Representation and selection of relative position.

Journal of Experimental Psychology:Human Perception and Performance1993. Vol. 19. No. 3, 488-516

Representation and Selection of Relative Position

Andrew Heathcote and D. J. K. Mewhort

Part 1 concerns representation: We demonstrate pop-out for a target that can be distinguished fromthe distractors only by the relative position of its components and thereby show that simple shapeinformation is represented preattentively. We discuss our findings in terms of the early- versuslate-selection debate and in terms of theories of search and texture segregation. Part 2 concernsselection: We demonstrate strong learning in a varied-mapping paradigm and show that preattentiveshape information can be used for selection. Finally, we suggest an account of the learning, namelythe group scale hypothesis, and present a final experiment to test it. Our results confirm and extendthe emphasis placed by Duncan and Humphreys' (1989) attentional engagement theory on groupingprocesses in visual search.

In a search task, subjects are asked to indicate whether ornot a target is present in a display. A search function is thetime to decide (response time, or RT) plotted against displaysize (the number of targets and distractors). In a consistent-mapping (CM) paradigm, the same target is used across alltrials. In a varied-mapping (VM) paradigm, by contrast, atarget on one trial may serve as a distractor on another. Theodd-man paradigm is an extreme variation of the VM par-adigm: The target is a one-of-a-kind object in a set of iden-tical distractors. The present work uses CM, VM, and odd-man search tasks to study the representation and selection ofrelative position information.

Following Neisser (1967), models of perception assumethat the visual world is decomposed by the sensory systeminto elementary units and that two levels of processing areneeded to synthesize perceptual objects from the units. Thefirst level is thought to operate in parallel across the visualfield independently of attention and to have no capacitylimits. The second level is thought to have capacity limits,to operate serially across the visual field, and to involveattention.

A parallel distinction about levels of processing is used todescribe the locus of selection. Early-selection theories claimthat simple properties of an object, such as color, orientation,and size, are identified preattentively. Complex properties of

Andrew Heathcote and D. J. K. Mewhort, Department of Psy-chology, Queen's University, Kingston, Ontario, Canada.

This research was supported by Natural Science and EngineeringResearch Council of Canada Grant AP-318 to D. J. K. Mewhortand by a Commonwealth Fellowship to Andrew Heathcote. Thefirst two experiments appeared in Andrew Heathcote's doctoraldissertation (Queen's University, 1990).

We thank Jackie Braun, Debbie Feldman-Stuart, and Beth Johnsfor comments on earlier versions of this article, and A. Treisman,M. Tarr, and G. Humphreys for detailed and insightful reviews. Wealso thank Irene Armstrong for conducting Experiments 3 and 4.

Correspondence concerning this article should be addressed toAndrew Heathcote, who is now at the Department of Psychology,University of Newcastle, Rankin Drive, Newcastle, New SouthWales 2308, Australia. Electronic mail may be sent [email protected].

an object, such as shape and meaning, are thought to beconstructed at the second level of processing. Moreover, onlyproperties that have been identified preattentively have ac-cess to processing at the second level (e.g., Broadbent, 1958,1971, 1982; Sperling, 1960; Treisman, 1960, 1969, 1988).Late-selection theories, by contrast, claim that both simpleand complex properties of an object can be identified pre-attentively and that limitations to performance reflect post-identification processes (e.g., Allport, 1977; Deutsch &Deutsch, 1963; Duncan, 1980; Mewhort, 1987; Shiffrin.1975).

When applied to the search task, early-selection theoriesimply that search for simple properties is capacity free,whereas search for complex properties will reach capacitylimits. The limits can be overcome, however, by attending toone object at a time. Hence, for early-selection theories,search for complex objects implies serial processing. As aresult, the rate of decision processing is limited by the rateof perceptual processing, and the search function mustincrease.

By contrast, when applied to search, late-selection theoriesimply that perceptual processing does not limit the rate ofdecision in visual search. Display size effects may occur, butthey are assigned to a postidentification stage. Thus, depend-ing on the way decision processing occurs, both flat andincreasing search functions are consistent with the late-selection position.

In summary, in search for complex objects, early-selectionmodels always predict increasing functions, whereas late-selection models can tolerate both flat and increasing searchfunctions. Thus, search for complex objects provides anasymmetric test for the locus of selection: Increasing searchfunctions are consistent with both early- and late-selectiontheories, whereas flat search functions falsify early-selectiontheories.

Part 1: Representation and Detection

Object Complexity- and Capacity Limitation

Early-selection theories claim that the process of identi-fying an object's shape and meaning is capacity limited. The

488

ATTENTION AND VISUAL SEARCH 489

leading early-selection model for visual search, feature-integration theory (FIT), makes a stronger claim, namely thatsimple properties—called features in the theory—are iden-tified preattentively but that their position in the visual fieldis not. Because the position of simple properties is not avail-able preattentively, an object defined by a conjunction ofsimple properties must be assembled using attention to in-tegrate its features. Attention operates serially on an object-by-object basis. Thus, FIT predicts that search for a targetdefined by a conjunction of simple properties must produceincreasing search functions (see Treisman, 1982; Treisman &Gelade, 1980; Treisman & Patterson, 1984; Treisman &Souther, 1985; Treisman, Sykes, & Gelade, 1977).

Although FIT is consistent with a large body of data, sev-eral investigators reported flat search functions (also calledpop-out) in search for targets defined in terms of conjunc-tions of simple features. Early findings of pop-out were ob-tained using binocular disparity and motion, dimensions thathad not been studied by proponents of FIT (McLeod, Driver,& Crisp, 1988; Nakayama & Silverman, 1986; Steinman,1987). More recently, Wolfe, Cave, and Franzel (1989) foundpop-out in search for conjunctions of color, orientation, andsize, provided that the simple features are sufficiently dis-criminable (see also Dehaene, 1989).

Both Treisman and Sato (1990; Treisman, 1988) and Wolfeet al. (1989; Cave & Wolfe, 1990) proposed modifications ofFIT to account for pop-out in conjunction search. Both mod-ifications aim at reducing the amount of serial processing.Cave and Wolfe (1990), for example, proposed a guidancemechanism that gives attention a hint where to start andthereby reduces the average number of attention fixationsnecessary to detect a target. Neither modification, however,includes preattentive localization of object features. Hence,in both theories, a feature can be localized only by attendingto its position in the visual field.

Localization, Attention, and Shape Detection

An object's shape can be determined by matching it witha template. As Neisser (1967) explained some time ago, tem-plate matching at an object level implies a combinatorialexplosion, especially if one adds templates to achieve trans-lation and size invariance (but see Larsen & Bundesen,1992). The combinatorial explosion can be reduced, how-ever, if we first identify simple features of objects, such astheir orientation and curvature, and then determine the ob-ject's shape from the relative position of the components. Theshape-from-relative-position mechanism requires only asmall alphabet of feature templates and is invariant over sizeand translation.

Early-selection models (including FIT and its variants) addan assumption to the shape-from-relative-position mecha-nism, namely that relative position is determined by atten-tion; that is, although the component features are identifiedpreattentively, their relative position is determined from at-tention's position when it is focused on each component inturn. Localization by attention predicts increasing searchfunctions for targets defined by shape. Note that if positionwere determined preattentively, however, the shape-from-

relative-position mechanism would be consistent with flatsearch functions.

To test the claim that attention is necessary for localization,Treisman and Gelade (1980) required subjects not only tosearch for a target defined by a feature or by a conjunctionof features but also to report the target's location. Search forfeature targets produced pop-out, and search for conjunctiontargets produced increasing search functions. Moreover,analysis of identification conditioned on correct localizationshowed that, if a conjunction target was incorrectly located,its identity was reported at chance. By contrast, correct lo-calization was not associated with correct identification offeature targets. Treisman and Gelade concluded that attentionis involved in conjunction search but not in feature search.

Johnston and Pashler (1990) argued, however, that Treis-man and Gelade's (1980) findings were confounded byguessing strategies and by a loss of location information frommemory. When they removed the confounds, they found astrong association between identification and localization insearch for feature-defined targets. Because Johnston andPashler's evidence suggests that differential localization forfeature and conjunction search does not depend on attention,their results weaken the case for localization by attention.

Localization by attention also predicts that object com-ponents are "free floating spatially" unless fixed by attention(Treisman & Gelade, 1980, p. 126). Treisman and Gelade(1980) tested the free-floating nature of unattended objectcomponents by comparing search in which a target could beproduced by illusory conjunctions of parts of the distractorswith search in which the target could not be formed fromparts of the distractors (but was similar to the distractors inoverall shape). For illusory conjunction displays, they usedan uppercase R target among P and Q distractors. Note thatan illusory R (the target) could be formed by attaching thediagonal tail of the Q to the P. For nonillusory conjunctiondisplays, they used an R target among P and B distractors.

Consistent with the idea that attention is needed to fix afeature's location, search for R among Ps and Qs (i.e., illu-sory conjunction displays) yielded steeper search functionsthan search for R among Bs and Ps, displays that do notafford illusory conjunctions. Note that the two types of dis-plays differ only in using B versus Q distractors.

An alternate explanation is suggested by results from Dun-can and Humphreys' (1989; see also Duncan, 1990) study. Intheir experiments, search was speeded when the distractorswere made more similar to each other. They suggested thatincreasing the similarity among distractors facilitates group-ing and that strongly grouped distractors have less chance ofbeing selected for attentive processing. As a result, the timeto find the target should be reduced as the similarity amongdistractors is increased.

In Treisman and Gelade's (1980) experiment, the distrac-tors in illusory conjunction displays (Ps and Qs) were lesssimilar to each other than were the distractors in nonillusoryconjunction displays (Ps and Bs). Thus, because of weakerdistractor grouping, Duncan and Humphreys' (1989) accountpredicts that search with illusory conjunction displays (i.e.,for R among Ps and Qs) will be harder than with nonillusoryconjunction displays (i.e., for R among Ps and Bs)—the same

490 ANDREW HEATHCOTE AND D. J. K. MEWHORT

prediction derived by Treisman and Gelade on quite differentgrounds. We conclude, therefore, that Treisman and Gelade'sresults do not provide unambiguous support for localizationof object components by attention (see also Butler, Mewhort,& Browse, 1991).

In summary, several considerations cast doubt on evidencefor the localization-by-attention mechanism. Even if local-ization does not require attention, however, shape identifi-cation may still require attentive processing. Hence, the mostdirect test for preattentive shape processing is to examinesearch performance when targets are differentiated from dis-tractors only by the relative positions of their componentfeatures.

Preattentive Detection of Shape

Early evidence from the texture-segregation literature fa-vored the idea that shape cannot be coded preattentively.Beck (1966), for example, demonstrated strong segregationfor regions of texture that differ only in the orientation oftheir constituent elements (e.g., upright Ts vs. tilted Ts) butweak segregation for regions that differ in the position oftheir elements (e.g., Ts vs. +s). If texture segregation is apreattentive process, Beck's demonstration indicates that theposition of components is not coded preattentively.

Similar results occur in visual search. Wolfe et al. (1989)asked subjects to search for a T in one of four rotations amongLs, again in any of four rotations (0°, 90°, 180°, or 270° inboth cases). The targets differed from the distractors only onthe positions of component features. Search functions wereincreasing, and Wolfe et al. concluded that processing ofposition information is capacity limited; that is, attentionmust be involved in localization (also see Beck & Ambler,1972). At first glance, then, the difficulty of search for targetsdefined by the relative positions of their components seemsa remarkably robust finding.

More recent data, however, favor preattentive coding ofshape. For example, Enns and Rensink (1991; see also Enns,1990a, 1990b) found pop-out using targets that differed fromdistractors in apparent three-dimensional properties, such asthree-dimensional orientation (slant) and direction of lightsource. Slant and direction of light source were conveyed bythe relative position of object components; hence, Enns andRensink concluded that the relative position of componentsis encoded preattentively, at least when it corresponds tosalient three-dimensional properties.

Duncan, Humphreys, and colleagues argued that, underappropriate conditions, the relative position of any arrange-ment of lines can be coded preattentively (Duncan & Hum-phreys, 1989; Humphreys, Quinlan, & Riddoch, 1989;Humphreys, Riddoch, & Quinlan, 1985). In their view,shape discriminability controls search performance: Pop-outwill occur whenever shape discriminability is adequate. Be-cause it indexes shape discriminability, the ratio of objectsize to retinal eccentricity (SER) predicts performance insearch for shape-defined targets (shapes at a large SER aremore discriminable than those at a small SER). Hence,pop-out should occur when the objects searched are largerelative to retinal eccentricity at which they are presented.

(The analogous measure for texture segregation—the ratioof element size to element spacing—has similar predictivevalue; see Northdurft, 1985.)

Duncan and Humphreys (1989) investigated search for anL among rotated L and T distractors. Humphreys et al. (1989,1985) investigated search for an inverted T target in rotatedT distractors. They found pop-out when SER was sufficientlylarge. Moreover, the search-function slope increased withdecreasing SER. In addition, pop-out was strongest when thedistractors were homogeneous (i.e., all of the same type),although performance close to pop-out occurred in one casein which the distractors were heterogeneous (Duncan &Humphreys, 1989, Experiment 4).

If Duncan and Humphreys's (1989) analysis is correct,preattentive coding of position is the norm, and previousexperiments failed to find pop-out because they used inap-propriate displays (e.g., heterogeneous distractors and largedisplay areas; Beck, 1966; Wolfe et al., 1989). Because suchdisplays involve small objects presented in large display ar-eas, the targets and distractors were not highly discriminable.

Of course, pop-out with targets defined by shape supportsthe late-selection position because it implies that attention isnot necessary to construct a representation of simple lettershapes. A basic tenet of early selection is that shape cannotbe coded preattentively.

Spatial Scales and Search Stimuli

The finding of pop-out based on the relative position of anobject's components seems to provide a strong falsificationof early-selection theories such as FIT and guided search(Cave & Wolfe, 1990). Both theories claim that shape isderived by the serial operation of attention. Pop-out based onshape differences, even simple shape differences betweenletters, falsifies such models because pop-out denies serialprocessing.

Unfortunately, existing evidence for pop-out based on rel-ative position is not convincing because the stimuli used inthe demonstrations are confounded with simple featureswhose detection should not require processing of positioninformation. Consider an L, a target used in the previousdemonstrations. It contains vertical and horizontal orienta-tions associated with its component lines. To anticipate theargument of the following section, if low spatial frequencycomponents of the L are considered, it also contains an ob-lique orientation. The vertical and horizontal componentsexist at small spatial scales, whereas the oblique componentexists at large spatial scales.

Perceptual processing operates at multiple spatial scales.A role for multiple spatial scales in perceptual processing hasbeen acknowledged by early models based on a global Fou-rier analysis of the image (e.g., Campbell & Robson, 1968).1

1 Fourier analysis transforms an image from the spatial domainto the frequency domain. In the frequency domain, an image can bedescribed by its power and phase spectra. The power spectrumdepends on the amplitude of the frequencies into which the imageis decomposed. The phase spectrum depends on the relative posi-


Psychophysical evidence indicates that local rather than glo-bal Fourier analysis is carried out by the perceptual system(Cavanagh, 1984).2

A potential mechanism by which a local spatial frequencyanalysis might be realized in biological vision has been sug-gested by physiological studies that map receptive fields ofneurons in the visual pathway. Populations of cells have beenfound that, taken together, are capable of discriminating size,orientation, and curvature at multiple spatial scales. We con-sider evidence for each in turn.

