Journal of Experimental Psychology: Human …...attention can be either central or peripheral...

Journal of Experimental Psychology:Human Perception and Performance

Copyright © 1984 by the American Psychological Association

VOL. 10, No. 5 OCTOBER 1984

Abrupt Visual Onsets and Selective Attention:Evidence From Visual Search

Steven Yantis and John JonidesUniversity of Michigan

The effect of temporal discontinuity on visual search was assessed by presentinga display in which one item had an abrupt onset, while other items wereintroduced by gradually removing line segments that camouflaged them. Wehypothesized that an abrupt onset in a visual display would capture visualattention, giving this item a processing advantage over items lacking an abruptleading edge. This prediction was confirmed in Experiment 1. We designed asecond experiment to ensure that this finding was due to attentional factors ratherthan to sensory or perceptual ones. Experiment 3 replicated Experiment 1 anddemonstrated that the procedure used to avoid abrupt onset—camouflage re-moval—did not require a gradual waveform. Implications of these findings fortheories of attention are discussed.

The human visual system is exquisitely apparent to anyone viewing a complex scenesensitive to relative movement and to flicker, with a moving object in it: Until the move-The generality of this observation is readily ment begins, the object can seem completely

invisible; at movement onset, the object'slocation is immediately and compellingly

Experiment 1 was reported at the 23rd Annual Meeting manifest almost without effort on the partof the Psychonomic Society, Minneapolis, November -,. ' _, ,. „„ , . ,. , .19g2. of the observer. The salience of flashing lights

This research was funded by Grant 82-0297 from the on ambulances and radio towers is similarlyAir Force Office of Scientific Research to John Jonides obvious. We are concerned in this articleand by a National Science Foundation Graduate Fellow- wjth determining what role, if any, spatialShiw™.diiS thank Richard Abrams for designing "elective attention plays in the detection ofthe apparatus used in Experiment 3 and Joanne Alexander such stimuli. In particular, we will test afor skillfully and meticulously running the experiments, hypothesis in the context of visual searchRichard Abrams and Peter Gordon made important which holds that isolated abrupt Stimulussuggestions during the course of the research for which t id d involuntary deploy.we are grateful. We also thank J. E. Keith Smith for help . . • , , f^, .. ,with modeling and Richard Abrams, William Banks, ment of attention to the locus of the temporalCarol Krumhansl, David Meyer, John Palmer, and two discontinuity,anonymous reviewers for comments on an earlier draftof the paper. Visual Search and Attention

Requests for reprints should be sent to John Jonides c . „_ . «7_.,.,_i /. nne.\ ~~A A tu;ncn«at the Human Performance Center, Department of Psy- „ ̂ stes and Wessel (1966) and Atkinson,chology, University of Michigan, 330 Packard Road, Ann Holmgren, and Juola (1969) were among theArbor, Michigan 48104. first to show that the time required to detect

Copyright 1984 by the American Psychological Association, Inc.

601

602 STEVEN YANTIS AND JOHN JONIDES

a prespecified alphanumeric character amongdistractors increases nearly linearly as thenumber of distractors increases. This displaysize effect is a very robust finding that holdsunder a variety of experimental conditions.The implication of the display size effect isclear: When there is uncertainty about thelocation of the target in a visual display, asearch must be conducted over locations oc-cupied by potential targets. Furthermore, thesearch is constrained in its speed; severalitems cannot be evaluated as quickly as one.Instead, an increase in the number of poten-tial targets taxes the capacity of the searchmechanism to detect the target.

This pattern of results may be contrastedwith those obtained in a visual search taskin which subjects are informed about thetarget's location in advance of its presentation.By now there are many demonstrations ofsubjects' ability to allocate attention to aparticular location while they maintain fixa-tion. Investigators have commonly demon-strated this by using a precue to specify alocation or set of locations that will containinformation relevant to the task at hand.This method was used by Sperling (1960) inhis classic study of visual persistence and hassince been used by others to study selectivespatial attention more directly (e.g., Colgate,Hoffman, & Eriksen, 1973; Engel, 1971; Er-iksen & Hoffman, 1972, 1973; Jonides, 1980,1981; Posner, 1980; Posner, Nissen, & Ogden,1978). In these experiments, subjects havetypically been asked to determine whether aprespecified target is present in a multielementdisplay. When no locational precue is pro-vided (Atkinson et al, 1969), reaction timesincrease roughly linearly with the number ofelements in the display. When a bar markerreliably indicating the location of an upcom-ing target is provided (Holmgren, 1974), bothoverall reaction time (RT) and the slope ofthe regression line relating display size andlatency are smaller than without a cue. Similarresults are seen in the accuracy of re-sponses under impoverished viewing condi-tions (Bashinski & Bacharach, 1980; Van derHeijden & Eerland, 1973). Furthermore, Shaw(1978) has shown that observers can takestimulus probabilities into account as theyoptimally allocate fixed attentional resourcesto multiple spatial locations.

These findings have generally been inter-preted as showing that subjects can directtheir attention to spatial positions in prepa-ration for a display and can thereby reduceprocessing time by not having to localizeprobable targets as they do in an uncuedsituation. Put another way, the cue allowssubjects to allocate scarce resources to aparticular spatial channel in advance, result-ing in rapid processing of whatever signalsubsequently occupies that channel. Thisconclusion is compelling given that subjectscannot take advantage of cues that specifytwo noncontiguous spatial positions (Engel,1971; Hoffman & Nelson, 1981; Posner, Sny-der, & Davidson, 1980; but see Shaw, 1978).

The locus of control of shifts of visualattention can be either central or peripheral(Jonides, 1981; Posner, 1980), That is, a shiftof attention may occur either because anobserver chooses to execute one or becausethe occurrence of a peripheral visual eventmay automatically capture attention. Thisdistinction was made by Jonides (1981) in aseries of experiments exploring the effects ofcentral and peripheral visual cues. He showedthat abrupt peripheral locational cues, ascompared with foveal ones, produced bothfaster reaction times and a greater differencebetween the benefits of a valid cue and thecosts of an invalid one. From this result andothers, Jonides concluded that shifts of visualattention elicited by peripheral stimuli areimperative and automatic, whereas those pro-duced by central cues are less efficient andsomehow less obligatory.

The identification of a mechanism thatautomatically directs attention to salient visualevents in the periphery is consistent with thefinding that simple stimulus features canoften be used as the basis for item selection.When items are processed this way, subjects'search behavior mimics their behavior whenthey are cued about the location of an im-pending target. For example, Treisman andGelade (1980) demonstrated that targets dif-fering from hontargets in a simple way (e.g.,Xs among Os) are easily detected, regardlessof the number of distractor elements. Onlywhen the target is defined by a conjunctionof features (e.g., a red X among red Os andgreen A's) must focal attention be applied toitems serially. Similarly, Egeth, Jonides, and

ABRUPT ONSETS AND ATTENTION 603

Wall (1972) and Jonides and Gleitman (1972)showed that the detection of digits amongletters could take place in parallel across theentire visual field, regardless of the numberof distractor letters. In fact, it is possible withpractice to develop new target categories thatallow rapid, automatic detection (Rabbitt,1967; Shiffrin & Schneider, 1977).

Results of this sort can be accommodatedby a theory in which certain stimulus features(e.g., color) could be used as a basis forpassing stimuli into a limited capacity systemfor further processing (Duncan, 1980). Eachstimulus can be interviewed for features rel-evant to the task at hand, and stimuli pos-sessing the appropriate features are immedi-ately admitted to the limited capacity system.

In the experiments reported above, weexplored a stimulus property—abruptness ofonset—that may be used as a basis for adrmission into the limited-capacity system. Thechoice of this property is not arbitrary. If weconsider some of the commonplace examplesin which visual search seems intuitively tobe directed by distal stimulus features, theubiquity of sudden stimulus onset becomesapparent. Cases in which flashing lights cap-ture our attention illustrate this point nicely.Furthermore, there is emerging electrophysi-ological and psychophysical evidence thatabruptness of onset is a stimulus property towhich the visual system is particularly sensi-tive. Let us briefly review some of this evi-dence before we describe our experiments.

Abrupt Onset in Visual Processing

Several investigators have explored the ef-fect of temporal waveform on the detectionof visual stimuli. This has been motivated inpart by the electrophysiological finding thatcells in the visual system are differentiallysensitive to abrupt and sustained visual events(e.g., Cleland, Levick, & Sanderson, 1973),and the finding in psychophysics of visualchannels or mechanisms that are selectivelysensitive to specific temporal and spatial fre-quencies (e.g., Kulikowski & Tolhurst, 1973).

Electrophysiology

In 1966, Enroth-Cugell and Robson iden-tified two distinct classes of ganglion cells in

the cat's retina; the firing frequency of oneclass (denoted Y or transient cells) was en-hanced at the abrupt onset and offset of astimulus and only then; another class of cells(called X or sustained cells) exhibited contin-uously enhanced firing throughout the dura-tion of the stimulus.

Subsequently, several other investigatorscorroborated and expanded on this finding.The two types of cells, it emerged, differ inseveral respects. Transient cell receptive fieldshave larger surrounds (Cleland et al., 1973)and are more sensitive to rapid flicker ormotion (Fukada & Saito, 1971) than sustainedcells are. Furthermore, the response latencyof transient cells is shorter and their axonalpropagation velocity is higher than that ofsustained cells (Cleland et al., 1973; Fukada,1973). Finally, Fukada and Stone (1974)found that the concentration of sustained-cell receptive fields is highest in the fovea,whereas transient receptive fields are moreevenly distributed about the retina.

