integration of the contours of stimulus and mask, or
because the arrival of the mask interrupts the processing of the stimulus,
remains an issL•e. But, if interruption is accepted as the cause, then
the visual system must register stimulus and mask as separate events, and
in general this does not occur at very brief separations. Therefore,
masking in the present study probably occurred because it was impossible
to extricate the contours of the stimulus from the mask at very brief
separations, but was caused by the interruption of processing as the
separation increased (see Turvey, 1973). This could account for t'Je
first leg of the curves by Interpreting the period of rapid Improvement
as that required for the separation of stimulus and mask. But, it cioes
not explain the plateau effect which can still be interpreted In terms
of a tv10-s tage mode 1.
4.4.5 Memory span limitations
Superficially similar negatively accelerated curves were
obtained by Sperling (1963, Figure 5), and he attributed the failure of
performance to improve after an exposure of 50 msec to l imitations in an
immediate-visual-memory. Sperling used a recognition task which required
subjects to report as many items as possible, and vs a consequence, memory
span was exceeded after 4 or 5 items had been p1·ocessed. However,
performance in the present study would not be restricted by a limited
memory span since subjects are not required to remember irrelevant ite~s.
-
52
4.5 Further investigation
There are two aspects of the accuracy/time curves
(Figure 9) which allow for further investigation. Firstly, the failL1re
to detect a significant improvement in performance between exposure-times
of approximately. 50 and 180 msec, suggests than an investigation of the
forms of the curves beyond 200 msec may be of interest, and, secondly,
if the period during \vhich rapid improvement in performance occurs, is
in fact determined by the time necessary to complete selection from the
icon, then it v1ould be of some interest to know if these times are
independent of the difficulty involved in distinguishing the target from
the irrelevent items.
-
53
5. EXPERIMENT 4
5. 1 Introduction
FolJm,ling the suggestions for further investigation made
in the discussion of Experiment 3, the aim of this study \vas to extend the
accuracy/time curves (Figure 9) by sampling search oehaviour over a range
of 100 to 500 msec.
The previously obtained relationships between accuracy of
target selection and stimulus exposure-time (Experiment 3, Figure 9),
shmved that as exposure-time increased performance irP.proved rapidly at
first, but then stabilised. It was argued that Neisser's (1967) tHo-stage
model of analysis gave a plausible account of these results, i.e., rapid
improvement occ,urred when sufficient time VIaS provided to enable the
comp 1 et ion of firs t•·s tilge precessing, and performance stab i l i sed wh i l e a
selected item underwent detailed analysis. This model 1vas used by
Neisser to explain the data obtained when subjects searched for targets
embedded in lists (see the previous discussion of visual search on
pages 3 to 6 ), and it was assumed that first-stage" processing •1as
repeated until the target was found. The extension of exposure-time
beyond 200 msec was intended to bridge the gap ben,een the time-] imi ted
(brief exposures) and list conditions, and it was hoped that
commencement of a second processing cycle •1ould be demonstrated as a rapid
Improvement in performance after the previously obtained period of
stab i 1 i ty.
-
5.2 Method
The task was again visual search and an experimental trial
was the same as previously described in Experi~ent 3 apart from the
fo 1 lowing: a) no reaction time measures were co I lected and Ss responses
were not speeded, b) subjects .vere not required to give confidence
ratings, and c) for stimulus exposure-times greater than 180 rr1ay des lgn.
Subjects were 8 undergraduates (6 rna les and 2 females)
with normal 6!6 vision. They undertook the study partly on a voluntary
basis and partly for course credit. All Ss searched for 7, 1-hour
sessions (1 practice 18 and 6 experimental), with 4 searching high-similarity
displays and 4 searching low-similarity displays.
17 Maintaining the physical presence of the stimulus for a maximum of 180 msec, ensured that the input into the processing system .vas from a single eye fixation.
18 \-lhen carrying out this study (and .vhen running Experiment 5) funds were not avai ]able for the paynent of subjects. Consequently, they were somewhat reluctant volunteers and practice, unfortunately, .vas more limited than .vas considered desirable.
-
55
5.2.2 Materials and apparatus
The materials and apparatus (minus a counter-timer) were
essentially the same as those used in Experiment 3.
5.2.3 Procedure
The nature of the experimental task was explained to S
upon his arrival at the practice session. He was told to mark the
location of the target on the sheet provided and to distribute his guesses
over the avai ]able positions, and he was then given 112 practice trials
with the similarity level relevant to his group. The practice trials
were administered at the experimental exposure-times with 7 trials at each
(i.e., 56 trials for both 12-item and 6-item displays). The presentation
order for the 12-item and 6-item displays was 12, 6 or 6, 12 (2 Ss in
each group received one of these orderings), and was held constant
throughout the study.
During experimental sessions i, 2, 3, and 4 Ss
completed 16 practice and 128 experimental trials, and during sessions 5
and 6