Wiesel (1960), for example, found retinal ganglion neu-rons that have circularly symmetrical receptive fields withexcitatory (inhibitory) centers and inhibitory (excitatory)surrounds. Such neurons code an object's size: When a stim-ulus is large relative to the receptive field, some of it will fallinto the inhibitory surround, decreasing the neuron's re-sponse. Likewise, a small stimulus will not produce as biga response as a larger stimulus that falls entirely within theexcitatory region.

Neurons in visual cortex are known to be selective fororientation and line length (e.g., Hubel & Wiesel, 1968).Cells that code orientation have a receptive field elongatedin the direction of its preferred orientation, with a centralexcitatory (inhibitory) region and inhibitory (excitatory) sidelobes. Line-length selectivity is coded by end-stopped cells(e.g., Dreher, 1972; Hubel & Wiesel, 1965). Such cells haveinhibitory regions at one or both ends of long excitatoryregions.

Finally, Dobbins, Zucker, and Cynader (1987) found end-stopped cells that detect curvature. Moreover, they devel-oped a computational model to illustrate how curvature canbe detected using end-stopped receptive fields (Dobbins,Zucker, & Cynader, 1989).3

Variation in the size of receptive fields produces a multiplespatial scale representation of an image. Large spatial scalescan also be coded by cooperation between cells with smallreceptive fields. Nelson and Frost (1985) demonstrated fa-cilitatory interactions between colinear and cooriented sim-ple receptive fields (see also von der Heydt, Peterhaus, &Baumgartner, 1984). Dobbins et al. (1989) suggested thatcells with cocircular, end-stopped receptive fields maysharpen tuning for curvature.

Implications of multiple spatial scales. Both FIT andCave and Wolfe's (1990) guided search model anticipate thattargets defined by size, orientation, and curvature will pop-out (provided, of course, that the feature is discriminable).Although neither model has incorporated representations atmultiple spatial scales, both can be extended to do so byadding feature maps operating at multiple spatial scales. Theextension is consistent with their main principles, especiallytheir early-selection assumptions. Hence, neither model canbe falsified convincingly without taking representation atmultiple spatial scales into account.

With the possibility of confounding at large spatial scalesin mind, consider the stimuli used by Duncan, Humphreys,

tions of the frequency components. Larger spatial scales corre-spond to larger spatial frequencies in Fourier analysis.

and colleagues (Duncan & Humphreys, 1989; Humphreys etal., 1985, 1989), namely Ls and Ts constructed of lines ofequal length (and their 90°, 180°, and 270° rotations). Suchstimuli are matched in the size and orientation of their com-ponents (i.e., vertical and horizontal) but are not matched atlarger spatial scales.

First, the Ls and Ts differed in size. The radius of theminimum circle enclosing an L is 1.13 times larger than theradius of the minimum circle enclosing a T. Hence, an Lamong Ts may be detected by a difference in size at largespatial scales (see Gurnsey & Browse, 1987, for a similaraccount of texture discrimination between regions of +s andLs). Second, Ts and Ls can also be discriminated on thebasis of orientation at large spatial scales: L has a strongorientation at 45° to the left of vertical, whereas T has nostrong orientation.

When targets and distractors differ only by rotation, sizecannot be used for search. Rotations of Ls that differ by 90°,however, can be discriminated by orientation at large spatialscales. For instance, L and J have strong orientations at 45°to the left and right of vertical, respectively. Ls that differ bya 180° rotation have the same large-scale orientation, butthey differ in the sign of curvature: L and ], for example, areconcave (e.g., 1) and convex (e.g., 1 ),4 respectively, at largespatial scales.

Rotations of T can be discriminated by differing sign ofcurvature or by a conjunction of sign of curvature and ori-entation. A T, for instance, has strong convex curvatures atlarge spatial scales with differing orientations (i.e., 1 and \ onthe left- and right-hand sides of the T, respectively). A 180°rotation has strong concave curvatures with the same ori-entations. Ts in 90° and 270° rotations, however, differ in aconjunction of curvature and orientation.

Humphreys et al. (1989) found pop-out using outlinesquares with a line extending from either the top or bottomof the squares. Differences in curvature and in conjunctionsof curvature and orientation—differences that differentiateTs and rotated Ts—also differentiate these stimuli.

In summary, existing evidence for pop-out based on rel-ative position is not convincing because the stimuli used inthe demonstrations are confounded with simple featureswhose detection should not require processing of positioninformation. Moreover, an account based on detection ofsimple features available at low spatial scales is consistentwith the discriminability of the emergent features. Pop-outputatively based on the relative positions of component fea-tures, for example, has been found only when stimulus sizewas large relative to its eccentricity (usually SER is 1:3 or

2 Global Fourier analysis requires integration of informationacross the entire image. By contrast, local analysis integrates oversmall, possibly overlapping, areas of the image.

3 Cells with asymmetrical end-stopped receptive fields can dis-criminate sign of curvature (concave or convex), whereas cellswith symmetrical end-stopped receptive fields discriminate curva-ture but not convexity (concavity).

4 We define a curve as concave if a straight line joining twopoints on the curve lies above the curve; the curve is convex if theline lies below the curve.


1:6). It is precisely under these conditions that orientation,size, and curvature at large spatial scales are most discrim-inable. Hence, the effect of SER is consistent with a spatialscale account. Furthermore, increasing search functionslopes with decreasing SER is consistent with the role oflarge spatial scale features as these features become less dis-criminable as SER decreases.

The finding of pop-out with Ts at different rotations is, ofcourse, the most difficult for a spatial scale account becauseit may require detection of a conjunction of curvature andorientation. Detection of conjunctions of features is not with-out precedent: Wolfe et al. (1989), for example, found pop-out for highly discriminable conjunctions of features. Hence,pop-out of a conjunction of features is not implausible, es-pecially when the features are highly discriminable, as arecurvature and orientation at large spatial scales when SER islarge. Note that Duncan and Humphreys (1989) demon-strated pop-out of inverted Ts only among upright Ts, a casethat can be discriminated purely by sign of curvature.

Humphreys et al. (1989) attempted to rule out certainemergent features as potential confounds in their experi-ments. To test for an upward-pointing terminator feature, forexample, they compared search for an inverted T targetamong homogeneous upright T distractors, among homoge-neous I distractors (formed by a conjunction of a T and in-verted T), and among a mixture of T and I distractors. Pop-outoccurred when each distractor was used separately but notwhen the distractors were mixed.

If pop-out reflects an emergent property in the homoge-neous distractor case, why does it fail in the mixed distractorcase? In particular, in the mixed distractor displays, theupward pointing terminator was a unique target cue; if sub-jects searched using it, pop-out should have occurred.Hence, the failure to find pop-out seems to argue against theidea of an emergent feature. Humphreys et al. used similararguments—arguments that rely on the internal consistencyof results across experiments—to rule out like emergentfeatures.

We agree that Humphreys et al.'s (1989) arguments ruleout the particular emergent features that they tested. More-over, their comparison of homogeneous versus heteroge-neous distractors indicates an important role for interdis-tractor similarity. We note, however, that their arguments donot rule out use of size, orientation, or curvature featuresemerging at large spatial scales, principally because de-creased interdistractor similarity can explain poorer perfor-mance with heterogeneous distractors relative to homoge-neous distractors.

The problem is that both homogeneous displays may offera unique cue—an inverted T can be distinguished from an Iby a size cue, and an inverted T can be distinguished froman upright T by a sign of curvature cue—but when the dis-tractors are mixed, there is no unique cue that can be usedto group targets separately from distractors. The emergentfeatures may not be as useful in the heterogeneous case asin the homogeneous case purely because of decreased in-terdistractor similarity; hence, the failure of pop-out in theheterogeneous case is no argument against their use in thehomogeneous case.

In short, previous demonstrations of pop-out based on therelative positions of components do not falsify FIT andguided search. Shape pop-out is a critical finding not onlyfor FIT and guided search but also for the long-standingearly- versus late-selection debate. Hence, it is important toverify pop-out unconfounded by features available at a largespatial scale.

Experiment 1

Stimuli Without Large Spatial Scale Cues

Figure 1 shows the stimuli used in the present experiments.The target and distractor stimuli cannot be discriminated bysize, orientation, or curvature at any spatial scale. Targetsdiffer from distractors only in the relative position of com-ponent features. In the luminance condition, one half of eachitem was black and the other half was white. In the colorcondition, one half of each item was red and the other halfwas blue. The red and blue areas were approximately equi-luminant. (See Figure 1 for further details of the displays.)

To ensure that subjects use relative position to detect tar-gets, interobject cues were controlled. In the regularly spacednontarget displays, such as shown in Figure 1 a, areas with thesame feature are colinear at positions marked 2 and 8, 3 and7, and 4 and 6. Areas with the same feature are also cocir-cular. Hence, the inclusion of a target in a display may bedetected by violations of colinearity and cocircularity. Thevalidity of cocircularity and colinearity as cues was elimi-nated by randomly perturbing the horizontal position of itemsin the displays.

An additional control was introduced to remove a cue sug-gested by texton theory (Julesz, 1981, 1984, 1986). Textontheory claims that attention is drawn to discontinuities in thefirst-order statistics or density of features. In the present dis-plays, feature density is proportional to the horizontal dis-tance between corresponding points in feature areas. Whenstimuli are positioned regularly around a circle, horizontalinterarea distance is a potential target cue. In Figure Ib, forinstance, the black areas in the target at Position 2 and thedistractor at Position 1 are closer than any of the black areasin the nontarget display in Figure la.

Perturbation of the horizontal position of stimuli reducesbut does not eliminate the density cue. If perturbation is com-pletely random, a target may still introduce a unique interareadistance and, hence, a unique density. To control for the den-sity cue completely, random perturbation was restricted sothat a target could not introduce a unique horizontal interareadistance.

Duncan and Humphreys (1989) found pop-out of shapewhen items in a search display were large relative to theirretinal eccentricity (see also Humphreys et al., 1989). In Ex-periment 1, large items were used to maximize the chance ofpop-out, and SER was either 1:3 or 1:6. SER was manipu-lated by varying the size of items at a fixed eccentricity.Specifically, SER was defined throughout the present workas the ratio of the object's to the average radius of presen-tation. If size (or SER) affects search for relative position inthe same way that it affects search forTs and Ls (i.e., Duncan


LI Jila

LILI

Distractor

I

a~ :

(a) (b)

Figure 1. A description of (a) nontarget and (b) target displays used in Experiment 1. (In thedisplays, the external boundary of the object was defined by a discontinuity with the backgroundinstead of the borders drawn in the figure. In the luminance condition, the background was gray, andthe targets and distractors were black and white. In the color condition, the background was black,the black areas in the stimuli were red, and the white areas in the figure were blue. The numeralsin Figure la did not appear in displays. The stimuli were positioned at approximately 45° intervalsaround an imaginary circle. The figure is not drawn to scale.)

and Humphreys' stimuli), the 1:6 SER condition should beharder than the 1:3 SER condition.

Criteria for Pop-Out

Treisman and Gelade (1980) suggested two criteria foridentifying pop-out: shallow search functions and nonlinearsearch functions. They identify pop-out with preattentiveprocessing and linearly increasing search functions with se-rial, attentional processing: "Rat or markedly bowed RTfunctions are ... difficult to reconcile with serial processing"(Treisman et al., 1977, p. 345), whereas the "diagnostic forserial search is a linear increase in search latency as dis-tractors are added to the display" (Treisman & Souther, 1985,p. 287).

Linearly increasing search functions do not, in fact, implyserial search. Townsend (1971, 1976) showed that limitedclasses of serial and parallel models can mimic each other.Vorberg (1977) and Townsend and Ashby (1983) proved thatwide classes of serial and parallel models can produce equiv-alent predictions. Townsend (1990) summed up the equiv-alence of serial and parallel models as follows:

The phenomena of linear increasing reaction time curves is nolonger considered a fundamental parallel-serial distinction be-cause it simply indicates, first and foremost, a limitation incapacity . . . due to seriality, limited capacity parallel, or evenhybrid processing mechanisms, (p. 47)

There is, therefore, an asymmetry in inference about par-allel and serial models: Flat search functions falsify serialmodels, but increasing search functions cannot falsify par-allel models. Serial models can explain flat search functions

only by accepting an implausible assumption, namely thatthe processing rate increases with display size. As a result,the meaning of a shallow search function for serial modelsdepends on what is considered a plausible rate of attentiveprocessing.

A slope criterion for pop-out is problematic because it re-lies on a lower limit to the plausible rate of attentive pro-cessing. For the present work, we used the slope of the targetsearch function as a criterion for pop-out: We treat a searchfunction slope of less than 10 ms per object as an indicatorof pop-out. We used the slope of target search functions be-cause of the greater variability of nontarget search functions.The same criterion has been adopted by other investigatorsof visual search (e.g., Enns & Rensink, 1991; Wolfe et al.,1989).

Our criterion for pop-out relies on the implausibility ofprocessing at a rate faster than 10 ms per object. If we assumeexhaustive processing of target displays, a slope of 10 ms perobject maps directly on the rate of processing. If we assumea self-terminating model, the same slope doubles the rate, buteach serial step now must include a terminate-or-continuedecision. The important point for the present purpose is thatour criterion is conservative relative to empirical estimatesof processing times derived from other paradigms (Briand &Klein, 1987; Colgate, Hoffman, & Eriksen, 1973; Eriksen &Collins, 1969; Eriksen, Webb, & Fournier, 1990; Jolicoeur,Ullman, & MacKay, 1986; Klein & Briand, 1986; Krose &Julesz, 1989; Sagi & Julesz, 1985; Tsal, 1983).

Finally, Wolfe et al. (1989) noted that search functionslopes vary widely between subjects in the same experimen-tal conditions. Because the slope of the search function isconsistent within each subject, the variation does not appearto be due to random factors. Therefore, we adopt Wolfe etal.'s practice of reporting slopes based on linear regressionfor individual subjects.


Method

Subjects. Twenty-four students from Queen's University par-ticipated in two 1 -hr experimental sessions on successive days. Sub-jects were paid $10 for their participation. All subjects had normalor corrected-to-normal vision.

Apparatus and stimuli. Stimuli were presented on a ZenithData Systems ZVM-1330 color monitor controlled by a ZenithZ-158 PC. Responses were obtained using two buttons. Millisecondtiming and the synchronization of stimulus presentation with screenrefresh were achieved using Heathcote's (1988) programs. Viewingdistance was 1.0 m. Room illuminance was 15 Ix. Trials were ini-tiated by pressing a foot pedal, and responses were made by pressingone of two buttons. The subjects were encouraged to think of thetask as a video game, specifically one involving competing space-ships. The button used for target responses was marked "FIRE," andthe button used for nontarget responses was marked "SHIELDS."The two labels refer to firing at enemy ships and lowering shieldsto save friendly ships, respectively.

Stimuli were squares divided vertically into equal black andwhite areas in the luminance condition and equal red and blue areasin the color condition. Targets had blade/red to the left and white/blue to the right. For distractors, the relative horizontal positions offeatures were reversed (see Figure 1). Stimulus side lengths wereapproximately 12 mm (0.69° of visual angle) and 6 mm (0.34°) inthe 1:3 and 1:6 SER conditions, respectively.