Psychophysics

Other investigators have found correspon-dences to the electrophysiological findingsdescribed above using psychophysical proce-dures and human subjects. Kulikowski andTolhurst (1973) distinguished two visualthresholds for temporally modulated sinusoi-dal gratings: the contrast at which flickercould be detected and the contrast at whichspatial structure could be identified. Flicker-detection sensitivity declined dramatically athigh spatial frequencies and low temporalmodulation frequencies. However, at low spa-tial frequencies, sensitivity to spatial structuredeclined, whereas changes in temporal fre-quency had little effect. Kulikowski and Tol-hurst concluded that there are (at least) twoindependent mechanisms in the visual system,one sensitive to rapid flicker and low spatialfrequency and the other sensitive to visualdetail and form and high spatial frequency.

Subsequently, Tolhurst (1975) investigatedthe onset and offset characteristics of sustainedand transient channels. In this experiment,sensitivity to a sinusoidal grating was mea-sured when it was flashed for 4 ms at varioustimes before, during, and after the presenta-tion of a subthreshold 800-ms grating of the


same spatial frequency and phase. Tolhurstdiscovered that when the gratings were of lowspatial frequency (2 cycles/deg), test flashsensitivity increased briefly at the onset anddecreased briefly at the offset of the longflash. However, at a higher spatial frequency(7.6 cycles/deg), the sensitivity to the testflash increased to a steady level for the du-ration of the long flash.

Another demonstration of the importanceof abrupt onset is a study by Breitmeyer andJulesz (1975) in which contrast sensitivity tosinusoidal gratings was assessed under eitherabrupt or gradual onset and offset. The grad-ual onsets were 200-ms linear ramps, and allgratings had the same time-averaged energy.Breitmeyer and Julesz found that sensitivitywas enhanced at spatial frequencies less than5 cycles/deg when the onset was abrupt ascompared to when it was gradual; there waslittle effect of offset waveform.

Wilson (1978) also obtained contrast sen-sitivity functions for gratings of various spatialfrequencies under two types of temporalmodulation. Mean intensity of the gratingsfollowed either a Gaussian function with atime constant of 250 ms (S-modulation), orone 125-ms cycle of a square wave (T-mod-ulation). In agreement with the findings ofBreitmeyer and Julesz (1975), T-modulationincreased sensitivity at low spatial frequenciesand reduced sensitivity at high spatial fre-quencies as compared to S-modulation. Theseexamples suggest that abrupt onset has aspecial status in the visual system.

Reaction Time and Detection

This conclusion is supported by other ex-periments that explored the effects of abruptchange on the processing of visual stimuli.Todd and Van Gelder (1979) measured re-sponse latencies to suprathreshold stimulihaving abrupt onsets or not. Rather thanusing a gradual onset contrasted with anabrupt one, as in the Breitmeyer and Julesz(1975) or Wilson (1978) studies, Todd andVan Gelder employed what they called theno-onset presentation procedure as a control.With this technique, stimuli are present beforetrial onset, but are completely camouflagedby irrelevant stimulus elements. The irrelevantelements are then removed at trial onset,

revealing the unchanging display ,items. Thiscondition was contrasted with a standardabrupt onset procedure. In one experiment,eight asterisks were arrayed horizontally be-fore the subject. At trial onset, seven of themdisappeared, and subjects were required toexecute a saccadic eye movement to theremaining target. Eye movement latency inthis no-onset condition was compared to thatin the onset condition in which a single targetappeared abruptly in what had been an emptydisplay field. Subjects were uniformly fasterin the onset condition; this difference in-creased as the complexity of the decision tobe made about the stimulus increased fromone of detection and localization to one ofcategorization.

Following Todd and Van Gelder (1979),Krumhansl (1982) reported a study using aprocedure similar to the onset-no-onset par-adigm. A linear array of characteristicswas presented in a prestimulus field (e.g.,|+x+X'+x+x|; • = fixation point). At trialonset, all but one location was extinguished;the remaining character was either identi-cal with the character in that position inthe prestimulus field, (no form change:I + • |) or was different (form change:| x • |). Pattern masks then terminatedthe display. Krumhansl found superior local-ization and identification accuracy in theform change than in the no-form changecondition. Although this finding is consistentwith the hypothesis that a form change rep-resents an abrupt onset (or apparent move-ment) at the target location and hence mayactivate transient mechanisms, thereby sum-moning attention, Krumhansl derived aquantitative information-processing modelthat does not invoke transient mechanismsor attention (although she points out that herdata do not rule out such mechanisms). Themodel assumes the initiation of a brief phaseof rapid encoding with stimulus onset, fol-lowed by a period of relatively less rapidprocessing. The rapid phase occurs duringthe prestimulus array in both conditions;however, the form change condition renewsthe rapid phase at trial onset, whereas theno-form change does not, resulting in anadvantage for the form change condition. Wewill return to this enhanced encoding modelin the General Discussion.


Conclusion

All of the experiments described previouslyhave consistently demonstrated a processingadvantage conferred by a temporally abruptleading edge. This has at times been explicitlyequated with the recruitment of visual atten-tion. Breitmeyer and Ganz (1976) assertedthat the transient system responds to abruptchanges in the periphery, such as movementor flicker, and thus composes "part of an'early warning system' that orients an organ-ism and directs its attention to locations invisual space that potentially contain novelpattern information" (p. 31). Similar sugges-tions have been made by Ikeda and Wright(1972, pp. 796-798) and by Broadbent (1982),who stated, "It seems plausible that a suddenstimulus onset acts to increase intake [ofinformation] from that sensory region and todecrease it from elsewhere . . . ." (p. 284)Indeed, all of these writers were anticipatedby Tichener (1908), who said,

sudden stimuli and sudden changes of stimulus exert afamiliar influence upon attention [p. 192] . . . suddenstimuli impinge upon nervous elements that have hithertobeen free from stimulation of their particular kind, i.e.,upon nervous elements of a high degree of excitability;and it is probable that the excitations which they set upsurfer less dispersion and diffusion, within the nervoussystem, than the excitations resulting from gradual ap-plication of stimulus, (pp. 204-205)

None of the studies reported above, how-ever, directly assess the extent to which anabrupt onset may capture attention. It is notenough to compare a single abrupt onsetwith multiple onsets; we must establish thatit is the presence of a single abrupt onsetitself that captures attention in the presenceof other (nonabrupt) irrelevant items. Oneway to conceptualize this capture is in termsof the monitoring of several visual channels.A channel in this context refers to a groupof receptive fields in the same retinal area orto a group of noncontiguous receptive fieldsall sensitive to the same spatial frequency(Graham, 1981). If the observer is limited inthe number of channels that can be monitoredsimultaneously, then an allocation model likethat described by Shaw (1978) may reasonablyaccount for the performance of the system.It is not difficult to imagine a model in whichthe activation of a so-called transient channel

causes a rapid shift of resources to the taskof monitoring events on the activated channel.

Two patterns emerge, then: (a) Abruptonset of an object tends to confer an advantagein visual processing, and (b) theories of atten-tion often assume that, although processingresources are scarce, the assignment of atten-tion by an efficient allocation schedule (e.g.,on the basis of a preparatory cue or .salientstimulus feature) can overcome these limita-tions, resulting in quite efficient performance.How are these patterns related?

We suggest a simple hypothesis. Abruptonset is a property to which the visual systemis particularly sensitive. According to thehypothesis, attention is rapidly and involun-tarily applied to visual stimuli (or channelscontaining stimuli) possessing this property(i.e., abrupt onset) when no other such stimuliare present. The first experiment tests oneaspect of this hypothesis, namely the rapiditywith which attentional resources are broughtto bear upon the relevant stimuli.

Experiment 1

We employed a standard visual search taskin which the targets and nontargets were notdistinguishable on the basis of simple physicalfeatures (e.g., color; cf. Treisman & Gelade,1980) or categorical status (e.g., alphanumericclass; cf. Egeth et al., 1972). Furthermore,the stimuli served as targets and as distractorsequally often, preventing practice with a con-sistent set of target items. Thus the mappingof stimuli to responses was varied (Schneider& Shiffrin, 1977).

Display size was two or four items. Wechose only two display sizes because we wereinterested in the relative magnitudes of thedisplay size effect under various conditions,and not in the linearity of the display sizefunctions. This goal dictates collection ofdata only at the endpoints of the display sizefunctions.

We used a version of Todd and Van Gelder's(1979, Experiment 5) no-onset procedure topresent items lacking an abrupt onset. Wedid this because of a fundamental problemassociated with using gradual onset as insome previous studies (e.g., Breitmeyer &Julesz, 1975; Wilson, 1978): Our dependentvariable is response latency, and the temporal


asynchrony inherent in using gradual versusabrupt onset would make interpretation oflatencies difficult at best. The no-onset pro-cedure circumvents this problem and at thesame time minimizes (indeed, eliminates)change in the nonabrupt items.

The variable of interest here is onset type.On every trial, exactly one item had anabrupt onset, and the remaining items werepresented via the modified no-onset procedure(Todd & Van Gelder, 1979) in which cam-ouflaging premasks were removed gradually.The use of gradual camouflage removal wasintended to reduce the possibility of atten-tional capture by abrupt offsets. On sometrials, the target was the abrupt onset item;on other trials, the target was revealed grad-ually; on still others, of course, the target wasabsent altogether.

If the aforementioned hypothesis is correct,then the abrupt onset item, whether or not itis the target, is identified first on every trial.If on a particular trial the abrupt onset itemis the target, scanning halts and a positiveresponse is emitted. Otherwise, scanning con-tinues in a serial, self-terminating manner orin an equivalent parallel limited capacitysearch (cf. Atkinson et al., 1969; Townsend,1972). Consequently, the hypothesis predictslittle increase in latency with display sizewhen the target is presented abruptly, but theusual increase of about 40 ms/comparisonwhen the target is of the no-onset type or isabsent.

Method

Subjects. Eighteen University of Michigan undergrad-uates were paid $7 for participation in two 50-minsessions.

Apparatus and stimuli. A Digital Equipment Corpo-ration (DEC) PDF-11/60 computer controlled stimuluspresentation and response collection. Stimuli were dis-played on the face of a DEC VT-11 graphics displaydevice (P4 phosphor), and responses were made on aHewlett-Packard Model 2621A terminal keyboard. Sub-jects were seated in a dimly lit sound-attenuating boothwith an illuminance at the display screen of 25 Ix.

The stimuli Consisted of the letters E, H, P, S, and U,constructed by illuminating the appropriate five segmentsof a seven-segment character (see Figure 1 for examples).Letters subtended a visual angle of 1 ° in width and 1,9°in height from a viewing distance of 45 cm. They weresituated 5.7° from fixation and occupied various subsetsof the vertices of an imaginary hexagon with sides of5.7°. This spatial separation is well outside the range oflateral interference typically found in masking studies

(Bouma, 1978; Wolford & Chambers, 1983) as well asoutside the estimated 1 ° focus of attention (Eriksen &Hoffman, 1972).

The luminance of a test patch with approximately thesame number of points as a letter was 11.1 cd/m2, andthe luminance of the blank screen was 0.7 cd/m2.

Procedure. Trial events, illustrated in Figure 1, wereas follows. Each trial began with a 1,000-ms intertrialinterval. The target item for the trial was shown for1,000 ms in the topmost display location and thenextinguished. A central dot then appeared, upon whichsubjects remained fixated throughout the trial. A 500-ms pause ensued, followed by the presentation of threefigure-8 premasks arranged in an upward-pointing equi-lateral triangle. After 1,000 ms, irrelevant display segmentsbegan to fade in a series of four luminance decrements.'The offset lasted 80 ms. During offset, all remainingsegments stayed at a constant intensity. When displaysize was'four, all three premasks changed gradually toletters; when it was two, one changed to a letter and theother two faded to blanks.

At the end of camouflage offset, one letter was abruptlyilluminated in what had been a blank display location.Subjects were told to scan the resulting display to deter-mine whether the prespecified target was present. If itwas indeed present, a prespecified key was pressed withthe right index finger; if the target was absent, anotherkey was pressed with the left index finger. Several typesof errors were possible, including an incorrect response,a failure to respond within 2,000 ms, or the depressionof an irrelevant key. When an error occurred, a singlebeep was emitted by the terminal as feedback. Thedisplay was then extinguished, and the next trial began1,000 ms later.

Reaction time was measured from the onset of theabrupt stimulus, that is, from the end of camouflageremoval. Thus, in principle, subjects had up to 80 ms of"free viewing" of the no-onset stimuli as they wererevealed over time. This timing of events was requiredto ensure that any advantage observed for onset itemscould be attributed only to attentional factors and not toperceptual ones. For instance, if the onset item appearedat the beginning of the gradual offset, an abrupt onsetadvantage could be interpreted as reflecting contrastdifferences rather than attentional capture.

Subjects were instructed to respond as quickly aspossible while keeping errors to a minimum. They alsowere advised to maintain fixation at center.

Design. The two main variables, trial type and displaysize, were completely crossed. The three trial types were

1 So as to minimize the possible attention-capturingeffects of abrupt offset, the camouflage was removedgradually, within the limits of the available equipment.This was accomplished by reducing the luminance of theirrelevant segments in the figure-8 premasks in fourdiscrete luminance decrements situated 27 ms apart.Thus the last step, which completely extinguished thecamouflage, occurred 81 ms after the first one. Thislooked like a constant, gradual fading. The objectionmay be raised that this staircase comprises several abruptchanges that could summon attention. This possibility isevaluated in Experiment 3.


Duration ONSET NO-ONSET

1000

1000

80

RT

B -•

B B

B•

B B

U p

*

5 - E

P

B B

B•

B B

P E

Figure 1. Sequence of events for a typical trial (displaysize = 4) in Experiment 1. (Dashes repre'sent fading linesegments. Duration in ms. The correct response for bothtrial types here is yes. Under display size = 2, not shown,two of the figure eights would fade to blanks.)

target present with an abrupt onset (or, for brevity, onset),target present with gradual camouflage removal (or no-onset) and target absent (or negative). Half the trials ineach session required a positive response, and the re-maining ones a negative response.

Display sizes of two and four occurred under each ofthe three trial type conditions. Exactly one onset itemappeared in every test display, and the remaining items

Table 2Mean Error Rates (%) For Each Condition andDisplay Size (2 and 4) in Experiment 1

Onset No-onset Negative

Day

12

3.63.2

4.93.1

7.95.4

7.36.2

3.43.2

3.43.0

were always no-onsets. When display size = 2, half thepositive trials were onset and half no-onset. When displaysize = 4, however, only one. quarter of the positive trialswere onset, and the remaining were no-onset. Theserelative frequencies were chosen so that the occurrenceof an abrupt onset was not itself correlated with theoccurrence of a target; we thereby precluded the possiblestrategy of intentionally using the abrupt onset as aselection cue.2 For clarity, the number of trials per blockin each of the six conditions is shown in Table 1.

Each of the five letters served as target equally oftenin each condition and as the onset stimulus approximatelyequally often. The one onset item that appeared on eachtrial occupied each of the three onset display locationsequally often. No letter was repeated in a display. Withinthese constraints, the trials were ordered randomly within80-trial blocks.

Each subject completed four 80-trial blocks on eachof two days, for a total of 640 observations per subject.Short rest periods were taken after every block.

Results and Discussion

The mean error rate across subjects andconditions was 4.4%. Error rates for eachcondition are shown in Table 2. In general,error rates were lower on Day 2 than on Day1, were slightly higher when set size = 4 thanwhen set size = 2, and were greater when thetarget was no-onset than when it was onset.These error rates iare positively correlatedwith reaction time (r = .23); a speed-accuracytrade-off was not evident.

An analysis of variance was conducted onthe mean reaction time data, the factors ofwhich were days (1 and 2), trial type (onset,

Table 1Number of Trials per Block in Each Conditionof Experiment 1

Displaysize

24

Onset

105

Trial type

No-onset

1015

Negative

20 .20

2 Note that (when display size = 4) this results in thetarget appearing in any one onset location only onetwelfth of the time, but in any single no-onset locationone fourth of the time. Were subjects aware of thiscontingency, they would be expected to optimally allocateattention toward the high-frequency (no-onset) locationsand away from the low-frequency (onset) locations (Shaw,1978), thereby reducing the probability of producing thepredicted result.


650-1

600-

0>

155°V

o'•§ 500O&

450-

400

-• observed-o predicted ,g negative

positive no-onset

positive onset

Display Size

Figure 2. Mean reaction time on Day 2 as a function of display size. (Curve parameter is trial type. Thenumber of trials per point is proportional to the entries in Table 1. See text for explanation of predictedpoints.)

no-onset, and negative), and display size (twoand four). Large main effects of all the factorswere found: For days, P(l, 17) = 179, MSe =4,442; for trial type, F(2, 34) = 109, MSe =1,840; and for display size, F(l, 17) = 169,MSe = 726; all ps < .001. No interactioninvolving day was significant (all Fs < I).Finally, the interaction between trial type anddisplay size was highly Significant, F(2, 34) =22.6, MSe = 508, p < .001.3

Given the lack of interaction involving day,we shall focus on data from the secondsession only, because they reflect more highlypracticed performance and are therefore lessvariable. Mean Day 2 reaction times acrossall subjects are shown as the observed datapoints in Figure 2. An inspection of the figureshows that the increase in mean RT withincreasing display size was not large whenthe target had an abrupt onset: The increment

in RT (or slope) was 7.9 ms/item. The onsetslope contains zero in its 95% confidenceinterval, which was 9.04. The slope was largerwhen the target was no-onset (slope = 24.5ms/item) and larger still when the target wasabsent (slope = 35.0 ms/item).4 The TrialType X Display Size interaction is readily ap-parent here.

3 Four of the subjects had exceedingly large error rates(greater than 10% in at least one condition). Becausethese subjects may have exhibited a speed-accuracytrade-off, the ANOVA was rerun excluding them. Thepattern of results was unchanged: All the main effectswere highly significant, and the only significant interactionwas Trial Type X Set Size, F(2, 26) = 22.15, MS, = 469,p< .001.

4 Excluding the four error-prone subjects (see Footnote3) gave slopes of 5.6, 23.3, and 36.4 ms/item for theonset, no-onset, and negative conditions, respectively.


Bonferroni multiple contrasts were con-ducted on the Day 2 data to determine thesources of the significant interaction. Thedifferences between the slopes of each pair oftrial types were all significant at the .05 level(the difference between the onset slope andthe no-onset slope was 16.6; between theonset and negative slopes was 27.1; and be-tween the no-onset and negative slopes was10.5; a 95% simultaneous confidence intervalis 9.04).