In the luminance condition, the luminances of the gray back-ground and the black and white stimulus areas were 40 cd/m2, 0.6cd/m2, and 125 cd/m2, respectively. In the color condition, the back-ground was black (0.6 cd/m2). The colors used were highly dis-criminable and approximately equiluminant (27 cd/m2 for red and30 cd/m2 for blue).

Stimuli were presented on adjacent positions of an imaginarycircle with eight possible positions at approximately 45° intervals(see Figure 1; note that Figure 1 is not drawn to scale). The radiusof the circle was 36 mm (2.06°). On each trial, the positions ofstimuli were perturbed randomly -3, 0, or 3 mm horizontally fromtheir position on the circle. The perturbations were constrained toeliminate feature-density cues.

Design. A 4 X 2 X 2 X 2 factorial design was used. Within-subjects factors were display size (two, four, six, or eight objects)and target presence versus absence. Between-subjects factors werefeature type (luminance or color) and SER (1:3, 1:6), with 6 subjectsper condition.

Procedure. At the start of the first session, subjects read thefollowing instructions:

Fighter Ship Training Simulation: Cadet LevelYou are piloting a fighter ship with hyperspace capabilitiesand a computer targeting system. After a X appears at thecenter of the screen, you may come out of hyperspace bypushing the foot pedal (remember to release it after pressingit). You will be surrounded by a number of ships (from two toeight). The ships may either be all friends, or there may be oneenemy. Your computer targeting system will automaticallypick the most likely target, but, like most computers, it is fairlystupid and cannot tell whether its target is really an enemy ora friend. You must make that decision by pressing the buttonmarked "FIRE" or the button marked "SHIELDS." If youpress "SHIELDS," your shields will be lowered and you maysave the friendly ships. If an enemy is present and you loweryour shields, you will be hit.

You should make your choice as quickly as you can whilemaking as few errors as possible. If you fire when no enemyis present, you will shoot a friend. If you lower the shields

when an enemy is present, you will be shot. Either way, it isnot going to look good on your record. Each game will consistof about 80 encounters. At the most, make only eight errors!:However, also try to respond as fast as you can safely becausespeed is also important! At the end of each game, you will begiven your ratings. For each error you will lose three points;for each correct choice you gain one point. Your score for eachgame is the sum of your points divided by your averagereaction time. At the end of your mission, you will be givenyour total score.

Before you begin to play, you will do some practice. First, youwill practice recognizing enemies and friends. Only one shipwill appear on each encounter. The object is to get 20 correctin a row. Do not worry about speed at first. Then you will dothe same again, but now you have to decide in less than 1 s.Next you will play a series of games.

Note from the technical division: It has been noted that betterresults are obtained if you do not move your eyes from theposition of the X before making your decision. If a target ispresent, you will tend to look at it after making a decision.This is okay, but try to minimize eye movements otherwise.Do not worry about this too much on the first practice session,but thereafter try to minimize your eye movements.

After the subject had read the instructions, the screen displayeda picture of a target and a distractor. While the subject examined thedisplay, the experimenter described the task and answered ques-tions. The experimenter explained speed-accuracy trade-off (SAT)and stressed that errors were to be kept to a minimum, while main-taining fast performance. Subjects were instructed to use one handfor target responses and the other hand for nontarget responses. Halfof the subjects used their dominant hand for target responses andhalf used their nondominant hand. Subjects were encouraged to takebreaks as often as they wished.

In the first session, subjects performed two practice blocks. Onlyone object was presented on each practice trial. To complete a prac-tice block, subjects had to produce 20 consecutive correct respons-es. In the first practice block, there was no time limit on responses.•In the second practice block, however, responses had to be made inless than 1 s. The 20 consecutive correct responses criterion wasimposed to encourage accurate responding. The response time limitin the second block was imposed to accustom subjects to speededresponding.

After completing practice, subjects performed eight blocks ofexperimental trials. On the second day, no practice and 12 blockswere performed. Error responses were followed by unrecordeddummy trials. Experimental trials resumed when a correct responsewas made on a dummy trial. All experimental error trials wererepeated later in the block. Hence, each experimental block con-sisted of 80 correctly answered trials, 10 from each within-subjectscondition in random order.

Subjects initiated a trial by pressing a foot pedal. After each trialhad been initiated, the pretrial display (a fixation cross) remainedon the screen for 250 ms and was replaced by the search display.When a subject pressed the target button, a "happy face" character,representing a missile, moved toward an object, and that objectflashed briefly. When the target response was correct, the objectselected was the target. When the target response was incorrect, theobject selected was a randomly chosen distractor. A correct responsewas signaled by an alternating 1000/1100-Hz tone on target trialsand signaled in nontarget trials by a tone increasing in steps of 100Hz from 700 to 1000 Hz. Incorrect responses were signaled by analternating 50/100-Hz tone.


At the end of each block, subjects were given their score and thenumber of correct and incorrect responses to target and nontargettrials. The score was designed to stress both speed and accuracy.Each correct response gained a point; each error response lost threepoints. The overall score was the total number of points divided bythe average response time in seconds.

Results

Mean correct RT. Analysis of variance (ANOVA) wascarried out on mean correct RT (MRT) to examine the effectof the two between-subjects factors—feature type (lumi-nance and color) and SER (1:3 and 1:6)—and the threewithin-subjects factors—display type (target and nontarget),display size (two, four, six, and eight objects), and practice.The practice factor was created by dividing experimentaltrials into five groups containing four blocks each. The firsttwo groups came from the first session, and the last threegroups of blocks came from the second session. The resultsare summarized in Figure 2.

MRT, averaged over the within-subjects factors, was 668ms and 731 ms for luminance condition at 1:3 and 1:6 SER,respectively. The corresponding values for the color condi-tion were 826 ms and 771 ms, respectively. Neither between-subjects main effects were significant: feature type, F(l, 20)= 3.03, p > .05; SER, F(l, 20) < 1, and there were nosignificant interactions involving the between-subjects fac-tors (ps > .1).

From Duncan and Humphreys' (1989) work, we antici-pated an increase in MRT with increased SER. The trend inthe means was consistent with the predicted effect for theluminance condition but not for the color condition. Becausethe trends were in opposite directions, the anticipated effectof SER may have been masked by averaging across featuretype. Although the null interaction of SER with feature typeargues against such a possibility, the data were reanalyzedhierarchically. In the hierarchical analysis, the main effect ofSER in both the feature conditions remained nonsignificant(F < 1 in both cases). We conclude that the SER manipu-lation was ineffective.

MRT increased with increasing display size, F(3, 30) =63.3, p < .001, and was greater for nontarget displays thanfor target displays, F(l, 20) = 57.6, p < .001. Finally, MRT

decreased with practice, F(4, 80) = 95.8, p < .001. Theseeffects are discussed in detail in connection with the analysesto follow.

Linear regressions. Table 1 presents the slopes producedby linear regression of MRT on display size. Regressions werecalculated separately for each subject at each level of practicefor both target and nontarget displays. The table also presentsslopes from regressions calculated on the data averaged oversubjects. The latter regressions reflect the A/RT presented inFigure 2.

The data, averaged over subjects, exhibit the same patternin all between-subjects conditions. Search functions wereinitially increasing, even in the SER 1:3 conditions. Withpractice, however, both the slope and the intercept of thesearch functions decreased. Pop-out, according to the 10-ms

per item target slope criterion, occurred in the last three prac-tice levels of the 1:3 SER luminance condition and in the lastpractice level of the 1:6 SER color condition.

Regressions for each subject indicated that some subjectsin each between-subjects condition produced pop-out. Threeor more subjects of the six had target search slopes less than10 ms per object in at least one practice level in all between-subjects conditions. In the first practice level, Subject 5 in the1:3 SER luminance condition exhibited pop-out. Subjects 2and 5 in the 1:3 SER luminance condition and Subjects 3, 5,and 6 in the 1:6 SER color condition exhibited negativeslopes! In short, pop-out occurred in all between-subjectsconditions, even those that did pass the criterion in the re-gressions on the mean data.

Error frequency and SAT. A five-factor ANOVA wasperformed on percentage of error. Neither the feature typenor the SER effect (the between-subjects manipulations)were significant, F(l, 20) < 1 for both. The three within-subjects factors were highly significant: Errors increasedwith increased display size, F(3,60) = 43.9, p < .001. Therewere more errors with target displays than with nontargetdisplays, F(l, 20) = 65.7, p < .001, and errors decreasedacross practice, F(4, 80) = 10.8, p < .001.

With one exception, the main effects for errors were par-allel to the corresponding effects for A/RT. The exception wasdisplay type: Nontarget displays took longer to process butproduced fewer errors than target displays. Hence, the dif-ference in MRT for target and nontarget displays may be con-taminated by an SAT (see Pachella, 1974).

An explanation for pop-out in terms of SAT implies thatthe increase in percentage of error with display size shouldbecome steeper as the slope of the search function decreaseswith practice. In fact, however, the increase in percentage oferror with display size became less steep with practice,F(12, 240) = 2.8, p < .01. The difference was stronger fortarget than for nontarget displays; Display Size X DisplayType X Practice, F(12, 240) = 3.8, p < .001. Thus, it isclear that pop-out did not reflect SAT for the critical targetcase. The ANOVA confirms the picture shown in Figure 2:Practice decreased the effect of display size both on MRT

and on errors.Probabilistic distance cues. As noted earlier, random

perturbation of object positions was restricted so that a targetdisplay could not contain a unique horizontal interarea dis-tance between objects in adjacent positions. Some interareadistances, however, remained more probable in a target dis-play, as illustrated in Table 2. In particular, distances of 0, 12,and 24 mm for the 1:3 SER conditions and 3, 15, and 27 mmfor the 1:6 SER conditions were more probable in targetdisplays. Although the probabilistic cues were small, subjectsmay have learned to take advantage of them through practice.

To determine whether or not probabilistic distance cuesmediated pop-out, trials in which targets generated the moreprobable distances were compared with those in which theydid not. In all cases, MRT from trials with less probable dis-tances was less than those with more probable distances;hence, pop-out was not mediated by probabilistic distancecues.


(a)

2 4 6 8 2 4 6

HOS

«

S

o

1500,

1000

500

15105

ra = 0.99

rp = 0.99

r, = 0.97

rp = 0.97

(b)

2 4 6 8 2 4 6 2 4 6 2 4 6 2 4 6

L.O!_

W

1500

1000

500

15105 <•»—e- t

_ 2= 0.95 ^=0.98

rp=1.00

2ra =

. 2

(c)

2 4 6 8 2 4 6 8 2 4 6 8 2 4 6 2 4 6

ecB

1500

1000

500

15105

2ra =

• 2ra = 0.99

8s

ra = 0.93

rp = 0.74

PHI mi .q.r

(d)

2 4 6 8 2 4 6 2 4 6 8 2 4 6 8 2 4 6

Display Size

Figure 2. Mean correct response time (RT; upper panels) and percentage of errors (lower panels)from Experiment 1 for each level of practice (four blocks per level, practice increasing from left toright panels) in (a) the 1:3 ratio of object size to retinal eccentricity (SER) luminance condition, (b)the 1:6 SER luminance condition, (c) the 1:3 SER color condition, and (d) the 1:6 SER colorcondition. (Dashed lines represent data from target-absent displays, and solid lines represent datafrom target-present displays, r^ and i2

u are the squared correlation coefficients from linear regressionson target-present and target-absent RT data, respectively.)


Table 1Slopes (Milliseconds per Object) From the Linear Regression of Mean Correct Response Time on Display Size inExperiment 1

Practice level

Target display

Subject

123456

Combined

123456

Combined

1

49302

22753235

293528643412853

2

714715412318

203312424110141

3

169-7141389

1414167244320

4

1:3 SER534712127

1:6 SER8231615172117

5 1

Nontarget display

2 3 4 5

luminance condition2

-1-5714135

667912381447669

7551716

1263251

3724

113762730

36-90768718

84

-1-2471111

luminance condition31851161810

647919

1311919685

68797

1062314872

3252949157739

1344142073021

194610811416

1:3 SER color condition123456

Combined

47667584646867

231865107353547

10173791223034

30

2167182322

1016734111015

91401269613010298

512

100135868076

291568105446655

5-320110415338

-2929109403436

1 :6 SER color condition123456

Combined

58304874328054

5927628155632

4617132618020

3422-07-5310

18261

-2-1955

90401529558182103

444245645113664

99143761306751

7126724202528

503-5326-813

Note. SER = ratio of object size to retinal eccentricity. Regressions were performed on data from individual subjects and on data obtainedby combining subjects within each between-subjects condition for each level of practice. There were 4 blocks of 80 trials per level ofpractice.

If subjects had to learn to use probabilistic distance cues,their effect should be more evident later in practice. Analyseson trials from the last four blocks, however, confirmed theoverall pattern of results: With one exception, MRT was fasterfor the less probable distances than for the more probabledistances. The exception was the 1:6 SER color condition;here, however, the facilitation was not significant, F(l, 5) =1.4, p > .25. Thus, even after practice, pop-out was not me-diated by probabilistic distance cues.

A final point before leaving the issue: Targets presented inthe diagonal positions (1, 3, 5, and 7) produced the moreprobable distances on all trials, whereas targets presented inthe remaining positions generated the more probable dis-tances on only 45% of trials. If subjects were slower whenthe target was in a diagonal position, the previous analysescould be biased. A supplementary analysis excluding trialswith targets in the diagonal positions showed no benefit asa result of the probabilistic distance cues.

Discussion

The results indicate clearly that relative position can yieldpop-out after practice. Improvement is rapid, requiring only1,600 trials to change the search function slope from a steeplyincreasing slope to a flat one. The decrease in the searchfunction slope cannot be explained in terms of SAT or ofsubjects' learning of probabilistic distance (density) cues.

A possible but trivial explanation for the null effect offeature type is that subjects detected targets in the color con-dition using luminance cues. Previous research has shownthat search function slope increases as both luminance andcolor-defined targets become less discriminable from dis-tractors (Duncan, 1990; Treisman & Gormican, 1988). Thestimuli in the color condition were made from highlydiscriminable colors, whereas the luminance difference be-tween the colors (3 cd/m2 or 10% contrast) was small. Ifsubjects used luminance cues in the color conditions,


Table 2Probability Associated With the Horizontal Distance Between the Closest Edges of Adjacent Areas With the SameFeature Type in Experiment 1

Distance (mm)

Target

AbsentPresentDifference

0/3

.056

.072

.016

3/6

.111

.102-.009

6/9

.167

.151-.016

9/12

.111

.102-.009

12/15

.111

.144

.033

15/18

.111

.102-.009

18/21

.167

.151-.016

21/24

.111

.102-.009

24/27

.056

.072

.016Note. Distances are specified in millimeters for ratios of object size to retinal eccentricity of 1:3/l :6. The average number of occurrencesof a distance in a display is probability x display size x 2.

performance should have been better in the correspondingluminance condition because its cues were more discrim-inable. The equivalence of performance in the color and lu-minance conditions argues, therefore, that luminance cuesdid not mediate pop-out in the color condition.

Granting that subjects did not use luminance cues in thecolor condition, the equivalence of performance in the twoconditions is important because it rules out accounts forperformance in the luminance condition that do not appealto relative position information. It is possible, for example,to devise an account of pop-out based on luminance differ-ences without appealing to relative position information(indeed, we consider some after presenting the next exper-iment). Such accounts fail, however, when applied to thecolor condition. Thus, the equivalence of performance in thetwo conditions reinforces our emphasis on relative positioninformation.