Quantitative model. We fit to the data asimple visual search model postulating thatthe abrupt onset item is scanned first onevery trial but that otherwise the search isserial and self-terminating. We shall refer tothis as the abrupt-capture model. This modelbears some resemblance to Schneider andShiffrin's (1977, p. 29) Model la. The modelis expressed as

RT**A + kT+5N, (1)where E(A) = a, E(T) = r, E(N) = v, RT isa random variable representing reaction time,A is a random variable reflecting the time forall mental operations (e.g., encoding, motorprogramming, response execution) not ac-counted for by the other terms in the equa-tion, k is the number of comparisons requiredon a trial, T is a random variable reflectingthe time required to complete one compari-son, <5 is an indicator variable that equals 1if the target is absent and 0 otherwise, N is arandom variable corresponding to the extratime required to deal with a negative trial(Nickerson, 1965), and E denotes expectedvalue. Under the model, k is exactly onewhen the target has an abrupt onset, nomatter how many other items are present inthe display. When the target is absent, k isequal to the display size d since the search isexhaustive in this case. When the target ispresent but has a gradual onset,

Table 3Parameters and Predictions of the Model, Day 2Data Only, Experiment 1

The initial 1 in this equation corresponds toscanning the abrupt onset item, which bystipulation is not the target. This leaves d -1 items to be scanned in a self-terminatingsearch. The mean number of subsequent

comparisons then will be - - - . (This

Trialtype

OnsetOnsetNo-onsetNo-onsetNegativeNegative

Displaysize

242424

k112324

S

000011

Mean RT (ms)

Obs Pred

441.8 450.0457.7 450.0482.9 488.1531.9 526.2524.4 521.3594.4 597.6

Note. In the no-onset-4 condition, A; is 3 on the average.All other values of k are exact. R2 = .987. RT = reactiontime; Obs = observed; Pred = predicted.

holds because the target appears in eachlocation equally often in a random order.)Thus in the present case, k = 2 when thedisplay size is two, and k has an expectedvalue of 3 when the display size is four.

The values of k and 5 under each of theconditions, as well as the observed responselatencies, are displayed in Table 3.

Multiple regression of the data under thismodel gave estimates of the parameters asfollows: a = 411.9, f = 38.1, and ? = 33.2.The fit was quite satisfactory, accounting for98.7% of the variance in the means with onlythree parameters.5 The predicted means areshown in Table 3 and Figure 2. The estimateof 38.1 ms/comparison for the search param-eter T is in good agreement with other esti-mates in the literature (e.g., Schneider &Shiffrin's (1977) estimate of 42 ms/compari-son and Sternberg's (1966) of 38 ms/compar-ison).

Root mean squared (RMS) error6 was 5.9ms, or less than 2% of mean RT. To further

5 The abrupt capture model also makes predictionsabout variances under each condition (cf. Schneider &Shiffrin, 1977, Appendix G). The fit of the predictedvariances to our data was not good. In particular, underconditions for which a disproportionately large amountof variance was predicted (i.e., those in which differentnumbers of comparisons are made across trials), weobserved even more variance than predicted. The standarderror of the observed variances was quite large, however,and so our ability to reject any plausible model wasweak.

6 RMS error is defined as

n " ~i"2

- 2(obs, - pred,)2 .L»/ - i -I


assess the adequacy of the model, fits wereperformed using the means from each con-dition for each subject individually. The me-dian proportion of variance accounted for bythe model within each subject was .933 (range:.5S-.99). The parameter estimates were sim-ilarly reasonable for individuals; for instance,the median T obtained was 38 ms/comparison(range: 15-61).7

We fit several other plausible models, in-cluding Model 1, a parallel, limited capacity,abrupt-capture model that mimics the serial,capture model (Atkinson et al., 1969; Town-send, 1972); Model 2, a parallel, unlimitedcapacity, capture model that does not allowfor display size effects; Model 3, a parallel-capture model that requires a serial re-searchof the display set on negative trials; andModel 4, a standard serial model with noprovision for abrupt capture. (See the Appen-dix for further description of these models.)None of these models fit the data as well asthe serial-capture model did, although Model1 did almost as well. The proportions ofvariance accounted for by Models 1 through4 were .985, .761, .895, and .772, respectively,as compared to .987 for the serial abruptcapture model. RMS error values for Models1-4 were 6.3, 25.1, 18.8, and 24.5 ms, re-spectively, whereas the value for the serialabrupt-capture model was 5.9 ms. The con-trasts in these fits underline the quality of theserial abrupt-capture fit. We prefer the serialabrupt-capture model to the parallel, limitedcapacity model because the latter predictsparallel slopes for the negative and no-onsetconditions, whereas our analysis revealed asignificant deviation from this pattern (p <.05; see previous discussion of results).

From these results we conclude that thedata are well described as being generated bya process that scans the single onset locationfirst and all other locations serially in arandom order until a target is found or thesearch is complete.

Experiment 2

The critical independent variable in Ex-periment 1 was onset type, and the conclu-sions to be drawn from the results of thatexperiment depend on the assumption thatgradual no-onsets as instantiated in this study

are not perceptually more difficult to processthan abrupt onsets are. But this assumptionmay be false. For instance, the visual receptorsmay adapt to the figure-8 premasks, resultingin reduced sensitivity at locations they occupy.Because we are claiming that the effect ofabrupt onset is to summon attention when itis not already directed to the abrupt onsetstimulus location, then the processing advan-tage an abrupt item confers should declineor disappear if attention is appropriately di-rected in advance. Indeed, in such a case thesubject may be able to make better use ofthe preview afforded by the gradual targetrevelation and thereby show an advantage forno-onset stimuli. However, if the advantageof abrupt onset in Experiment 1 was not dueto attentional factors, then the advantageshould remain when attention is appropriatelydirected in advance.

Experiment 2 was conducted to test thesepredictions and thys to distinguish betweena perceptual and an attentional explanationfor the effects observed in Experiment 1.Subjects were required on each trial to identifya single letter that occupied a known location;in some blocks, the item had an abrupt onset,and in others, it appeared via the modifiedno-onset procedure used in Experiment 1.Eye position was monitored to ensure thatdepartures from fixation were detected andadmonished. Thus the retinal eccentricitiesof Experiment 1 were maintained.

Method

Subjects. Twenty undergraduates with uncorrectednormal vision (11 female, 9 male) each served in a single30-min session. None had served in Experiment 1.Compensation was $3.

Apparatus and stimuli. In addition to the equipmentused in Experiment 1, a Gulf and Western model 200scleral reflectance eye movement monitor was used to

7 One aspect of the capture model that remains un-specified is whether search is consistently ordered, and ifso, how it is ordered. To investigate this issue, wecalculated mean response times to no-onset targets whenthe display size = 4 (all other cases are degenerate) as afunction of onset item location. No consistent orderingof response times was observed, including (a) clockwiseor counterclockwise from the onset location and (b)locations near the onset item first and opposite the onsetitem last. Insofar as our data can reveal, then, search inthis task is not consistently ordered.


determine eye position. Eye movement data were collectedthrough an analog-to-digital (A/D) converter on the PDP-11/60 computer.

Head position was maintained by a chinrest. Subject'seye position was calibrated at the beginning of each 30-trial block and again whenever departures from fixationwere detected on 3 trials since the last calibration run.A simple position criterion (±2° from fixation) was usedto detect eye movements toward any of the displaylocations. Because the monitor was configured to measureonly horizontal movements, glances directly upward ordownward could not be detected. Subjects were unawareof this. Nevertheless, in the analyses below, trials arepartitioned into those with vertical and those with non-vertical display locations in order to reveal possibledifferences due to undetected vertical movements.

Eye movement calibration. All eye position samplingvia the A/D converter was at 1 kHz. Calibration consistedof sampling for 50 ms at each of five eye positions,corresponding to fixation, ±2°, and ±4°, run in randomorder. Conversion from A/D units to gaze position wasthrough linear interpolation. Preliminary tests showedthat observers could not voluntarily move their eyes orheads without such movement being detected.

Procedure. The sequence of events on each trial (seeFigure 3) was as follows. A single target letter appearedat fixation for 1,000 ms. The target was replaced by afixation cross, which remained throughout the durationof the trial. After 150 ms, an indicator rectangle appearedat one of the six possible display locations. During no-onset blocks, the indicator was a block figure 8; duringonset blocks, the indicator was a configuration of six dotsat the vertices of a block figure 8. The indicator waspresent for 1,000 ms. The procedure for the two typesof blocks differs somewhat at this point. For noronsetblocks, irrelevant segments began to fade as in Experiment1 over a period of 80 ms, gradually revealing the displayitem (see Footnote I). For onset blocks, the six-dotconfiguration began to fade as a whole over an 80-msperiod. At the instant the dots were gone, the displayitem was abruptly illuminated at that location. Thedisplay item remained until a response was made, atwhich time feedback appeared for 500 ms. A 1,000-msintertrial interval followed.

Subjects were told that the purpose of the indicatorwas to allow them to prepare in advance for the displayitem: They were asked to "pay attention" to the spatiallocation occupied by the indicator, because the displayitem would appear there after a consistent delay. Theywere further informed that they should do this whilemaintaining fixation. The working principles of the mon-itor were briefly described.

Subjcts were instructed to respond as quickly as possiblewhile making very few errors. A positive response wasrequired if the display item was identical to the target;otherwise a negative response was called for. As inExperiment 1, response latencies were measured fromthe end of the camouflage removal, or, equivalently, fromthe onset of the abrupt-onset item. A beep after eacherror served as feedback. If an eye movement was detected,the beep sounded, and the trial response was discarded.All trials during which departures from fixation weredetected were rerun at the end of each block. As men-tioned earlier, if movements were detected on three trials

within a block, the calibration routine was automaticallyinvoked.

Design. All five letters served equally often as thetarget. One half of the trials required positive responses,the rest negative. On negative trials, the display item waschosen randomly from the four nontarget items for thattrial. Each target also appeared at each display positionequally often. Two blocks of 30 trials each (plus reruntrials) were completed with abrupt onsets, and two withno-onsets. Onset type (no-onset, N, or onset, O) andresponse (positive or negative) were completely crossedwithin-subject variables. Order of blocks (onset or no-onset first) was a between-subject variable. Half thesubjects ran with blocks ordered ONON, and the restwith NONO.


Eye movements were detected on 8.7% ofthe trials. Many of these can probably beattributed to inadvertent head movements

Duration ONSET NO-ONSET

1000

1000

80

RT

P

B

B

Figure 3. Sequence of events for a typical trial inExperiment 2. (Dashes represent fading line segments.The correct response for both trial types here is no.)


permitted by the unintrusive chinrest used.Nevertheless, we discarded all trials duringwhich movements were detected; these trialsare not included in what follows.

The mean error rate across all conditionswas 7.2%. The somewhat higher error rateobserved here is probably due to the relativelysmall amount of practice in this experiment.Overall, subjects were less accurate in thecondition with which they began. Errors werepositively correlated with RT (r = .75),indicating the absence of speed-accuracytrade-off.

Mean RT for each group and condition isshown in Table 4. The practice effect evi-denced in the error data is clearly apparenthere: Subjects were slowest in the conditionin which they started. Overall, however, sub-jects were 10.6 ms faster when the displayitem appeared in the gradual, no-onset fashionof Todd and Van Gelder (1979) than via thestandard abrupt-onset method.

To assess the magnitude of these effects,an analysis of variance was conducted withorder as a between-subjects variable (onset orno-onset block first) and onset type (onset orno-onset) and response (positive or negative)as within-subjects variables. Subjects werefaster on positive than on negative trials, F( 1,18) = 42.4, MSe = 1,258,;? < .001. The maineffect of onset type, however, did not attainsignificance, F(l, 18) =1.4, MSe = 1,624,p > .20, although no-onset RT was 10.6 msfaster than onset RT (SE = 10.1 ms). Thepractice effect is seen in the significant Or-der X Trial Type interaction, F(l, 18) = 5.6,MSe = 1,623, p < .05: Subjects improvedmore when they began with an onset blockthan when they began with a no-onset block.This interaction suggests that with practice,subjects were better able to take advantage ofthe 80 ms of preview afforded by the gradualcamouflage removal.

The analysis was repeated excluding verticaltrials, that is, trials on which the target ap-peared directly above or below fixation. Theseare trials on which any eye movements wouldhave escaped detection. The pattern of resultswas the same: Positive responses were fasterthan negative ones, F(l, 18) = 16.5, MSe =2,268, p< .001, but again, although no-onsetRTs were faster than onset RTs by 2.2 ms,this difference did not attain significance

Table 4Mean Reaction Time and Standard Error (ms) inEach Condition of Experiment 2

Order ofconditions

Onset-No-onsetNo-onset-Onset

Combined

Abrupt onset

M

526534530

SE

212516

Gradualonset

M

492546519

SE

232417

(F< I). Apparently, then, eye movementscannot account for these results.

To further evaluate the onset-no-onset dif-ference, we attempted to remove the practiceeffect seen as an interaction in Table 4. Meanreaction times for the four blocks, collapsedacross conditions, were 596, 520, 508, and513 ms, respectively, suggesting asymptoticperformance by Block 3. An analysis of vari-ance as above was therefore conducted onBlocks 3 and 4 only to determine whether,after practice, gradual presentation producesresponses as fast or faster than abrupt onsetswhen attention is appropriately allocated.Mean no-onset RT was 13 ms faster thanmean onset RT, but once again this differencewas ndt significant, F(l, 18) = 2.0, MSe =1,676, p > .10. Of the remaining effects andinteractions, only the response type F ratioexceeded unity, ̂ 1,18) = 26.2, MSe = 1,999,p < .001.

In summary, the results of Experiment 2indicate that when attention is appropriatelydirected in advance of a visual display, thereis no advantage for abrupt onsets over gradualno-onsets. The pattern of results obtained inExperiment 1 cannot be attributed to percep-tual or sensory factors. Instead, an explanationof that experiment based on attentional cap-ture is supported.

These results may seem somewhat incon-sistent with some findings reported by Toddand Van Gelder (1979). In particular, Exper-iments 3 and 4 in that article involved theuse of precues that provided some locationalinformation about upcoming stimulus events.In principle, following the reasoning usedhere, subjects in those experiments ought tohave been able to allocate their attention tothe cued locations in advance and thereby


eliminate the latency difference between onsetand no-onset. In contrast, Todd and VanGelder found a large effect of onset type oncued trials. Two aspects of their procedure,however, may have precluded the effectiveassignment of attention before trial onset.First, the cues were only partially valid: On25% of the trials, the stimulus appeared inan uncued location. Second, the cues indi-cated two spatial locations that were at least5° and as much as 20° of visual angle apart.Hoffman and Nelson (1981), Posner et al.(1980), and others have shown that subjectscannot simultaneously attend to multiplenoncontiguous spatial locations efficiently.Thus Todd and Van Gelder's cuing procedureshould not be viewed as an effective atten-tional manipulation.

The fact that no-onset RT was only 10.6ms faster than onset RT deserves some com-ment. One might expect that because no-onset stimuli began to be revealed 80 msbefore the clock started, they might result inresponses that were as much as 80 ms fasterthan those to onset stimuli. There are tworeasons to doubt this. First, the no-onsetitems did not become clearly discriminableuntil perhaps 20-40 ms before the clockstarted. Thus the 80 ms figure is an overes-timate of the "preview time." Second, thereis some reason to suspect that RTs to onsetstimuli may indeed be faster than to no-onsetstimuli after attentional effects have beenminimized (Todd & Van Gelder, 1979, Ex-periment 3). Thus the latency advantage forno-onset stimuli is smaller than the previewduration might suggest. The important pointfrom this experiment is that the advantagefor onset stimuli vanishes when attention isappropriately directed in advance of an abruptevent, demonstrating the attentional natureof the advantage.

Experiment 3

In Experiment 1, we altered Todd and VanGelder's (1979) no-onset procedure by re-moving the camouflage gradually rather thanabruptly. Our reason for doing this was tominimize the possible attention-capturing ef-fects of abrupt offsets. There is evidence fromboth electrophysiology (Ikeda & Wright, 1974)and psychophysics (Kulikowski & Tolhurst,

1975) that the transient system responds tooffsets as well as to onsets. Consequently, tooptimize our test, we avoided abrupt offsets.

However, there is reason to believe thatthis precaution was not necessary. For in-stance, Breitmeyer and Julesz (1975) foundenhanced contrast sensitivity to gratings withabrupt leading edges and gradual trailingedges, but not to ones with only abrupttrailing edges, suggesting the visual system isnot as sensitive to offsets as to onsets. Fur-thermore, Todd and Van Gelder (1979) usedonly abrupt offsets in their no-onset experi-ments and still obtained large effects. In fact,they found that increasing the number ofirrelevant abrupt offsets facilitated RTs to no-onset items. Offsets, then, may have a fun-damentally different attentional status thanonsets do.

In Experiment 3, we replicated Experiment1 twice within subjects: Camouflage was re-moved either with an 80-ms Gaussian offsetfunction or abruptly. These two offset typeswere randomly intermixed. The apparatusused in Experiment 3 was an improvementover that used in Experiment 1: Here weused seven-segment light-emitting diodes(LEDs) driven by a digital-to-analog converterallowing nearly continuous intensity changesand thus smooth temporal waveforms. Wecould therefore assess the adequacy of theoffset step function employed in Experiment1 (see Footnote 1).

Method

Subjects. Nineteen University of Michigan under-graduates participated in two SO-min sessions. None ofthese people had served in Experiments 1 or 2, and eachwas paid $7 for his or her time.

Apparatus and stimuli. In order to effect greatercontrol over the temporal waveform of the onsets andoffsets of the stimuli, we constructed a display deviceconsisting of six 7-segment LEDs (Jimpak #DL-750) andan array of digital logic to control the individual LEDsegments. The hardware logic was under software controlvia a digital output interface. The LED segments weredriven by the digital-to-analog (D/A) converter of a DECPDP-11/34 computer. The D/A converter was configuredto output voltages at 1 kHz, so that voltages could bealtered once per millisecond. There were 1,024 functionalvoltage levels. Letters with abrupt onsets were illuminatedthrough the action of a single digitally controlled relay.

In removing the camouflaging LED segments, weemployed a Gaussian offset function similar to thatemployed by Wilson (1978); its duration was 80 ms from


1.5 -I

.M

I

QL 0.5-

20 40 60

Time (msec)80 too

Figure 4. Temporal luminance profile of camouflage removal in Experiment 3.

completely on to entirely off. The equation for theGaussian (with t in ms) was

7(0 = exp[-(f/40)2],

where / is relative intensity and 0 si t =s 80.8 Figure 4shows the temporal luminance profile. At their maximumbrightness, each letter had a luminance of 3.6 cd/m2.The luminance of the blank display screen was 1.8 cd/m2, and general booth illuminance was 190 Ix at thedisplay surface.