Because we found a null effect of SER, our results differfrom those of Duncan, Humphreys, and their co-workers(e.g., Duncan & Humphreys, 1989; Humphreys et al., 1985,1989). Two potential reasons come to mind. First, the nulleffect of SER may reflect the small range over which it wasvaried. Because our stimuli were constructed from widelines, they may have been too discriminable at both 1:3 and1:6 SER. A more interesting alternative concerns the type ofinformation that mediated pop-out in the two cases. As wesuggested previously, T and L stimuli can be discriminatedby large spatial scale size, orientation, and curvature cues,whereas our stimuli can be discriminated only by relativeposition. SER may affect the two types of information dif-ferently. Experiment 2 was designed to resolve the issue.

Experiment 2

Humphreys et al. (1989) showed that SER governs thediscriminability of their stimuli. Decreased discriminabilitybecause of eccentric presentation was corrected by increas-ing the size of the stimuli to equate SER. Similar effects ofsize and eccentricity on recognition thresholds are found, forexample, in letter discrimination, contrast discrimination,and grating discrimination (e.g., Anstis, 1974).

Is the discriminability of our stimuli determined by SER?The first experiment found a null effect of SER, and otherevidence is consistent with that finding. Position acuity—judgments of the relative positions of two features—isamong the strongest hyperacuities (e.g., a minimum resolu-tion down to 5 s of arc, much less than the average width of

a photoreceptor). Over the same range of eccentricity, how-ever, position acuity decreases more rapidly (by a factor oftwo) than contrast and grating acuities; that is, position acu-ities are very strongly affected by retinal eccentricity (e.g.,Westheimer, 1982).

In experiments requiring discrimination of two sinusoi-dally modulated luminance profiles, phase acuity is alsostrongly affected by retinal eccentricity (P. J. Bennett &Banks, 1987;Rentschler&Treutwein, 1985). Rentschler andTreutwein (1985) found, for example, that discrimination ofmirror image gratings decreased strongly with retinal ec-centricity. The decrease in discriminability with eccentricitywas greater than the decrease for gratings differentiated bycontrast cues (i.e., that were not mirror symmetrical). Rent-schler and Treutwein argued that the mirror-symmetricalgratings were discriminated by phase differences and, there-fore, that phase acuity decreases with eccentricity more rap-idly than contrast acuity.

When the stimuli were composed of a number of periodicfunctions, however, Morrone, Burr, and Spinelli (1989)found that phase sensitivity scaled with retinal eccentricityin the same way as contrast sensitivity. In their stimuli, phasechanges altered the nature of local features (from bars toedges and from edges to bars) but did not change their rel-ative position. Morrone et al. suggested that "the supposedpoor peripheral phase resolution reported by previous in-vestigators is more a consequence of positional uncertaintyrather than a deficit in phase sensitive mechanisms" (pp.442-443).

In summary, if subjects use relative position cues to per-form search with our stimuli, evidence from threshold tech-niques suggests that performance will be degraded by in-creased retinal eccentricity, even when the stimulus size isscaled to maintain a constant SER.

Experiment 2 was designed to consider the effect of retinaleccentricity unconfounded by SER. In the experiment, weused the same kind of stimuli as in Experiment 1 but doubledthe stimulus size and retinal eccentricity relative to the 1:6SER condition of Experiment 1. If SER determines the dis-criminability of the stimuli, performance in Experiment 2should be equivalent to performance in the 1:6 SER condi-tions of Experiment 1. If scaling to maintain a constant SERdoes not compensate for increased retinal eccentricity, how-ever, performance should be slower in Experiment 2 than inthe 1:6 SER conditions of Experiment 1. Such a findingwould indicate that the discriminability of our stimuli—


unlike those used by Duncan, Humphreys, and co-workers—is not determined by SER alone.

Method

The method was identical to Experiment 1, except that the radiusof the imaginary circle on which stimuli were presented was in-creased from 36 mm to 72 mm (4.12°). The stimuli were the sameas those used for the 1:3 SER conditions in Experiment 1 (12 mm/0.68° wide squares). Twelve students from Queen's University par-ticipated in the study. Six subjects participated in the luminancecondition and 6 in the color condition. Subjects were paid $10 fortheir participation. All subjects had normal or corrected-to-normalvision.

Results

Figure 3 shows A/RT for correct trials and percentage oferror as a function of display size and display type for boththe color and luminance conditions.

Correct RT. ANOVA was carried out on A/RT from cor-rect trials to examine the effect of feature type (luminance orcolor, administered between subjects), display size, display

type (target or nontarget), and practice (administered withinsubjects). The practice factor was created in the same way asin Experiment 1.

The main effect of feature type was not significant (F <1), and, with one exception, there were no interactions withfeature type (ps > .05). The exception was interaction ofdisplay type with display size and practice: Practice de-creased the slope of search functions more quickly in thecolor condition than in the luminance condition, F(\2, 120)= 2.5, p < .05.

The main effects of all the within-subjects factors werehighly significant. MRT increased with increasing displaysize, F(3,30) = 40.1, p < .001, and was greater for nontargetthan for target displays, F(l, 10) = 25.0, p < .001. Finally,MRT decreased with practice, F(4,40) = 36.0, p < .001. Thelatter effects are discussed in detail in connection with linearregression analyses.

Linear regressions. Table 3 presents the slopes producedby linear regression of A/RT on display size. For each subject,separate regressions were calculated at each level of practicefor target and nontarget displays. Table 3 also presents sloperegressions calculated on the average over subjects. Theselatter describe the correct RT data presented in Figure 3.

E•>~s

H

I

o

W

1500 -

1000 -

(a)

2 4 6

so

v

i.Ou

1500

1000

500

15105

i i I

0.98 = 0.99

2 4 6 8 2 4 6 8 2 4 6 8 2 4 6

Display Size

2 4 6

(b)

Figure 3. Mean correct response time (RT; upper panels) and percentage of errors (lower panels)from Experiment 2 for each level of practice (four blocks per level, practice increasing from left toright panels) in the (a) luminance condition and (b) color condition. (Dashed lines represent datafrom the target-absent displays, and solid lines represent data from the target-present displays, r, andrj are the squared correlation coefficients from linear regressions on target and nontarget RT data,respectively.)


Table 3Slopes (Milliseconds per Object) From the Linear Regression of MKT on Display Size in Experiment 2

Practice level

Target display

Subject

123456

Combined

123456

Combined

1

1153255102839280

9946110781442283

2

1213332496710368

5247431101071462

3

11328683546458

37558360656

51

4

8815347793544

1410415817324

5

Luminance6516349913343

Color4-6245418016

1

140963778103194108

20313219917627159173

Nontarget display

2

16478398996222115

10810114419126715

138

3

1833898411312792

7711219418917010

125

4

182351671999182

2651177130125786

5

119542470

1 1 14170

161110611986757

Note. Regressions were performed on data from individual subjects and on data obtained by combining subjects within each between-subjects condition for each level of practice. There were 4 blocks of 80 trials per level of practice.

Slopes derived from data averaged over subjects showedclear evidence of learning, but none were below the pop-outcriterion in either luminance or color conditions. The slopefor the color condition at the final level of practice was closeat 15.8 ms/item.

Slopes derived for individual subjects showed large vari-ability: Subject 3 in the luminance condition exhibited pop-out consistently after the second practice level. Subject 6 inthe color condition also exhibited pop-out after the secondpractice level; Subjects 1 and 2 in the color condition ex-hibited pop-out in the last practice level.

Error frequency and SAT. The main effect of feature typeon percentage of error was not significant (F < 1); themeans of the luminance and color conditions were 4.8% and4.2%, respectively. Feature type did not interact with anyother variable (ps > .05). The null difference in MRT be-tween luminance and color conditions, therefore, was notdue to SAT.

The main effect of all within-subjects factors on percent-age of error was significant. Percentage of error increasedwith display size, F(3, 30) = 9.4, p< .001, and was greaterfor target than for nontarget displays, F(l, 10) = 17.3, p <.001. Finally, errors decreased with practice, F(4,40) = 10.6,p < .01.

As in Experiment 1, the main effects of display size andpractice paralleled the corresponding effects on MRT. Non-target displays took longer to process but produced fewererrors than target displays. Hence, the difference in MRT fortarget and nontarget displays may be contaminated by SAT.

As we noted earlier, SAT can account for pop-out if theincrease in percentage of error with display size becomessteeper as the slope of RT search functions decreases withpractice. Although the interaction of display size with prac-tice was not significant, F(12, 120) = 1.5, p > .15, there wasa significant interaction of display size and display type with

practice, F(12, 120) = 3.9, p < .001. The three-factor in-teraction indicates that, for target displays, the increase inpercentage of error with display size became shallower withpractice, denying an explanation of pop-out by SAT. In-spection of Figure 3 indicates that, for nontarget displays,errors tended to decrease with display size, denying an ex-planation for the effect of practice with nontarget displays interms of SAT.

Probabilistic distance cues. Analyses of distance cues,described for Experiment 1, were performed to determinewhether or not probabilistic distance cues confounded thepresent results. As in Experiment 1, there was no evidencefor confounding (all ps > .25).

Comparison with Experiment J. Experiment 2 was con-ducted to consider whether retinal eccentricity affects searchperformance independently of SER. The effect of eccentric-ity with SER held constant was assessed by comparing per-formance in the present experiment with that from the 1:6SER condition in Experiment 1. For the ANOVA, thebetween-subjects factors were eccentricity (2.06° in Exper-iment 1 and 4.12° in Experiment 2) and feature type. Thewithin-subjects factors were display size, display type, andpractice.

MRT was greater in the large eccentricity condition (953ms) than in the small eccentricity condition (751 ms), F( 1,20) = 7.0, p < .05. The difference between the luminance(860 ms) and color (843 ms) conditions was not significant(F < 1) and did not interact with eccentricity (F < 1).

The advantage in MRT for target over nontarget displayswas larger in the large eccentricity condition than in the smalleccentricity condition, F(l, 20) = 6.3, p < .05. Moreover,the slope of search functions was steeper in the large ec-centricity condition than in the small eccentricity condition,F(3, 60) = 9.0, p < .005. Finally, the rate of decrease in theslope of the search functions across practice was greater in


the small eccentricity condition than in the large eccentricitycondition for luminance stimuli, F(3, 60) = 5.3, p < .005.No other interactions involving eccentricity were significant(ps > .35).

There was no significant effect of eccentricity or featuretype on percentage of error (F < 1 in both cases). There wasonly one interaction with eccentricity: Percentage of error inthe large eccentricity conditions decreased more quickly withpractice than in the small eccentricity condition, F(4, 80) =3.2, p < .05. SAT, therefore, cannot explain the differencein speed between the large and small eccentricity conditions.

Discussion

Although increased retinal eccentricity made discrimina-tion of relative position harder than in Experiment 1, strongpractice effects still occurred, and some subjects achievedpop-out. Neither the practice effect nor the decrease in per-formance (caused by increased retinal eccentricity) can beexplained in terms of SAT or probabilistic distance (feature-density) cues.

Although SER was the same (1:6) in both experiments,performance was slower in the present experiment than inExperiment 1. The difference reflects the larger eccentricityin the present experiment: Performance with our stimuli wasnot equated by scaling their size to equate SER.

As we noted earlier, evidence from threshold experimentssuggests that discrimination of relative position is stronglyaffected by retinal eccentricity. It is also worth noting thatsensitivity (measured by d') for position acuities, in contrastto other acuities, improves with practice (McKee & West-heimer, 1978; see R. G. Bennett & Westheimer, 1991, for arecent review). The strong effects of both eccentricity andpractice in Experiments 1 and 2 are, therefore, consistentwith detection by relative position.

By contrast, the experiments by Duncan, Humphreys, andco-workers used different stimuli and found that SER con-trolled the discriminability of their stimuli. Further evidenceis needed to confirm the idea, but the different effect of ec-centricity in the two situations suggests that detection is me-diated by different information.

Theoretical Implications of Experiments 1 and 2

Because relative position is the sole basis of discriminationwith the present stimuli, Experiments 1 and 2 provide a directtest of the position-from-attention mechanism associatedwith FIT and guided search. Both models assume that po-sition is derived from attention, and, for that reason, neithermodel is consistent with pop-out based on relative positioninformation. Both acknowledge that practice may speed themovement of attention and, hence, may speed the process ofcomparison. Nevertheless, for both models, attention re-mains inherently serial, and because it implies that the speedof attentional scanning is too fast to be plausible, pop-outwith our stimuli implies that relative position is availablepreattentively.

Before we conclude against FIT and guided search models,we consider some counterarguments. Pop-out after practice

might be explained within FIT or guided search, for exam-ple, if we assume that practice yields rapid adaptation in apreexisting dimension. Learning an entirely new dimensionseems implausible, however, because search functionschanged from steeply increasing to flat within only twosessions.

Enns and Rensink's (1991) finding of pop-out of directionof light source provides a candidate dimension for rapidadaptation. Their stimuli were hexagons formed of black,gray, and white areas consistent with a projection onto twodimensions of a three-dimensional cube. The three areascorresponded to three faces of a cube. Targets and distrac-tors differed in the relative position of the three areas. Thestimuli used in our luminance condition could be interpretedas a projection of a cube showing only two faces, withtargets lit from the right and distractors from the left. If so,subjects might, with practice, have learned a three-dimen-sional interpretation.

Learned pop-out in the color conditions, however, arguesagainst such an account. A red/blue-colored light source doesnot produce a blue/red-colored shadow; therefore, the di-rection of the color boundary is not a cue to direction of thelight source. Hence, the color condition cannot be explainedby a direction-of-light feature. Parsimony and the null dif-ference in performance between luminance and color con-ditions suggest that the same mechanism should apply to bothconditions. Hence, FIT and guided search cannot escape theimplications of pop-out by appealing to preattentive codingof the direction-of-light cue.

Direction of contrast is another candidate dimension withwhich to explain pop-out. Burr, Morrone, and Spinelli (1989)presented psychophysical evidence for two classes of de-tectors in human vision with odd and even symmetrical re-ceptive fields.5 An even symmetrical receptive field, or linedetector, has peak excitation (inhibition) in its central regionand symmetrical flanking areas of inhibition (excitation). Anodd symmetrical receptive field or edge detector has exci-tation (inhibition) on one side and inhibition (excitation) onthe other. An edge detector, therefore, is sensitive to directionof contrast.

The stimuli used in our luminance conditions could bediscriminated by edge detectors. An appropriately positionedreceptive field with excitation on the right and inhibition onthe left would respond to a target but not a distractor. Withpractice, subjects may be able to learn to isolate the outputof such an edge detector.

Although an explanation based on edge detection mightwork in the luminance condition, it is less plausible in thecolor conditions. Michael (1978) found complex cells inmonkey striate cortex that responded strongly to opponent(red-green) color edges. We know of no evidence, however,for edge detectors based on nonopponent chromatic con-trasts, such as those used in the present experiments; As be-fore, parsimony and the null difference between the lumi-nance and color conditions argue against an explanation by

5 Receptive field mapping of cat and monkey striate cortex,however, indicates a continuum between odd and even symmetry(Kato, Bishop, & Orban, 1978; Parker & Hawken, 1988).


direction of contrast. Again, FIT and guided search cannotescape the implications of pop-out by appealing to preat-tentive coding of direction of contrast. In summary, then, thepresent experiments imply that the relative position of bothluminance and chromatic features is coded preattentively.