The LEDs were arrayed at the vertices of an imaginaryhexagon centered about fixation. The LEDs were 5.7°from fixation and 5.7° apart center to center. Each ofthe five-segment letters (E, H, P, S, and U) subtended1.9° of visual angle in height and .95° in width. TheLEDs were mounted in a flat black display surface 60cm (53°) high and SO cm (48°) wide. All visual angleswere measured from a viewing distance of 45 cm.

The LED display was controlled and responses collectedby a DEC PDF 11/34 computer. Subjects responded bypressing the appropriate keys on a Hewlett-Packard 2621Aterminal keyboard. Subjects were seated in a sound-attenuating booth illuminated by a fluorescent bulblocated above and behind them.

Procedure. The procedure was identical to that usedin Experiment 1, with the exception that on half thetrials the irrelevant segments were removed abruptly,whereas on the remaining trials the irrelevant segmentswere removed via an 80-ms half-Gaussian (see Figure 4).The design of Experiment 1 was thus replicated twice,once with abrupt camouflage removal and once withgradual camouflage removal, with the two trial typesrandomly intermixed. Subjects were not informed of thisfact, and upon debriefing none reported having any

awareness of it. Subjects completed 6 blocks of 80 trialson each of 2 days, for a total of 960 trials per subject.


The mean overall error rate collapsed acrossdays, subjects, and conditions was 5.4%. Errorrates for each condition are shown in Table5. As before, error rates for the variousconditions correlated positively with the as-sociated reaction times (r = .41), indicatingthe absence of a speed-accuracy trade-off.

An ANOVA was performed on the meanRTs with the following factors: day (1 or 2),camouflage removal waveform (abrupt orgradual), trial type (onset, no-enset, or neg-ative), and display size (two or four). All ofthe main effects were highly significant: F(l,18) = 59 for day, F(l, 18) - 106 for cam-ouflage waveform, f\2, 36) = 133 for trial

8 The intensity of LEDs is known to vary linearly withcurrent but not with voltage (cf. Nygaard & Frumkes,1982). Because we intended to vary the intensity ofindividual LED segments via voltage changes in the D/A converter, we empirically determined the functionrelating voltage to intensity (or current). The relationbetween voltage and current was quadratic. We compen-sated for this nonlinearity in our voltage output function.


Table 5Mean Error Rates (%) for Each Condition andDisplay Size (2 and 4) in Experiment 3

OnsetCamouflage

removal

No-onset Negative

Day 1

GradualAbrupt

5.74.5

5.26.6

8.27.9

8.712.1

8.96.2

7.08.2

Day 2

GradualAbrupt

3.12.7

1.73.0

3.64.4

5.25.2

3.03.4

2.13.0

type, and F(l, 18) = 146 for display size, allps < .001. No interaction involving day wassignificant (all ps > .05). As in Experiment1, the interaction between trial type anddisplay size was significant, F(2, 36) = 23.2,MSe = 753, p < .001; again, the display sizeeffect depended upon the trial type.

We shall once again concentrate on themore highly practiced Day 2 performance inwhat follows, given the lack of interactioninvolving day. Mean response times are shownas the observed points in Figure 5. Of partic-ular importance in this experiment was theeffect of camouflage offset waveform on theTrial Type X Display Size interaction. Thisthree-way interaction did not attain signifi-cance, F(2, 36) = 0.4, MSe = 278. In agree-ment with this result, the Camouflage Wave-form X Display Size interaction also failedto reach significance, F(l, 18) = 0.1, MSe =261. Thus the abrupt-capture effect seen inExperiment 1 was uninfluenced by whetherthe camouflage was removed gradually or allat once.

The Camouflage Waveform X Trial Typeinteraction, however, did attain significance,F(2, 36) = 4.65, MSe = 515, p < .05; therelative positions of the three trial type func-tions in the two panels of Figure 5 dependupon the offset waveform. In particular, al-though there was a general increase in latencywith abrupt camouflage removal comparedto gradual removal (as revealed by the maineffect of camouflage waveform), the no-onsetfunction climbed more than the onset func-tion did; this suggests that gradual removaldid give subjects some usable preview infor-

mation. Nevertheless, abrupt onset targetswere always detected more rapidly.

An inspection of Figure 5 will confirmthat both abrupt and gradual Camouflageremoval produced the effect seen in Experi-ment 1: In both cases, we observed a negligibledisplay size effect for onset targets and alarger one for no-onset targets. The slopes forthe Day 2 onset, no-onset, and negative con-ditions, respectively, were 9, 12, and 27 ms/item for gradual camouflage removal, and 7,14, and 25 ms/item for abrupt camouflage

•—• observedo—o predicted negative

positive no-onset

-• positive onset

Display Size

negative

positive no-onset

—_-——s8 positive onset

2 4Display Size

Figure 5. Mean reaction time on Day 2 as a function ofdisplay size. (Curve parameter is trial type. Panel a:gradua| camouflage removal. Panel b: abrupt camouflageremoval.)


Table 6Parameter Estimates for Serial Capture Model,Experiments 1 and 3

Dataset a f v R

Experiment 1Experiment 3

Gradual removalAbrupt removal

412

475488

38.1

28.835.4

33.2

50.342.5

.987

.987

.965

removal. Bonferroni multiple contrasts con-firmed that the onset slopes in each camou-flage waveform condition were not signifi-cantly larger than zero (p > .05). Althoughno-onset slopes were significantly larger thanzero (p<.01), they were not significantlylarger than the onset slopes (p > .05), incontrast to the findings of Experiment 1.Finally, there was a significant difference be-tween no-onset and negative slopes and be-tween onset and negative slopes (p < .01).

The data were fit to the serial abrupt-capture model described in Experiment 1.The predicted means are illustrated in Figure5. Table 6 summarizes the resulting parameterestimates, along with those from Experiment1 for comparison. The model accounts for98.7% and 96.5% of the variance in themeans for the gradual and abrupt camouflage-waveform conditions, respectively. RMS errorin the mean data was 5.4 ms and 9.8 ms forthe gradual and abrupt waveform conditions,respectively; these values are less than 2% ofthe mean response times. Individual fits foreach subject yielded median proportions ofvariance accounted for by the model withineach subject of .923 (range: .6S-.99) and .884(range: .6S-.99) for graudal and abrupt cam-ouflage removal conditions, respectively. Theparameter estimates were again reasonable;

for instance, the median ^ obtained was 32ms/comparison (range: 8-69). The fits aresatisfactory and the parameter estimates arein rough agreement with those from Experi-ment 1.

We once again fit to the data the fouralternative models described in the Appendix.Table 7 summarizes the goodness of fit forthe data from Experiments 1 and 3 for theserial abrupt-capture model as well as for thefour alternatives. Again, the serial, abrupt-capture model prevails. As in Experiment 1,the parallel, limited capacity, capture modelwas fit quite well. However, because the no-onset and negative slopes were significantlydifferent (p < .01)—a violation of the parallelmodel's predictions—the serial, abrupt-cap-ture model is preferred.

These results indicate that our precautionof removing the camouflage gradually in Ex-periment 1 was not crucial to the outcomeof the experiment. At least in the context ofthis study, the presence of several abruptoffsets (sometimes as many as two completefigure eights) did not dramatically disrupt theapparent attention-capturing property of asingle abrupt onset. This is consistent withTodd and Van Gelder's (1979, Experiment 2)finding that with stimulus-response uncer-tainty held constant, increasing the numberof abrupt offsets in a display does not interferewith responses to remaining items.

One inconsistency between Experiments 1and 3 remains: The Bonferroni contrastsshowed that in Experiment 1, no-onset slopeswere significantly larger than onset slopes,whereas in Experiment 3, the two were notstatistically distinguishable. This result maybe accounted for by the fact that the searchrate T for subjects in Experiment 1 (38.1 ms/

Table 7Goodness of Fit for Alternative Models, Experiments 1 ami 3

Serial capture Parallel limitedParallel

unlimitedParallel

re-search Standard serial

Experiment R2 RMSE RMSE RMSE RMSE RMSE

13

GA

.987

.987

.965

5.9

5.49.8

.985

.965

.961

6.3

8.910.1

.761

.862

.893

25.1

17.616.7

.895

.954

.885

18.8

10.217.5

.772

.843

.684

24.5

18.829.0

Note. RMSE = root mean squared error; G = gradual camouflage removal; A = abrupt camouflage removal.


comparison) was somewhat slower than thatfor subjects in Experiment 3 (28.8 and 35.4ms/comparison). The more rapid search ratesobserved in Experiment 3 may have renderedtrue differences between the two slopes un-detectable. The satisfactory model fits wehave observed throughout support this view.

General Discussion

The results of these experiments are welldescribed by the abrupt-capture model out-lined at the end of Experiment 1: On everytrial of visual search, attention is rapidlyassigned to the channel containing the abruptitem. A comparison is made with the targetitem residing in memory, and on a match, apositive response is emitted; on a mismatch,the search continues in the standard serialself-terminating fashion. In Experiment 2, wetested and rejected the possibility that subjectscould more easily process the onset itemsthan the no-onset items due to some physicaldifference between the two. In that experi-ment, when attention was appropriately al-located in advance, there was no differencein detection latency. Finally, in the thirdexperiment, we established that the no-onsetprocedure of Todd and Van Gelder (1979) iseffective whether or not camouflage is re-moved gradually.