Implications for Theories of Texture Segregation

Texton theory (Julesz, 1981, 1984, 1986) is based on ev-idence from texture-segregation experiments. In this theory,features are called textons and, like FIT and guided search,texton theory claims that the relative position features mustbe derived using attention. In Rentschler, Hubner, and Cae-lli's (1988) words, "the essence of texton theory is the dis-tinction between two modes of visual processing [attentiveand preattentive] with respect to their susceptibility to po-sitional information" (p. 289).6

Easy segregation of texture and pop-out in visual searchhave often been taken as complementary evidence for pre-attentive representation (Treisman & Gelade, 1980; Treis-man & Patterson, 1984). Texton theory predicts that stimulithat have identical second-order (dipole) statistics cannot bediscriminated preattentively. Because they are mirror sym-metrical, the targets and distractors in the present experi-ments have identical second-order statistics. Hence, pop-outwith our stimuli argues against texton theory.

Other exceptions have been found to the rule that stimulithat do not differ in second-order statistics are not preatten-tively discriminable (e.g., Caelli & Julesz, 1978, 1979). Theother exceptions can be explained by local feature detectorsencoding properties such as colinearity and orientation. Thestimuli used in the present experiments, however, cannot bedifferentiated by such cues.

Our results conflict with texture-segregation data fromRentschler et al. (1988). Their stimuli were periodic gratingpatches made from compound Gabor signals. In commonwith our stimuli, their targets and distractors were mirrorimages that differed in phase by 180° but had identicalpower in the frequency domain. In the spatial domain, theirtargets and distractors differed in the relative position offeatures (elongated blobs of light and dark) but not in con-trast magnitude.

Rentschler et al. (1988) found that the regions of mirrorimage textures produced poor segregation and concluded that"neither positional relationships within . . . luminance pro-files nor phase relationships . . . are directly registered" (p.289). Instead, they proposed that position information is ob-tained by combining the output of odd and even symmetricalfilters and that position information is not available to pre-attentive vision, because the outputs of the two filters arecombined by attention (see also Rentschler, 1985; Rentschler& Treutwein, 1985).

Although our stimuli have properties similar to those usedby Rentschler et al. (1988)—the targets and distractors havemirror symmetry, differ in the spatial domain only in therelative position of features, and differ in the frequency do-main only in phase—they produced pop-out. Hence, our re-sults suggest that position information is available to preat-tentive vision. We suspect that the difference reflects the

relative discriminability of the stimuli in the two studies:Specifically, we suggest that they did not find easy texturesegregation because their stimuli were not discriminableenough.

Implications for Theories of Learning in Search

The results show strong practice effects. Although practiceeffects are common in visual search, they are often ignored.Much of Duncan and Humphreys' (1989) data, for example,came from a final session that was preceded by several prac-tice sessions. Perhaps because their bias is based on a per-spective derived from perceptual theory, none of the theoriesreviewed so far predict learning in visual search. In this sec-tion, we consider a theory of search based on performancetheory in which learning plays a central role, namely Shiffrinand Schneider's (1977) controlled and automatic processingtheory (CAPT).

CAPT, a late-selection account, claims that pop-out is aconsequence of learning; specifically, it occurs when a targetbecomes associated with a strong attention-attracting re-sponse. Several procedural factors limit learning in search.In particular, CAPT claims that pop-out can be learned witha CM paradigm but not with a VM paradigm. CM supportslearning (and pop-out) because it permits an attention-attracting response to become attached to the target. VM doesnot support learning (and pop-out) because it attachesattention-attracting responses to all objects; that is, attentioncannot be drawn preferentially to the target.

Note that CAPT anticipates learned pop-out in Experi-ments 1 and 2 because both used CM search. Moreover,because it is a late-selection theory, CAPT assumes that rel-ative position—the feature that defines the target—is avail-able. Thus, CAPT is consistent with the major results in bothexperiments.

CAPT is not consistent, however, with all details of thepresent experiments. CAPT predicts that, at intermediate lev-els of practice, search functions will be negatively acceler-ated. Negative acceleration occurs because CAPT claims thatsearch performance reflects a race between two statisticallyindependent processes: a controlled serial search process andan automatic attention-attracting process. The time to com-plete the serial search is a linearly increasing function ofdisplay size, whereas the time to complete automatic detec-tion is not affected by display size.

6 Texton theory does not directly deal with practice effects,although Julesz suggested that "with long practice one learns topay attention to ... weakly stimulated texton detectors" (1984, p.604). Texton theory cannot explain learning in the present exper-iments because there are no textons that encode relative position.Gurnsey and Browse (1987) demonstrated learning in texture seg-regation of line stimuli that differed only in the relative positionsof component lines. Hence, easy segregation of micropatterns de-fined by the relative position of components may be possible afterpractice. The implications of evidence from the textures segrega-tion paradigm must be treated with caution if levels of practice arenot specified.


The speed of automatic detection increases with practice.Initially, serial search is faster than automatic detection, evenat large display sizes; as a result, the search function is lin-early increasing. At intermediate levels of practice, however,automatic detection has increased in speed, so it is faster thanserial search in large displays but remains slower than serialsearch in small displays. As a result, search functions shouldbe negatively accelerated. Pop-out is the limiting case: Au-tomatic detection is faster than serial search even for smalldisplay sizes.

Inspection of Figures 2 and 3 suggests that target searchfunctions in our experiments remained linear throughoutpractice. In particular, we did not obtain negative accelera-tion at the critical intermediate levels of practice. Consistentwith the practice in the literature, we have supplied a measureof linearity, r, the proportion of total variance accounted forby a linear trend, but we note that r2 is not a reliable indexof linearity when search functions are shallow.

When the search functions are shallow in our experiments,i2 was low, a pattern taken in the literature to be a sign ofnonlinearity. However, r2 is not a good index of linearity forshallow functions. The problem is that r2 decreases with totalvariance, that is, r2 = (1 - o^/vVcrV, where crV is noise ornonlinear variance and cr2/- is total variance. As a result, whenthe slope of the search function decreases, cr^also decreases.Hence, r2 also decreases, even if a2/^ remains constant. Thedecrease in r2 is not, therefore, a reliable indicator of non-linear processing.

To investigate search function linearity in an unbiasedmanner, trend analyses were performed on target searchfunction data from Experiments 1 and 2. Of the 30 analysesperformed, only two produced significant nonlinear trends.7

Both the paucity of nonlinear trends and the large r2 valuesat intermediate levels of practice reject CAPT's prediction ofnegatively accelerated search functions.

In summary, CAPT predicts learned pop-out in Experi-ments 1 and 2, but it fails to predict the form of the searchfunction across practice. Although negatively acceleratedsearch functions are commonly reported in memory search(e.g., Jones & Anderson, 1982; Schneider, 1985), visualsearch functions are usually linear. Two well-known excep-tions (Teichner & Krebs, 1974; Treisman & Gelade, 1980)are open to confounding because of eye movements inducedby large display areas. Hence, neither a detailed analysis ofthe present experiment nor results from the visual searchliterature in general support CAPT's explanation of learnedpop-out.

Part 2: Selection

Learning Selection in VM Search

Van der Heijden (1987) criticized early-selection theory onthe grounds that it too often merges identification and se-lection. Specifically, he argued that early-selection theoristsassume that identification implies selection. We take van derHeijden's point: Preattentive representation of an attribute isnot sufficient to cause pop-out. The representation must alsobe able to control selection. Experiments 1 and 2 showed that

relative position can be identified preattentively. The nextexperiments examine how subjects learn to select preatten-tively using relative position.

With one exception, the conditions in Experiment 3 werethe same as in the 1:3 SER luminance condition of Exper-iment 1. The exception was that both types of object couldserve as a target or as a distractor on successive trials; thatis, Experiment 3 involved a VM paradigm. Subjects weregiven the target's identity before each trial. The 1:3 SERluminance stimuli were used because they produced stronglearning with CM in Experiment 1.

CAPT predicts that learning will not occur in Experiment3 because equal attention-attracting responses will be learnedto both types of objects. The prediction is supported by alarge body of data from experiments using a wide range ofstimuli (e.g., Frisk & Lloyd, 1988; Frisk & Schneider, 1984;Myers & Frisk, 1987; Schneider & Frisk, 1982a, 1982b;Schneider & Shiffrin, 1977; Shiffrin & Schneider, 1977).Indeed, the CM-VM distinction is one of the most stableregularities in the performance literature.

Method

All methods were identical to the 1:3 SER luminance conditionof Experiment 1 with the following exceptions: Subjects were toldthat either type of object (black area to the left or right of the whitearea) could be an enemy (target) and that, at most, one enemy wouldappear in any trial. The following sentence was added to the in-structions: At the beginning of each encounter, a picture of theenemy ship will appear at the center of the screen. Once you haveremembered the enemy's appearance, you may come out of hy-perspace by pushing the foot pedal.

All trials began with a picture of the target for that trial at thecenter of the screen. Subjects pressed the foot pedal to bring up thefixation cross and pressed the foot pedal again to begin a trial.

Subjects participated in three 1-hr sessions on consecutive days.The first day began with two practice blocks. Each block terminatedafter 20 consecutively answered trials with no RT limit. Practiceused CM: In the first block, the target had the black area to the left,and in the second block, the target had the black area to the right.Subjects then performed 8 experimental blocks. In the followingtwo sessions, they performed 12 experimental blocks per sessionwith no practice. Both types of objects were used equally often astargets in each block. Eight students from Queen's University werepaid $15 for their participation. All subjects had normal orcorrected-to-normal vision.

Results

A practice factor with eight levels was created by dividingtrials into consecutive groups of four blocks. There was nodifference between the two target types in either MRT forcorrect responses (F < 1) or for percentage of error (F < 1),and target type did not interact with other factors in either

7 The cases were the quadratic trend for the third practice levelof the 1:3 SER luminance condition in Experiment 1, F(\, 5) =7.2, p < .05, and the quadratic and cubic trends for the 1:3 SERcolor condition in Experiment 1, F(l, 5) = 8.1,p < .05, and F(l,5) = 7.9, p < .05, respectively.


measure. Hence, we collapsed the data over target type. Fig-ure 4 shows both MRT for correct trials and percentage oferror as a function of display size and of display type.

Mean correct RT. ANOVA was carried out on MRT forcorrect trials to assess the effect of display size, display type,and practice. As in both Experiments 1 and 2, MRT increasedwith increasing display size, F(3, 21) = 22.7, p < .001, anddecreased with practice, F(7,49) = 28.3,/?< .001.The slopeof the search functions decreased with practice, F(21, 147)= 9.8, p < .001.

In contrast to Experiments 1 and 2, nontarget responseswere not significantly slower than target responses, F(3, 7)= 3.1, p < .1. In addition, nontarget search functions weresteeper than target search functions but only early in practice,a pattern revealed by the interaction of display type withdisplay size, F(3, 21) = 5.6, p < .01, and of display type withpractice, F(7, 49) = 3.7, p < .005.

Linear regressions. Table 4 presents the slopes of searchfunctions calculated by linear regression of MRT on displaysize. For each subject, regressions were calculated for targetand nontarget displays at each level of practice. Table 4 alsopresents the slope from regressions averaged over subjects.

The latter regressions apply to the RT data presented inFigure 4.

Slopes derived from regression on data averaged over sub-jects were below the pop-out criterion in the last two practicelevels. Regressions on individual data confirmed pop-outwith practice. All subjects except one (Subject 4) passed thepop-out criterion in at least one practice level. Several sub-jects (Subjects 3,7, and 8) displayed pop-out in most practicelevels after the third.

Error frequency and SAT. Subjects were remarkably ac-curate (4.2% errors overall). The number of errors increasedwith display size, F(3, 21) = 6.1, p < .005, and was largerfor target than for nontarget displays, F(3, 7) = 23.2, p <.005. The difference between present and absent errors in-creased with display size, F(3, 21) = 20.7, p < .001. Errorsdid not, however, decrease significantly with practice, F(l,49) = 1.5, p > .15.

Practice did not increase errors, as predicted by SAT;hence, we reject an account based on SAT. We note, however,that because practice did not reduce errors, the evidenceagainst an SAT account is not as strong as in Experiments 1and 2.

H

05O)

C

a

1500 •

1000

500

PCe«o>

1500

1000

u 500Ou 15

£ 10

= 0.97

.-•O..... -o —

= 0.96 0.89

2 4 6 8 2 4 6 2 4 6 8 2 4 6

Display Size

Figure 4. Mean correct response time (RT; upper panels) and percentage of errors (lower panels)from Experiment 3 for each level of practice (four blocks per level, practice increasing from left toright panels, the top row containing the first four levels and the bottom row the last four levels ofpractice). (Dashed lines represent data from target-absent displays, and solid lines represent datafrom the target-present displays, r; and ru are the squared correlation coefficients from linearregressions on target and nontarget RT data, respectively.)


Table 4Slopes (Milliseconds per Object) From the Linear Regression of MRT on Display Size in Experiment 3

Practice level

Subject

12345678

Combined

1

445744948633762257

2

108030724212702342

3

15451463142229-425

Target

4

1438264147

-13-215

display

5

273413273

11-1214

6

-23110361514-3213

7

-1833824

-147

8

10163331011

-25119

1

4560571299430803266

2

5561391277243861

60

Nontarget display

3

36482649243442-732

4

41552613272731-527

5

46582245341834032

6

22572233181636-225

7

24469616518-315

8

3335115201619

-1216

Note. Regressions were performed on data from individual subjects and on data obtained by combining subjects at each level of practice.There were 4 blocks of 80 trials per level of practice.

We computed linear regressions of percentage of error ondisplay size (averaged over subjects); the regressions cor-respond to the lower panels on Figure 4, and, as shown inTable 5, the effect of display size on error was constant overpractice. Hence, confounding of the practice effect (and theresulting pop-out) by SAT is unlikely. Note that nontargeterrors tended to decrease with display size. Figure 4 showsthat the decrease was due mainly to high error rates at thesmallest display size.

Probabilistic distance cues. Analyses of probabilisticdistance cues, as described for Experiment 1, were conductedto determine whether or not they confounded the presentresults. As in Experiments 1 and 2, we found no evidence thatsubjects used the distance cues. MRT from less probable trialswas less (although not significantly so) than MRT from themore probable conditions. Hence, probabilistic distance cuescannot account for pop-out in the present experiment.

Comparison with Experiment 1. The results show astrong practice effect in the present experiment. To determinewhether the practice effect is as strong as in a CM paradigm,we compared the first five levels of practice in the currentdata with their counterpart from Experiment 1, namely thedata from the 1:3 SER condition of Experiment 1.

Although MRT in Experiment 1 (668 ms) was faster thanin the present experiment (824 ms), the difference was notsignificant (F < 1). The CM-VM (experiment) factor con-tributed to only one interaction: between display size anddisplay type, F(4,48) = 4.3, p < .01. Nontarget search func-tions were shallower in the present experiment than in Ex-periment 1 (compare Figures 2 and 4).

Discussion

The present experiment showed strong pop-out in VMsearch after practice. Learned pop-out was not confoundedby probabilistic distance (feature-density) cues and was notdue to SAT. VM search performance was slower than in thecorresponding CM condition (from Experiment 1), but thedifference was not significant.

CAPT was consistent with the major finding of the firsttwo experiments: learned pop-out based on relative position.CAPT predicted, however, that learning would not occur inthe VM paradigm. Hence, the present results appear to offerstrong evidence against CAPT.

An advantage of CM over VM is one of the strongestregularities in the performance literature. We know of noother demonstration of learning in a VM search paradigm.Nevertheless, the present experiment demonstrated the stron-gest type of learning in search, namely learned pop-out.

CAPT's predictions are based on an object-level analysis:Learned pop-out occurs when a target object becomes as-sociated with a strong attention-attracting response. If thetarget and distractors trade roles across successive trials,learning (and hence pop-out) should not occur.

Is an object-level analysis appropriate? The question ismotivated by a simple idea we call the group scale hypoth-esis: At the level of the display as a whole, one cue wasconsistently mapped, namely the grouping structure of thedisplay. Target displays could be partitioned into two groups:a group of one (the target) and a group of many (the dis-tractors). Nontarget displays contained only one group. If

Table 5Slopes (in Percentage of Error per Object) From Linear Regressions of Percentage of Error on Display Size inExperiment 3

Practice level

Target

PresentAbsent

1

0.9-0.5

2

0.9-0.5

3

0.7-0.2

4

0.5-0.5

5

0.8-0.2

6

0.60.1

7

0.8-O.I

8

0.5-0.1


subjects responded on the basis of the group structure of thedisplays, the present data are consistent with the well-established CM versus VM regularity.

Note that the displays used by Schneider and Shiffrin(1977) did not afford group scale cues because distractorswere heterogeneous. Note also that search on the basis ofgroup scale reinforces our conclusion that relative positionis coded preattentively: The appropriate groups could beformed only if information about relative position is avail-able preattentively. Hence, the group scale idea is consistentwith the data from all three experiments.

Search Using Group Scale Cues

Grouping is widely acknowledged to be a fundamentaloperation of preattentive processing. All current theories ofvisual search assume preattentive grouping, at least at thelevel of objects. Moreover, all claim that subjects search inthe object domain; that is, subjects do not search blank screenareas between objects.

Our idea is that subjects can select groups of objects andprocess them as a single unit. Moreover, we claim thatpreattentive grouping is an adaptive process that can beperformed on the basis of any discriminable attribute, in-cluding relative position. In addition, we claim that selec-tion by attention occurs at the level of groups and that thesize of a group (group scale) itself is but one of many groupattributes that can control selection. The basic idea, then, isthat experience in a particular search task leads subjects tochoose dimensions with a similarity structure that producesuseful groups (i.e., groups that facilitate performance in thetask). Subjects select the dimensions that control groupingand then group objects that have similar values on thechosen dimensions.

In the natural world, luminance, color, motion, and posi-tion are useful dimensions for grouping because points withsimilar values often belong to the same object. Hence, whentargets are defined by discriminable differences on such di-mensions, subjects show pop-out with little or no practice.Because these dimensions are useful in day-to-day experi-ence, subjects likely try to use them first when performinga search task, even if they do not yield useful groups. Relativeposition is a less useful cue in the natural world, and, withoutpractice, search for objects defined by relative position isdifficult: Subjects have to discover that grouping by relativeposition solves the problem.

The present experiments indicate, then, that grouping cri-teria are not fixed, that subjects are able to use relative po-sition as a grouping criterion, and that subjects can selectgroups according to their scale. With our stimuli, relativeposition is a potent basis for grouping, and when the displaysare grouped using relative position, small groups tend to con-tain targets. Hence, when subjects select small groups, per-formance is facilitated. When preattentive grouping reliablyproduces only one small group (i.e., the target), search usingthe group scale cue yields pop-out.

Previous results support our emphasis on search usinggroup structure rather than using individual objects. We notetwo examples. First, Egeth, Virzi, and Garbart (1984) found

that the number of distractors sharing a target's color in con-junction search predicted search performance better than thetotal display size (see also Treisman, 1982). Second, randomvariation in dimensions irrelevant to target distractor dis-crimination does not usually affect performance (e.g., Treis-man, 1988). When an irrelevant variation is unique to onedistractor, however, Pashler (1988) showed that target dis-crimination is degraded. The interference can be explainedin terms of group scale: If the target and the unique distractorboth formed groups of the same spatial scale, the interferencereflects the difficulty of discriminating the target and theunique distractor on the group scale dimension.

Search using group scale cues also explains a puzzlingfeature of Experiment 3, namely the similarity of target andnontarget search functions. Usually, nontarget responses areslower than target responses, and the difference increaseswith display size (a pattern often attributed to self-terminating search). When subjects search on the basis ofgrouping structure, nontarget displays can be detected by theoccurrence of a single large-scale group (or by the absenceof a small-scale group). Therefore, detecting a nontarget dis-play requires only one discrimination, and the number ofdiscriminations does not increase with display size.

Humphreys et al. (1985) proposed a very similar idea:When they used distractors that were homogeneous andspaced at regular intervals, they found that nontarget re-sponses were faster than target responses and suggested thathomogeneity and regularity facilitated grouping of distrac-tors. Grouping allowed subjects to detect nontarget displaysby the occurrence of a large-scale group. The present ex-periment showed a pattern intermediate between the usualresult (slower nontarget search) and Humphreys et al.'s re-versal of that result, perhaps because our distractors werehomogeneous but not spaced regularly.

Experiment 4

In Experiment 4, we test a counterintuitive prediction ofthe group scale hypothesis, namely that subjects can learnpop-out even when they do not have prior knowledge of thetarget's identity. Experiment 4 was identical to Experiment3 except that subjects were not cued with the target's identitybefore each trial. Instead, they were told that target displayscontain one object that differs from the others: the odd man.

Under the group scale hypothesis, pop-out can be learnedbecause subjects do not need to know the feature value ofthe target; they need only group on the relevant dimension.Use of group scale cues also predicts that nontarget perfor-mance will be equivalent to target performance. Becauseodd-man search makes serial search less attractive than inExperiment 3, nontarget performance may even be fasterthan target performance.

Tsotsos (1990) claimed that odd-man search is difficultbecause of its computational complexity. He assumed thatsubjects search for mismatches between objects (see alsoSagi & Julesz, 1987). Mismatches are detected by comparingindividual objects with each other. On that assumption, odd-man search is difficult because of the number of comparisons


required. Target detection for the largest display size (8 ob-jects) in Experiment 4, for instance, requires up to 22 com-parisons among objects (7 of the possible 28 comparisonswill reveal the odd man if it is present). When the target isspecified, by contrast, a maximum of 8 comparisons to thetarget template are required. If subjects search using groupscale cues, however, the odd-man task should be no moredifficult than search in which the target is specified and maybe easier because subjects will have less incentive to use aninefficient serial strategy.

Method

All methods were identical to Experiment 3 with one exception.The subjects were told to look for a target defined as the odd man.Specifically, the instructions used in Experiments 1 and 2 wereexpanded: Your display will show the enemy (if it appears) as the"odd man out," that is, of the opposite type to the other ships.

As in Experiments 1 and 2, trials began with a fixation cross.Subjects pressed a foot pedal to begin a trial.

Results

A practice factor with eight levels was created by dividingtrials into consecutive groups of four blocks. There was no

difference between the two target types in either MRT (F <1) or percentage of error, F(l, 7) = 1.8, p > .2. Target typedid not interact with any other factors in either measure.Hence, we collapsed the data over the target type. Figure 5summarizes both A/RT for correct trials and percentage oferror as a function of display size and display type.

Mean correct RT. ANOVA was carried out on MRT fromcorrect trials to assess the effects of display size, display type,and practice. As in the previous experiments, A/RT increasedwith increasing display size, F(3, 21) = 4.3, p < .05, anddecreased with practice, F(7, 49) = 17.7, p < .001. In par-ticular, the slope of the search function decreased with prac-tice, yielding" a strong interaction of practice with displaysize, F(21, 147) = 2.6, p < .001.

As in Experiment 3, there was a null effect of display type,F( 1, 7) = 2.2, p > . 1. In contrast to Experiment 3, however,target displays were slower than nontarget displays (758 msand 737 ms, respectively), but the difference was not sig-nificant. Search functions were initially steeper for targetsthan for nontargets. Later in practice, they became equal. Theshift produced an interaction of display type with display sizeand practice, F(21, 147) = 1.6, p < .05.

Linear regressions. Table 6 presents the slopes of thesearch functions calculated by linear regressions of MRT on

««

£

1500

1000

50015105

ra = 0.97

= 0.83

2 4 6

= 0.82

•••o ^

ra = 0.97 = 0.93- 2

2 4 6 2 4 6 8 2 4 6

Vi

ase

Iw

1500

1000

50015105

= 0.75

= 0.42

ra = 0.81 = 0.93

0.33

2 4 6 2 4 6 2 4 6 2 4 6

Figure 5. Mean correct response time (RT; upper panels) and percentage of errors (lower panels)from Experiment 4 for each level of practice (four blocks per level, practice increasing from left toright panels, the top row containing the first four levels and the bottom row the last four levels ofpractice). (Dashed lines represent data from the target-absent displays, and solid lines represent datafrom the target-present displays, r^ and rf are the squared correlation coefficients from linearregressions on target and nontarget RT data, respectively.)


Table 6Slopes (Milliseconds per Object) From the Linear Regression of MRT on Display Size in Experiment 4

Practice level

Target display Nontarget display

Subject

12345678

Combined

1

20225812116870

133

2

028352151144417

3

-31 13331

-4481213

4

-3122753

-748-510

5

262610-43

21-47

6

31

23-3

1433-18

7

13-101468

-12337

8

-1-92147322

-114

1

254581-24-106459230

2

141356

-18-0-548-613

3

-84862

-17-10-11197

11

4

134852

1-16-651-118

5

84955-7-7-2420512

6

131967-3-14-2541-012

7

17964-5-8-28382

11

8

91362-65

-1926-511

Note. Regressions were performed on data from individual subjects and on data obtained by combining subjects at each level of practice.There were 4 blocks of 80 trials per level of practice.

display size. The regressions were calculated for each subjectfor target and for nontarget displays at each level of practice.Table 6 also presents the slopes derived from data combinedover subjects.

Regressions using the data averaged over subjects showedslopes below the pop-out criterion in the last four practicelevels. Regressions on individual subject data confirm pop-out with practice. All subjects except two (Subjects 3 and 7)passed the pop-out criterion in at least one practice level.Several subjects (Subjects 1,4,5,6, and 8) displayed pop-outin most practice levels after the third.

Error frequency and SAT. The percentage of errors in-creased with display size, F(3, 21) = 10.8, p < .001, wasgreater for target than nontarget displays, F(3, 7) = 7.2, p <.05, and decreased with practice, F(l, 49) = 4.1, p < .005.Early in practice, target errors increased with display size,whereas nontarget errors decreased with display size; later inpractice, the effect of display size decreased for both typesof error (see Figure 5). The pattern yielded an interaction ofdisplay type with display size, F(3, 21) = 23.6, p < .001,with practice, F(7, 49) = 2.6, p < .05, and with both displaysize and practice together, F(21, 147) = 2.1, p < .01.

As in both Experiments 1 and 2, an explanation for thepractice effect based on SAT can be discounted because per-centage of error decreased with practice. Nevertheless, weperformed linear regressions of percentage of error on dis-play size (as in Experiment 3). Each regression correspondsto an error search function, and, as shown in Table 7, targeterror slopes were high over the first five practice levels butdecreased slightly thereafter. As in Experiment 3, nontargeterrors decreased with display size.

Probabilistic distance cues. As before, we checkedwhether or not subjects used probabilistic distance cues. Asin the previous experiments, there was no evidence that sub-jects used the density cues: MRT from the less probable trialswere less than MRT in the more probable conditions (althoughonly significantly so in one of the four comparisons). Weconclude that subjects' use of probabilistic distance cues can-not account for pop-out in the present experiment.

Comparison with Experiment 1. As in Experiment 3,there was a strong practice effect in the present experiment.To determine whether the practice effect is as strong withodd-man search as it was with CM, we compared data fromcorrect trials at the first five levels of practice in the presentexperiment with the corresponding data in the 1:3 SER lu-minance condition of Experiment 1 (see Figures 2 and 5).

MRT was faster in Experiment 1 than in Experiment 4 (668ms and 748 ms, respectively), but the difference was notsignificant (F < 1). The CM-OM (experiment) factor in-teracted with all other factors, F(12, 144) = 3.6, p < .001.A four-factor interaction is difficult to describe in detail. Inthe present case, it reflects steeper search functions for Ex-periment 1 than for Experiment 4 early in practice and op-posite advantages for target and nontarget displays, also earlyin practice.

Comparison with Experiment 3. We suggested thatsearch may be faster in the present experiment than in Ex-periment 3 because the availability of the target cue in Ex-periment 3 may have tempted subjects to use an inefficientserial search strategy. To check, we compared correct trialsin the present experiment with the corresponding data inExperiment 3 (see Figures 4 and 5). The comparison exam-

Table 7Slopes (in Percentage of Error per Object) From Linear Regression of Percentage of Error on Display Size inExperiment 4

Target

PresentAbsent

1

1.6-0.8

2

1.6-1.0

3

2.4-0.7

Practice

4

1.7-0.1

level

5

1.60.)

6

0.90.0

7

1.4-0.2

8

1.20.3


ines two kinds of VM search, namely search with and withouta predefined target.

A/RT in Experiment 3 was slower than in Experiment 4(means of 824 ms and 748 ms, respectively), but the differ-ence was not significant (F < 1). There was a complex pat-tern of interactions: Early in practice, subjects were morestrongly affected by display size in Experiment 3 than Ex-periment 4. With practice, both conditions converged to anequal and small effect of display size. Second, early in prac-tice, subjects in Experiment 3 made slower nontarget thantarget responses, whereas subjects in Experiment 4 madeslower target than nontarget responses. With practice, thedifference between target and nontarget responses tended tozero. The pattern of results produced an interaction of ex-periment with practice, F(7,98) = 3.5, p < .005, with displaytype, F( 1, 14) = 5.1, p < .05, with display size and practice,F(21, 294) = 3.0, p < .001, and with display type and prac-tice, F(7, 98) = 4.8, p < .001.

Discussion

As in the earlier experiments, practice produced strongpop-out. Learned pop-out was not confounded by probabi-listic feature-density cues and was not due to SAT. Moreover,odd-man search was faster, although not significantly so,than the simple VM search used in Experiment 3. The resultsare consistent with the group scale hypothesis. They are notconsistent with theories, such as with Tsotsos's (1990) ac-count, that rely on an object-level analysis.

Comparison among the corresponding 1:3 SER luminanceconditions in Experiments 1, 3, and 4 showed consistent dif-ferences early in practice but similar performance after prac-tice. The differences early in practice likely reflect the sub-ject's use of a serial search strategy. Serial self-terminatingsearch predicts steep target and nontarget search functions,with slower performance for nontarget displays than for tar-get displays. That pattern was most evident in the CM searchand least evident in odd-man search, in which serial searchhad little utility.

In all paradigms, subjects abandoned the serial strategywith practice. Late in practice, all conditions showed a pat-tern consistent with pure group scale search; that is, thesearch functions were shallow, and target and nontarget per-formance was effectively the same.