A consistent finding in both Experiments1 and 3 was not predicted by the abrupt-capture model. In all three replications, weobserved small positive slopes in the abrupt-onset conditions, where a slope of zero waspredicted. A possible explanation for thisresult9 is that attention is captured on almostall of the trials, but on some small proportionp of the trials, capture is not effective, and aserial search of the entire display ensues. InExperiment 1, for instance, p was estimatedat about 10%. The addition of this parameter,however, improved R2 from .987 to only .989,hardly justifying its inclusion in the model.Nevertheless, this mixture assumption appearsto give a reasonable post hoc account of theotherwise unpredicted nonzero abrupt onsetslopes.

It is important to remark that the occur-rence of an abrupt onset does not by itselfconstitute a sufficient stimulus for capturingattention. Almost all experiments involvingthe presentation of visual stimuli, and in

particular nearly all the visual search exper-iments conducted over the last decade and ahalf, involved only stimuli with abrupt onsets.Because we know that attention cannot besimultaneously committed to multiple non-contiguous spatial channels (Hoffman & Nel-son, 1981; Posner et al., 1980), then clearlyattention cannot be summoned to all abruptvisual events at once. Apparently only whenthe visual field contains but one such event(as when a relatively static scene is viewedand a moving object appears in the visualperiphery) can attention be engaged in themanner illustrated by these experiments.

A demonstration of this idea was reportedrecently by Kahneman, Treisman, and Burkell(1983). Subjects in these experiments wereasked to read or to detect the presence ofwords appearing at spatially uncertain loca-tions in the visual field. In some conditions,the relevant stimulus appeared alone, and inothers it appeared simultaneously with irrel-evant and highly discriminable distractors.Kahneman et al. found that even when thedistractors were patches of random dots withalmost no features in common with the targetword, subjects were significantly slower innaming the word as compared to conditionsin which the word appeared alone. We inter-pret this as representing attentional captureby the single item, an event that is impossiblewhen more than one abrupt onset occurs.This view is supported by the results ofExperiment 5 by Kahneman et al. (1983), inwhich distractor elements were presented atvarious asynchronies with respect to the targetword. When either an abrupt onset or anabrupt offset occurred simultaneously withtarget onset, response latencies were signifi-cantly slower than in a target-alone condition.In contrast, continuously present distractorsor distractors removed well in advance oftarget onset resulted in RTs as fast or fasterthan those in the target-only condition.10

9 We thank J. E. K. Smith for this idea.10 This finding is somewhat inconsistent with the results

of Experiment 3, which indicated that abrupt offsets didnot interfere with capture by abrupt onsets. However,because the Kahneman et al. (1983) experiment includedsome conditions that are not analogous to any in thepresent experiments, direct comparisons may not beappropriate.


Another illustration of this point is foundin an experiment by Kowler and Sperling(1983). Subjects were shown a series of 5 X5 character arrays, one of which contained23 letters, one digit, and a central fixationpoint; the rest consisted of 24 letters and afixation point. Their task was to identify andlocate the digit. Each array was presented inone of several ways: either with abrupt onsetand ramped offset, ramped onset and abruptoffset, or abrupt onset and offset. Note thatin contrast to the present experiments, theentire array followed a single temporal course.In this respect, the Kowler and Sperling pro-cedure is similar to that employed by Breit-meyer and Mesz (1975). Kowler and Sperlingfound no enhancement or benefit of any kindfor abruptly presented arrays over those pre-sented gradually. This study was designed todetermine whether the abrupt changes inretinal projection due to saccadic eye move-ments might functionally enhance processingof a newly fixated scene. Kowler and Sperlingconcluded that they do not. This is consistentwith the hypothesis advanced in this articlethat isolated abrupt onsets capture attention.If all onsets, regardless of context, enhancedvisual processing, the attentional capture ex-planation for the current results would bewithout merit, and an alternative model wouldbe required.

One such alternative to be consideredis Krumhansl's (1982) enhanced encodingmodel that we described earlier. This modeldoes not employ the concept of attention,nor does it use the sensitivity of transientchannels to abrupt onset or change. Thecritical property of Krumhansl's model is theinitiation of a phase of rapid encoding atstimulus onset that gradually changes to aperiod of less rapid processing as the stimulusremains in view. This model describedKrumhansl's data quite well But although itdoes predict a difference between the onsetand the no-onset conditions, it has no pro-vision for the observed difference in the dis-play size effects because stimulus encoding inthe model occurs in all locations simulta-neously and in parallel. Thus Krumhansl'smodel cannot as it stands account for thepresent data (although it could be expandedto do so; C. L. Krumhansl, personal com-munication, 16 January 1984). Some appeal

to attention (scarce resources, a limited ca-pacity processing mechanism) seems requiredby the data reported here.

The current model would account forKrumhansl's data in the following manner.Initially, attention is evenly distributed overthe prestimulus array. When the prestimulusarray disappears, the subject must localizeand identify the remaining stimulus. If theremaining object is of the form-change typeand therefore exhibits an abrupt change attrial onset, attention is immediately securedby the appropriate input channel, and therequired processes may proceed at once withtheir tasks. In contrast, when the remainingobject does not contain an abrupt change attrial onset, there is no attentional imperative"marking" the relevant channel. Thus timemust be occupied with a rather slower local-ization process (which frequently is not com-plete at mask onset) before stimulus identifi-cation can commence.

On the current model, an abrupt onsetamong gradual ones has an attentional statussimilar to that of an X among 0s (Treisman& Gelade, 1980) or a digit among letters(Egeth et al., 1972). That is, it is a perceptualfeature of the stimulus that can be "encoded"at some early level of the visual system inparallel and without restriction across theentire field and that can then be used as abasis for passing the associated stimulus intoa limited-capacity system (Duncan, 1980). Inour view, the abrupt onset differs from theseother stimulus properties (e.g., shape, cate-gory) in that it is an imperative; it summonsattention (in Duncan's terms, it requires im-mediate, passage into the limited-capacity sys-tem), perhaps without intention, effort, orawareness on the subject's part, These aresome of the commonly held criteria for pro-cessing automaticity (Posner & Snyder, 1975;Schneider & Shiffrin, 1977).

Some predictions of this model may makeit more concrete. One is that with briefexposures, objects spatially adjacent to theabrupt object should be processed more effi-ciently than those farther from it (cf. Hoffman& Nelson, 1981). Other predictions have todo with the involuntary nature of automatic-ity: If a subject is asked to attend to aparticular spatial location (e.g., Eriksen &Hoffman, 1973) and if that location contains


a target, then we would expect that an abruptnontarget in some other location would bemore likely to interfere with target processingthan a no-onset nontarget would. Conversely,if an unattended object is the target, wewould expect performance to be neverthelessquite good if it has an abrupt onset. In otherwords, objects with abrupt onsets should beefficiently processed regardless of the subject'sintentional allocation of attention. These pre-dictions hold only, of course, if there is atmost one abrupt item in each display.

A finding of Todd and Van Gelder (1979)that is also predicted by this model involvesthe increased complexity of the task requiredof subjects across experiments from one ofsimple detection to more difficult letter dis-crimination. The relative advantage of onsetover no-onset presentation increased withcomplexity. This is consistent with the finding(Shaw, 1984) that letter discrimination cannotbe accomplished with divided attention,whereas simple luminance increment detec-tion can be. If abrupt onsets capture attention,the observed effects of such capture wouldbe expected to be greatest when attention ismost needed. This is just what Todd and VanGelder (1979) observed.

Conclusion

We have shown in these experiments thatisolated abrupt onsets are rapidly detected invisual search, and we have offered an atten-tional capture model that satisfactorily ac-counts for the data. The model provides forthe immediate recruitment of attentional re-sources by visual channels containing signalswith abrupt onsets.

ReferencesAtkinson, R. C, Holmgren, J. E., & Juola, J. F. (1969).

Processing time as influenced by the number of ele-ments in a multielement display. Perception & Psycho-physics, 6, 321-326.

Bashinski, H. S,, & Bacharach, V. R. (1980). Enhancementof perceptual sensitivity as the result of selectivelyattending to spatial locations. Perception & Psycho-physics, 28, 241-248.

Bouma, H. (1978). Visual search and reading: Eye move-ments and functional visual field: A tutorial review. InJ. Requin (Ed.) Attention and performance VII (pp.115-147). Hillsdale, NJ: Erlbaum.

Breitmeyer, B. G., & Ganz, L. (1976). Implications ofsustained and transient channels for theories of visual

pattern masking, saccadic suppression, and informationprocessing. Psychological Review, 83, 1-36.

Breitmeyer, B. G., & Julesz, B. (1975). The role of onand off transients in determining the psychophysicalspatial frequency response. Vision Research, 15, 411-415.

Broadbent, D. E. (1982). Task combination and selectiveintake of information. Acta Psychologica, 50, 253-290.

Cleland, B. G., Levick, W. R., & Sanderson, K. J. (1973).Properties of sustained and transient cells in the catretina. Journal of Physiology, 228, 649-680.

Colgate, R. L., Hoffman, J. E., & Eriksen, C. W. (1973).Selective encoding from multielement visual displays.Perception & Psychophysics, 14, 217-224.

Duncan, J. (1980). The locus of interference in theperception of simultaneous stimuli. Psychological Re-view, 87, 273-300.

Egeth, H., Jonides, J., & Wall, S, (1972). Parallel processingof multielement displays. Cognitive Psychology, 3, 674-698.

Engel, F. L. (1971). Visual conspicuity, directed attention,and retinal locus. Vision Research, 11, 563-576.

Enroth-Cugell, C., & Robson, J. G. (1966). The contrastsensitivity of retinal ganglion cells of the cat. Journalof Physiology, 187, 517-552.