In contrast to Experiments 1 and 2, both Experiments 3 and4 produced negatively accelerated search functions for targetdisplays. Because larger displays allow the formation oflarger distractor groups, negative acceleration is consistentwith the group scale hypothesis. Larger distractor groupsprovide a more discriminable group scale cue; hence, thelarge groups are less likely to be selected, speeding perfor-mance. Paradoxically, although negative acceleration is pre-dicted by CAPT in some conditions, it occurred only in con-ditions in which CAPT predicts no learning. The group scalehypothesis, then, is the only account consistent with the over-all pattern of the data.

Group scale and error data. We have been carefulthroughout to provide analyses of errors as well as of MRT

for correct trials. The analyses were motivated by the po-

tential of a confound with SAT. In fact, the error data tell adifferent story.

When MRT increases with display size, target errors in-crease, whereas nontarget errors are both less frequent thantarget errors and are not affected by display size. The patternis incompatible with a serial search model in which theterminate-or-continue decision is error prone, even thoughserial search may be consistent with MRT for correct trials.

Eriksen and Spencer (1969) were the first to note that anerror-prone serial process predicts that nontarget errorsshould increase with display size, whereas target errorsshould remain constant. Target errors are independent of dis-play size because only one target occurs in a display; hence,regardless of the number of distractors, there is only onechance of missing the target. By contrast, as display sizeincreases for nontarget displays, there are an increasing num-ber of chances to mistake a distractor for a target. Note thatthe pattern predicted by serial search is opposite to the patternusually found in studies of visual search, including our ex-periments. Thus, error data reject an error-prone serial searchprocess.

The predictions made by Cave and Wolfe's (1990) guidedsearch model are also inconsistent with the error data. In theiraccount, only objects that produce an activation greater thana threshold are selected. Target errors occur when the target'sactivation does not exceed the threshold. Because target ac-tivation is an increasing function of the sum of the differencesbetween its features and the features of all distractors (i.e.,its uniqueness value), the chance that a target's activationwill exceed the threshold should increase with display size.Hence, guided search incorrectly predicts that target errorswill decrease with display size. Guided search has no mech-anism for predicting nontarget errors. If an error-prone serialprocess is assumed, guided search erroneously predicts thatnontarget errors should increase with display size. Thus,guided search provides a less successful account than a sim-ple serial search model.

According to the group scale hypothesis, by contrast, tar-get errors occur when a target is grouped with distractors. Thechance of a target error increases with display size, bothbecause there are more distractors with which the target cangroup and because the groups that the target joins can belarger (and, hence, less likely to be selected). The group scalehypothesis is consistent with the target data from all fourexperiments (and with data of the visual search literature, ingeneral). The group scale hypothesis is also consistent withsubjects' reports that target errors occur when the target be-comes "lost among the distractors."

In the nontarget case, the group scale hypothesis predictsslightly higher error rates at small display sizes. A nontargetdisplay of two objects, for instance, forms a group not muchlarger than a single target, so subjects are likely to report thepresence of a target erroneously. As predicted, subjects mademost nontarget errors for display size two in both Experi-ments 3 and 4. The same pattern was not evident in Exper-iments 1 and 2, perhaps, as we noted earlier, because subjectsused a mixture of group scale and serial search. Becauseserial search predicts increasing nontarget errors with display


size, the effect of group cue search on nontarget error rateswould be masked.

Theoretical Implications of Experiments 3 and 4

In the following discussion, we review the implications oflearned pop-out in the VM paradigms for the main theoriesof visual search, namely FIT, guided search, and attentionalengagement theory (AET; Duncan & Humphreys, 1989). Al-though none of the theories address learning, all can predictthe learning observed. Only AET, however, is consistent witha detailed analysis of Experiments 3 and 4, although somerefinements in grouping mechanisms are suggested by thedata. We offer a modification of AET that accounts for thepresent data by incorporating the group scale hypothesis.

FIT and modified FIT claim that pop-out occurs when at-tention is spread across the entire search display (Treisman,1988; Treisman & Gormican, 1988; Treisman & Sato, 1990;Treisman & Souther, 1985). A target defined by a simplefeature can be detected by determining the total activation ina map tuned to the target feature.

FIT predicts learned pop-out if subjects can tune separatefeature maps to prefer each type of target. In Experiment 3,subjects could attend to pooled activation of the map indi-cated by the target cue. In Experiment 4, however, subjectsmust attend to the pooled activation of both maps in turnbecause they do not know the target's identity in advance.Hence, FIT predicts that performance in Experiment 4 shouldhave been slower than performance in Experiment 3. FIT isinconsistent with the data on two grounds: First, it assumesrelative position is derived from attention; second, perfor-mance was not slower in odd-man search.

Guided search assumes that a target gains activation, andhence a greater chance of being selected, from both its matchto a target template and its uniqueness in the display. Guidedsearch could deal with learning in visual search by allowingthe system to tune the target template. In Experiment 3, sub-jects could apply the template indicated by the target cue. Inodd-man search, however, each template must be applied inturn because subjects have no prior knowledge of the target.Hence, guided search using target-template matchingwrongly predicts slower performance in odd-man search.

Guided search could be modified to predict equivalent per-formance in Experiments 3 and 4 by assuming that subjectssuppress template matching and detect targets using the ac-tivation resulting from uniqueness. The uniqueness mecha-nism does not rely on prior knowledge of the target. Indeed,in reply to Tsotsos's (1990) claim that odd-man search iscomputationally difficult, Heathcote and Mewhort (1990)pointed out that simple algorithms that do not rely on priorknowledge of the target can make odd-man search tractable.As an example of such an algorithm, we cited the uniquenessmechanism of guided search. Unfortunately, the uniquenessmechanism incorrectly predicts that target errors will de-crease with display size. A further problem is that the unique-ness mechanism is not consistent with the redundant targeteffect in search.

The redundant target paradigm compares performance onsingle- and multiple-target displays. Faster performance for

multiple targets can be caused by statistical facilitation(Raab, 1962) and by coactivation (Miller, 1982). Statisticalfacilitation occurs when search is random and, therefore, thechance of randomly sampling a target increases as the num-ber of targets increases. Coactivation occurs when multipletargets increase the speed of selection of any one target.

Coactivation and statistical facilitation can be differenti-ated experimentally because coactivation predicts that thefastest RTs from multiple-target displays are less than thefastest RTs from single-target displays (Miller, 1982). Sta-tistical facilitation predicts only that multiple-target displayswill produce more fast RTs than single-target displays.Egeth and Mordkoff (1991) and Mordkoff, Yantis, andEgeth (1990) supported a form of coactivation, an interac-tive channels model, using the fast RT test (see also Fournier& Eriksen, 1990).

Neither simple serial search nor guided search are able toexplain coactivation. For both, the fastest RTs for a singletarget cannot be faster than for multiple targets, because thefastest RTs for both models occur when the target is selectedfirst. In fact, guided search provides a less successful accountthan simple serial search because multiple targets reduce tar-get uniqueness and, hence, reduce the chance of selectingeach target. As was the case for the error data, the uniquenessmechanism of guided search predicts the wrong pattern ofdata. It is, therefore, not a viable explanation for the nulldifference between Experiments 3 and 4.

The theories of visual search reviewed so far predict slowerperformance in odd-man search because of mechanisms thatrequire prior knowledge of the target's identity. Search us-ing group scale cues does not require prior knowledge of thetarget's identity, only knowledge of the dimension that fa-cilitates separate grouping of targets and distractors. Sup-port for efficient selection without prior knowledge of thetarget comes from a recent texture-segregation study byCaelli (1991). He found that priming subjects with the shapeor the texture of a disparate quadrant did not improve seg-regation performance; he concluded that prior knowledgedoes not influence texture segregation, a prototypical group-ing process.

Like the group scale hypothesis, AET (Duncan & Hum-phreys, 1989) emphasizes preattentive grouping in search.However, like CAPT, AET explains pop-out using anattention-attracting mechanism (attention-attracting strengthis called activation in AET). Selection acts on a perceptualdescription that is grouped hierarchically at multiple spatialscales. A group can be an object or a collection of objects.Grouping of objects occurs with a strength proportional totheir similarity and spatial contiguity. The unit of selectionby attention is a group.

When a display is first presented, AET assumes that allobjects have the same activation. Parallel target-templatecomparisons increase the activation of good matches.Changes in activation for objects in the same group are cor-related, and the size of the correlation depends on the strengthof grouping. Processing capacity is limited because the sumof activations is conserved (i.e., an increase in one group'sactivation leads to a decrease in other groups' activations).


A group's activation determines both the probability that itwill be selected and the speed with which it is selected.

AET can explain learned pop-out using two mechanisms.First, learning may improve target-template matching and,hence, gives targets greater activation. Second, learning maylead to stronger grouping and thereby decrease the chancethat distractors are selected.

In the original version of AET (Duncan & Humphreys,1989), the relation between the grouping and the (target)template-matching mechanism was not specified fully. Oneinterpretation is that grouping is driven by target-templatecomparison. If grouping is driven by target-template match-ing, however, it is unclear how subjects can perform an odd-man search because a target template is not provided.

If subjects can apply each template in turn, both potentiallearning mechanisms predict pop-out in odd-man search.Both, however, also incorrectly predict slower odd-mansearch than search in a simple VM paradigm. Even if learningaffects only the grouping mechanism, odd-man searchshould be slower because grouping is driven by target-template matching. An inappropriate template should lead toa highly activated distractor group that can be mistaken forthe single large distractor group characteristic of nontargetdisplays. Hence, even if subjects were to rely on strongergrouping, AET erroneously predicts slower performance inExperiment 4 than in Experiment 3.

A more recent version of AET (Duncan & Humphreys,1992) places a stronger emphasis on the independence ofgrouping and template-matching processes. They stated that:

When nontargets are completely homogeneous, subjects relylargely and perhaps entirely on a strategy of local mismatchdetection (i.e., deciding simply whether any element in thearray is unlike its neighbors), rather than true search using atarget template ... the slope of the search function can be veryflat even when subjects do not know in advance the identity ofany target, (p. 580)

The recent version of AET can explain pop-out in the odd-man condition, but it is somewhat at odds with the data, giventhat it proposes a local mismatch-detection mechanism. Asnoted in relation to Tsotsos's (1990) account of search, localmismatch detection predicts slower performance in the odd-man condition than in a conventional VM condition becausedetecting a mismatch requires more comparisons than tem-plate matching. Other versions of mismatch detection are,however, possible, so our results do not reject all possiblemismatch models.

Alternately, greater emphasis on grouping mechanisms,especially if group structure can form without target-template matching, allows AET to explain the present data.In the following section, we develop a modified version ofAET that explains our data by incorporating the group scalehypothesis.

Modified AET

Of the theories reviewed, AET is closest to the group scalehypothesis because it assumes late selection and emphasizesgrouping as a determinant of search performance. Our mod-ification of AET assumes that grouping is a preattentive pro-

cess that occurs according to the proximity and similarity offeatures. We also assume that grouping is adaptive: Theweights given to proximity and to each dimension of sim-ilarity can be modified by practice to produce useful groups.Although we suggest that grouping is preattentive, we retainthe idea of an attention-attracting mechanism that is con-trolled, at least in part, by target-template matching. In ad-dition, we retain both the assumption that changes in acti-vation within a group are correlated and the assumption thatthe speed with which a group is selected is proportional tothe group's activation.

Our six modifications to AET are designed to combine theexplanatory power of the group scale hypothesis with theexisting theory.

First, search using groups based on proximity and simi-larity explains a wide range of existing data including (a)search restricted to a subset of objects formed on the basisof a target feature (Egeth et al., 1984), (b) the effect of targetand distractor proximity (Bundesen & Pedersen, 1983;Carter, 1982; Farmer & Taylor, 1980; Treisman, 1982), (c)the effect of target and distractor similarity and of interdis-tractor similarity when distractors are not homogeneous(Duncan, 1990; Farmer & Taylor, 1980), (d) the effect ofso-called illusory conjunction displays (see Part 1), (e) theincrease in target errors with display size, and (f) the oc-currence of fast, nontarget responses (Humphreys et al.,1985; present Experiments 3 and 4).

Second, there are clear cases in which attention-attractingresponses are learned and in which prior knowledge of thetarget influences performance. Although we argued earlieragainst an attention-attracting mechanism, our dispute con-cerned the linkage between that mechanism and grouping.By retaining the mechanism, we acknowledge its role in suchcases.

Shiffrin and Schneider (1977), for example, found stronginterference in a CM search task performed after subjectshad extensive training with the mapping of the target anddistractor reversed. The interference indicates that subjectshad learned to select objects that previously were targets.Pashler (1987) showed that distractors similar to a targetincreased MRT in search, even for displays that did notcontain that particular target. The increase indicates thatsubjects' prior knowledge of possible targets influencedtheir search performance.

Note that the utility of attention-attraction through target-template matching depends on the demands of the task. OurExperiments 3 and 4 discouraged use of the attention-attraction mechanism because it could only degrade perfor-mance. Experiments 1 and 2 (and virtually all CM visualsearch tasks), by contrast, encouraged the use of the mech-anism. By including an attention-attracting mechanism con-trolled by target-template matching, we can deal with bothCM and VM performance.

Third, when multiple targets are presented, they can begrouped and, therefore, reinforce each other's activation. InAET, the speed with which a group is selected is proportionalto its activation (not just to the probability with which it isselected, as in guided search). When targets are grouped to-gether, therefore, the fastest RTs for multiple-target displays


will be less than the fastest RTs for single-target displays.Because the speed with which a group is selected is corre-lated with its activation, the model can address coactivationby exploiting the correlation in changes of activation amongmembers of a group.

Fourth, like AET, we assume that perceptual processing ishierarchical, with both feed forward and feedback interac-tions among representations ranging from simple features tomeanings and categories. Because we assume that the de-velopment of the perceptual representation and group acti-vation is a dynamic process, we acknowledge that some rep-resentations may take longer to activate than others, eitherbecause they are higher in the hierarchy (e.g., meanings) orbecause they are used rarely (e.g., relative position).

Representations that develop slowly should take longer toinfluence both grouping and group activation; therefore, suchrepresentations should not affect performance in tasks thatrequire rapid responding, such as visual search. Our exper-iments indicate that relative position has such a representa-tion but that practice can speed its development. In short, ourmodification of AET suggests that the use of a representationin perceptual processing depends on both discriminabilityand learned utility.

Fifth, in AET, each group is assigned an attention-attracting strength or activation. A group's activation is de-termined by its match to a target template. The target tem-plate can be a complex specification of attributes such ascolor, luminance, motion, and shape and meaning. We sug-gest that a target template can also specify the spatial prop-erties of the grouping structure itself (e.g., the group's scale).The inclusion of properties of the grouping structure in thetarget template extends AET to explain learning in the nom-inally VM conditions of Experiments 3 and 4. The expla-nation is achieved without having to reject a CM learningadvantage.

Sixth, like AET, we claim that attention selects groups; theprobability and speed of selection depend on the group'sactivation. Note that selection occurs on a unified perceptualrepresentation; it is not used to perform perceptual (identity)processing. Selection from a unified representation is con-sistent with the null difference between performance in thecolor and luminance conditions of Experiments 1 and 2. Oursis a late-selection position; we assume that objects are se-lected to control actions (see Duncan, 1980). A group's ac-tivation, therefore, indexes the strength with which it com-petes to control actions ranging from a saccade to productionof a manual response.