Eriksen, C. W., & Hoffman, J. E, (1972). Temporal andspatial characteristics of selective encoding from mul-tielement displays. Perception & Psychophysics, 12,201-204.

Eriksen, C. W., & Hoffman, J. E. (1973). The extent ofprocessing of noise elements during selective encodingfrom visual displays. Perception & Psychophysics, 14,155-160.

Estes, W. K., & Wessel, D. L. (1966). Reaction time inrelation to display size and correctness of response inforced-choice visual signal detection. Perception &Psychophysics, 1, 369-373.

Fukada, Y. (1973). Differentiation of the principle cellsof the rat lateral geniculate body into two groups; fastand slow cells. Experimental Brain Research, 17, 242-260.

Fukada, Y, & Saito, H.-I. (1971). The relationshipbetween response characteristics of flicker stimulationand receptive field organization on the cat's optic nervefibers. Vision Research, 11, 227-240.

Fukada, Y, & Stone, J. (1974). Retinal distribution andcentral projections of Y-, X-, and W-cells of the cat'sretina. Journal of Neurophysiology, 37, 749-772.

Graham, N. (1981). Psychophysics of spatial-frequencychannels. In M. Kubovy & J. R. Pomerantz (Eds.),Perceptual organization (pp. 1-25). Hillsdale, NJ: Erl-baum.

Hoffman, J. E., & Nelson, B. (1981). Spatial selectivityin visual search. Perception & Psychophysics, 30, 283-290.

Holmgren, J. E. (1974). The effect of a visual indicatoron rate of visual search: Evidence for processingcontrol. Perception & Psychophysics, 15, 544-550.

Ikeda, H., & Wright, M. J. (1972). Receptive field orga-nization of "sustained" and "transient" retinal ganglioncells which subserve differential functional roles. Journalof Physiology, 227, 769-800.

Ikeda, H., & Wright, M. J. (1974). Evidence for "sus-


tained" and "transient" neurones in the cat's visualcortex. Vision Research, 14, 133-136.

Jonides, J. (1980). Toward a model of the mind's eye'smovement. Canadian Journal of Psychology, 34, 103-112.

Jonides, J. (1981). Voluntary versus automatic controlover the mind's eye's movement. In J. B. Long &A. D. Baddeley (Eds.), Attention & performance IX(pp. 187-203). Hillsdale, NJ: Erlbaum.

Jonides, J., & Gleitman, H. (1972). Conceptual categoryeffect in visual search: 0 as letter or as digit. Perception& Psychophysics, 12, 457-460.

Kahneman, D., Treisman, A., & Burkell, J. (1983). Thecost of visual filtering. Journal of Experimental Psy-chology: Human Perception and Performance, 9, 510-522.

Kowler, E., & Sperling, G. (1983). Abrupt onsets do notaid visual search. Perception & Psychophysics, 34, 307-313.

Krumhansl, C. L. (1982). Abrupt changes in visualstimulation enhance processing of form and locationinformation. Perception & Psychophysics, 32, 5.11-523.

Kulikowski, J. J., & Tolhurst, D. J. (1973). Psychophysicalevidence for sustained and transient detectors in humanvision. Journal of Physiology, 232, 149-162.

Nickerson, R. S. (1965). Response times for "same"-"different" judgements. Perceptual and Motor Skills,20, 15-18.

Nygaard, R. W., & Frumkes, T. E. (1982). LEDs: Con-venient, inexpensive sources for visual experimentation.Vision Research, 22, 435-440.

Posner, M. I. (1980). Orienting of attention. QuarterlyJournal of Experimental Psychology, 32, 3-25.

Posner, M. I., Nissen, M. J., & Ogden, W. C. (1978).Attended and unattended processing modes: The roleof set for spatial location. In H. L. Pick & E. J.Saltzman (Eds.), Modes of perceiving and processinginformation (pp. 137-157). Hillsdale, NJ: Erlbaum.

Posner, M. I., & Snyder, C. R. R. (1975). Attention andcognitive control. In R. L. Solso (Ed.), Informationprocessing and cognition (pp. 55-85). Hillsdale, NJ:Erlbaum.

Posner, M. I., Snyder, C. R. R., & Davidson, B. J. (1980).Attention and the detection of signals. Journal ofExperimental Psychology: General, 109, 160-174.

Rabbitt, P. M. A. (1967). Learning to ignore irrelevant

information. American Journal of Psychology, 80, 1-13.

Schneider, W., & Shiffrin, R. M. (1977). Controlled andautomatic human information processing: I. Detection,search, and attention. Psychological Review, 84, 1-66.

Shaw, M. L. (1978). A capacity allocation model forreaction time. Journal of Experimental Psychology:Human Perception and Performance, 4, 586-598.

Shaw, M. L. (1984). Division of attention among spatiallocations: A fundamental difference between detectionof letters and detection of luminance increments. InH. Bouma & D. G. Bonwhuis (Eds.), Attention andperformance X (w. 109-121). Hillsdale, NJ: Erlbaum.

Shiffrin, R. M., & Schneider, W. (1977). Controlled andautomatic human information processing: II. Perceptuallearning, automatic attending, and a general theory.Psychological Review, 84, 127-190.

Sperling, G. (1960). The information available in briefvisual presentations. Psychological Monographs, 74(11,Whole No. 498).

Sternberg, S. (1966). High-speed scanning in humanmemory. Science, 153, 652-654.

Titchener, E, B. (1908). Lectures on the elementarypsychology of feeling and attention. New York:MacMillan.

Todd, J. T, & Van Gelder, P. (1979). Implications of atransient-sustained dichotomy for the measurementof human performance. Journal of Experimental Psy-chology: Human Perception and Performance, S, 625-638.

Tolhurst, D. J. (1975). Sustained and transient channelsin human vision. Vision Research, 15, 1151-1155.

Townsend, J. T. (1972). Some results concerning theidentifiability of parallel and serial processes. BritishJournal of Mathematical and Statistical Psychology,15, 155-197.

Treisman, A. M., & Gelade, G. (1980). A feature-integration theory of attention. Cognitive Psychology,12. 97-136.

Van der Heijden, A. H. C., & Eerland, E. (1973). Theeffect of cuing in a visual signal detection task. QuarterlyJournal of Experimental Psychology, 25, 496-503.

Wilson, H. R. (1978). Quantitative characterization oftwo types of line-spread function near the fovea. VisionResearch, 18, 971-981.

Wolford, G., & Chambers, L. (1983). Lateral masking asa function of spacing. Perception & Psychophysics, 33,129-138.


Appendix

Description of the Alternative Models

A brief description of the four alternative modelswe tested is given here. Theoretical predictions forthese models, as well as for the serial abruptcapture model described in the text, are shown inFigure A1. All of these predictions are based onvarying the definition of k in Equation 1, whileleaving its other terms intact.

1. Parallel, limited capacity abrubt-capturemodel. As Atkinson et al. (1969) and Townsend

-o A

AJtG

DISPLAY SIZEFigure Al. Theoretical predictions of the five alternativemodels considered in this article. (For convenience, wehave set v = T, Panel a: serial abrupt capture. Panel b:parallel, limited capacity capture. Panel c: parallel, un-limited capacity capture. Panel d: parallel, unlimitedcapacity re-search capture. Panel e: standard, serial, self-terminating. A = abrupt onset, G = gradual no-onset,N = negative.)

(1972) have pointed out, a parallel process canperfectly mimic a serial process, with the inclusionof certain assumptions. For instance, a parallelprocess can produce the display size effect—whichis the paradigmatic case of a serial process—if itis assumed that (a) processing resources are limited,(b) processing rate is directly proportional to theamount of available processing resources, and (c)as each item is identified, the rate of processingfor all unfinished items increases as limited pro-cessing resources become available. Thus, as thenumber of nontarget items in the display increases,the expected response latency also increases, be-cause available resources for any one item (andhence processing rate) are declining.

We instantiated the parallel mimic by assumingthat the abrupt onset item, as in the serial capturemodel, summons attention and is processed first.If the abrupt item is not the target, all other itemsin the array are scanned in parallel, with processingrate inversely proportional to the number of re-maining items. No-onset and negative slopes aretherefore predicted to be equal. Referring to Equa-tion 1, when the target is the abrupt onset item,k = 1; otherwise, k = d,

2. Parallel, unlimited capacity capture model.This model assumes that the onset item is processedfirst, and if it is not the target, all other items areprocessed in parallel with processing rate indepen-dent of display size. Here, k = 1 when the abruptitem is the target and k = 2 otherwise. This modelpredicts zero slopes for all three trial types, withintercept differences of r between the onset andno-onset funtions and v between the gradual andnegative functions.

3. Parallel, unlimited capacity re-search capturemodel. In order to produce nonzero slopes for thenegative trials, we evaluated a model identical toModel 2 with the added assumption that a serialre-search of the display occurs on negative trials.Under this model, k is as in Model 2 except onnegative trials, when k = 2 + d.

4. Standard serial model. This model assumesthat all the items in the display are treated iden-tically, so the onset and no-onset data are predictedto be the same. A serial, self-terminating searchpredicts a positive slope that is one half the mag-nitude of the negative slope. In terms of Equation1, k = (d + 1 )/2 on positive trials and k = d onnegative trials.

Received January 13, 1984Revision received April 27, 1984 •

Journal of Experimental Psychology: Human …...attention can be either central or peripheral...

Documents

Transcript of Journal of Experimental Psychology: Human …...attention can be either central or peripheral...