We also claim that selected groups are treated as a unit. Ifa group containing one target and many distractors is se-lected, the target's properties will be masked by the distrac-tors' properties in the average. Subjects should, therefore,tend to miss targets when they have been grouped errone-ously and selected with distractors. Target errors should bemore frequent than nontarget errors because they can becaused both by misidentification (from noise in the activationof representations) and by averaging. Nontarget errors, bycontrast, are caused only by misidentification. Note that er-rors could be decreased by searching at a finer group scalebut only at the cost of slower performance.

Modified AET Applied to the Present Experiments

In our account, subjects can use both group structure andgroup activation depending on the task they must perform.In the CM paradigm (Experiments 1 and 2), we assume thatsubjects developed an attention-attraction response to the tar-get and grouped objects using relative position. Initially, use-ful groups were slow to develop, and target errors occurredfrequently, especially with large display sizes that encour-aged targets and distractors to group together. Group acti-vation provided a poor target cue, and search was almostrandom within the smaller groups. With practice, however,both group scale and group activation became better targetcues; as a result, both A/RT and target errors decreased, andthe search function and it equivalent in errors flattened.

In the VM experiments, by contrast, group activation re-sulting from relative position information remained a poortarget cue throughout practice, especially in the odd-mansearch. Hence, there were more errors, and MRr was slowerthan in Experiment 1. Group size was the only reliable targetcue; as a result, small groups tended to have larger activation,and performance reflected the grouping mechanism morethan the attention-attracting mechanism.

Note that our account explains differences in performancebetween the VM and CM paradigms by changing the relativecontribution of the grouping and the attention-attractingmechanisms. In the VM paradigms, learning mainly reflectsthe development of grouping, whereas in the CM paradigm,learning reflects both mechanisms. Because the speed ofgrouping is not strongly affected by both display size andtype, the effects of display size and display type in the VMexperiments were weaker than in the CM experiments. Be-cause the identity of the target is less certain in a VM par-adigm, there were more errors in the VM experiments thanin the CM experiments. Because target uncertainty remainedstable across practice, the VM paradigm remained more errorprone across practice than the CM paradigm.

In our account, grouping and group activation are inter-dependent. Because grouping and group activation are in-terdependent, the theory is not constrained to predict nega-tive acceleration in search functions. It can tolerate negativeacceleration in the VM experiments, in which performancemainly reflected the development of grouping, and linearsearch functions in the CM paradigms, in which performancereflected both mechanisms.

General Discussion

The arguments can be summarized as follows: Previousputative demonstrations of preattentive representation of rel-ative position (and, we argue, simple shape) are flawed be-cause they used stimuli that, when analyzed at large spatialscales, contain simple size, orientation, and curvature cues.We constructed stimuli that avoid such confounding andfound learned pop-out. Moreover, we obtained learned pop-out of relative position using a CM paradigm and two VMparadigms, one in which the subjects were cued with thetarget's identity and one in which they were not (the odd-mantask).


In all cases, pop-out with our stimuli could be achievedonly on the basis of relative position information. Therefore,our results show that attention is not necessary for localiza-tion. Hence, our results favor late selection.

Learned pop-out under VM conditions appears to falsifyone of the most stable empirical generalizations in the per-formance literature, namely the learning advantage of CMover VM. Instead, we suggested an explanation that is con-sistent with the search literature to date and that preserves theempirical generalization favoring CM. Specifically, we sug-gested that a group scale cue was used by subjects to performsearch and that the group scale cue was mapped consistently,even in the nominally VM conditions.

According to our theory, the group scale hypothesis, per-ceptual processes preattentively group stimuli according tofeature similarity and contiguity. Although any discriminablefeature could serve as the basis for grouping, we assume thatsubjects have preexperimental biases to group along partic-ular dimensions (e.g., common fate in motion). Subjectslearn, however, to select the dimensions that support usefulgrouping in a particular experimental context. Finally, wesuggest that subjects can use the size of the groups as a cue.When the target forms its own small group and distractorsform a large group, for example, a correct target response canbe made by detecting the small group.

Before practice, subjects found search using our stimulidifficult, and performance was slowed by increasing displaysize. We claim that the initial difficulty was due to the un-familiarity of the stimuli, not to difficulty in discriminatingtargets and distractors. Strong learning occurred becausepractice made the dimension for grouping familiar. We takeperformance after practice to be the true measure of the dis-criminability of the stimuli. Because our stimuli were highlydiscriminable, practice produced pop-out. In short, initialperformance was limited by lack of familiarity, whereas as-ymptotic performance was limited by discriminability.

Our results demonstrate that simple shape information canbe represented preattentively. A different pattern of resultsmay apply for complex shapes, although it is difficult toenvisage how this would be tested without confounding fromdiscriminability factors, as may have occurred in the work ofBeck (1966) and Wolfe et al. (1989).

A strong late-selection position claims that meaning, aswell as shape, is determined preattentively. Our resultsstrengthen the case for the preattentive representation ofshape but do not address the representation of meaning. Be-cause we view perceptual processing as dynamic and hier-archical, however, we suspect that the effects of represen-tations further up the hierarchy are difficult to detect usinga speeded response task, such as visual search.

When responses are not speeded and the perceptual hier-archy can converge on an optimal solution using informationfrom all levels of the hierarchy, the effects of higher levelrepresentations on perceptual processing should be easier toobserve. Work by Prinzmetal (1990; Prinzmetal & Millis-Wright, 1984), for example, showed an effect of word struc-ture on the perception of simple features such as color. Withinour theoretical framework, such results support strong lateselection.

A final point deserves comment. Throughout our discus-sion, we have appealed to grouping as an explanatory prin-ciple. Grouping is clearly a potent phenomenon; its status asan explanatory principle is less clear. Skeptics may take ouremphasis on grouping without a corresponding description ofthe algorithm to be a deus ex machina. We are encouragedby the work of Caelli (1985, 1988; Caelli & Oguztoreli,1988), who described an adaptive computational model ofthe grouping process. His model can group texture regionsaccording to the correlations in the outputs of adaptive fea-ture detectors or filters. Local cooperative and competitiveinteractions group similar representations and split dissimilarrepresentations.

We view the challenge for future work to be explication ofalgorithms used by humans not only to group the outputs ofsimple detectors into objects but also to group objects intolarger units based on all levels of knowledge available to theobserver.

References

Allport, D. A. (1977). On knowing the meaning of words we areunable to report: The effects of visual masking. In S. Dornic (Ed.),Attention and performance VI (pp. 505-533). Hillsdale, NJ:Erlbaum.

Anstis, S. M. (1974). Achart demonstrating variations in acuity withretinal position. Vision Research, 14, 589-592.

Beck, J. (1966). Perceptual grouping produced by changes in ori-entation and shape. Science, 154, 538-540.

Beck, J., & Ambler, B. (1972). Discriminability of differences inline slope and in line arrangement as a function of mask delay.Perception & Psychophysics, 12, 33-38.

Bennett, P. J., & Banks, M. S. (1987). Sensitivity loss in odd-symmetric mechanisms and phase anomalies in peripheral vision.Nature, 326, 873-876.

Bennett, R. G., & Westheimer, G. (1991). The effect of training onvisual alignment discrimination and grating resolution. Percep-tion & Psychophysics, 49, 541-546.

Briand, K. A., & Klein, R. M. (1987). Is Posner's "beam" the sameas Treisman's "glue"? On the relation between visual orientingand Feature Integration Theory. Journal of Experimental Psy-chology: Human Perception and Performance, 13, 228-241.

Broadbent, D. E. (1958). Perception and communication. London:Pergamon Press.

Broadbent, D. E. (1971). Decision and stress. London: AcademicPress.

Broadbent, D. E. (1982). Task combination and selective intake ofinformation. Acta Psychologica, 50, 253-290.

Bundesen, C, & Pedersen, L. F. (1983). Color segregation andvisual search. Perception & Psychophysics, 33, 487^1-93.

Burr, D. C., Morrone, M. C., & Spinelli, D. (1989). Evidence foredge and bar detectors in human vision. Vision Research, 29,419-431.

Butler, B. E., Mewhort, D. J. K., & Browse, R. A. (1991). Whendo letter features migrate? A boundary condition for feature-integration theory. Perception & Psychophysics, 49, 91-99.

Caelli, T. M. (1985). Three processing characteristics of visual tex-ture segmentation. Spatial Vision, 1, 19-30.

Caelli, T. M. (1988). An adaptive computational model for texturesegmentation. IEEE Transactions on Systems, Man, and Cyber-netics, 18, 9-17.

Caelli, T. M. (1991). Region and boundary processing character-istics of visual texture segmentation. Unpublished manuscript,


Department of Computing Sciences, University of Melbourne,Australia.

Caelli, T. M., & Julesz, B. (1978). On perceptual analyzers under-lying visual texture discrimination: Part I. Biological Cybernet-ics, 28, 167-175.

Caelli, T. M., & Julesz, B. (1979). Psychophysical evidence forglobal feature processing in visual texture discrimination. Jour-nal of the Optical Society of America, 69, 675-677.

Caelli, T. M., & Oguztoreli, M. N. (1988). Some task and signaldependent rules for spatial vision. Spatial Vision, 2, 295-315.

Campbell, F. W., & Robson (1968). Applications of Fourier analysisto the visibility of gratings. Journal of Physiology, 197, 551-556.

Carter, R. C. (1982). Visual search with color. Journal of Experi-mental Psychology: Human Perception and Performance, 8,127-136.

Cavanagh, P. (1984). Image transforms in the visual system. In P.C. Dodwell & T. Caelli (Eds.), Figural synthesis (pp. 185-218).Hillsdale, NJ: Erlbaum.

Cave, K. R., & Wolfe, J. M. (1990). Modeling the role of parallelprocessing in visual search. Cognitive Psychology, 22, 225-27'1.

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

Dehaene, S. (1989). Discriminability and dimensionality effects invisual search for featural conjunctions: A functional pop-out. Per-ception & Psychophysics, 46, 72-80.

Deutsch, J. A., & Deutsch, D. (1963). Attention: Some theoreticalconsiderations. Psychological Review, 70, 80-90.

Dobbins, A., Zucker, S. W, & Cynader, M. S. (1987). Endstoppingin the visual cortex as a substrate for calculating curvature. Na-ture, 329, 438-441.

Dobbins, A., Zucker, S. W., & Cynader, M. S. (1989). Endstoppingand curvature. Vision Research, 29, 1371-1387.

Dreher, B. (1972). Hypercomplex cells in the cat's striate cortex.Investigate Ophthalmology, 11, 355-356.

Duncan, J. (1980). The locus of interference in the perception ofsimultaneous stimuli. Psychological Review, 87, 272-300.

Duncan, J. (1990). Boundary conditions on parallel processing inhuman vision. Perception, 18, 457-469.

Duncan, J., & Humphreys, G. W. (1989). Visual search and stimulussimilarity. Psychological Review, 96, 433-458.

Duncan, J., & Humphreys, G. W. (in press). Beyond the searchsurface: Visual search and attentional engagement. Journal ofExperimental Psychology: Human Perception and Performance.18, 578-588.

Egeth, H., & Mordkoff, J. T. (1991). Redundancy gain revisited:Evidence for parallel processing of separable dimensions. In J.Pomerantz & G. Lockhead (Eds.), The perception of structure(pp. 131-143). Washington, DC: American PsychologicalAssociation.

Egeth, H., Virzi, R. A., & Garbart, H. (1984). Searching for con-junctively defined targets. Journal of Experimental Psychology:Human Perception and Performance, 10, 32-39.

Enns, J. T. (1990a). Sensitivity to three-dimensional orientation invisual search. Psychological Science, 1, 323-326.

Enns, J. T. (1990b). Three-dimensional features pop out in visualsearch. In D. Brogan (Ed.), Visual search (pp. 37-45). London:Taylor and Francis.

Enns, J. T., & Rensink, R. A. (1991). Preattentive recovery ofthree-dimensional orientation from line drawings. PsychologicalReview, 98, 335-351.

Eriksen, C. W., & Collins, J. F. (1969). Temporal course of selectiveattention. Journal of Experimental Psychology, 80, 254-261.

Eriksen, C. W., & Spencer, T. (1969). Rate of information pro-cessing in visual perception: Some results and methodological

considerations. Journal of Experimental Psychology Monograph,79(2, Pt. 2).

Eriksen, C. W, Webb, J. M., & Fournier, L. A. (1990). How muchprocessing do nonattended stimuli receive? Apparently very little,but . . . . Perception & Psychophysics, 47, 477-488.

Farmer, E. W., & Taylor, R. M. (1980). Visual search through colordisplays: Effects of target-background similarity and backgrounduniformity. Perception & Psychophysics, 27, 267-272.

Fournier, L. R., & Eriksen, C. W. (1990). Coactivation in the per-ception of redundant targets. Journal of Experimental Psychol-ogy: Human Perception and Performance, 16, 538-550.

Frisk, A. D., & Lloyd, S. J. (1988). The role of stimulus-to-ruleconsistency in learning rapid application of spatial rules. HumanFactors, 30, 35-49.

Frisk, A. D., & Schneider, W. (1984). Consistent attending versusconsistent responding in visual search: Task versus componentconsistency in automatic processing development. Bulletin of thePsychonomic Society, 22, 330-332.

Gurnsey, R., & Browse, R. A. (1987). Micropattern properties andpresentation conditions influencing visual texture discrimination.Perception & Psychophysics, 41, 239-252.

Heathcote, A. (1988). Screen control and timing routines for theIBM microcomputer family using a high-level language. Behav-ior Research Methods, Instruments, & Computers, 20, 289-297.

Heathcote, A., & Mewhort, D. J. K. (1990). Is unbounded visualsearch intractable? Brain and Behavioral Sciences, 13, 449.

Hubel, D. H., & Wiesel, T. N. (1965). Receptive fields and func-tional architecture in two nonstriate visual areas (18 and 19) ofthe cat. Journal of Neurophysiology, 28, 229-289.

Hubel, D. H., & Wiesel, T. N. (1968). Receptive fields and func-tional architecture of monkey striate cortex. Science, 195, 215-243.

Humphreys, G. W., Quinlan, P. T., & Riddoch, M. J. (1989). Group-ing processes in visual search: Effects with single- and combined-feature targets. Journal of Experimental Psychology: General,118, 258-279.

Humphreys, G. W., Riddoch, M. J., & Quinlan, P. T. (1985). In-teractive processes in perceptual organization: Evidence from vi-sual agnosia. In M. I. Posner & O. S. M. Marin (Eds.), Attentionand performance XI (pp. 301-318). Hillsdale, NJ: Erlbaum.

Johnston, J. C., & Pashler, H. (1990). Close binding of identity andlocation in visual feature perception. Journal of ExperimentalPsychology: Human Perception and Performance, 16, 843-856.

Jolicoeur, P., Ullman, S., & MacKay, M. (1986). Curve tracing: Apossible basic operation in the perception of spatial relations.Memory & Cognition, 14, 129-140.

Jones, W. P., & Anderson, J. R. (1982). Semantic categorization andhigh-speed scanning. Journal of Experimental Psychology: Hu-man Learning and Memory, 8, 237-242.

Julesz, B. (1981). Textons, the elements of texture perception, andtheir interactions. Nature, 290, 91-97.

Julesz, B. (1984). A brief outline of the texton theory of humanvision. Trends in Neuroscience, 7, 41^45.

Julesz, B. (1986). Texton gradients: The texton theory revisited.Biological Cybernetics, 54, 245-251.

Kato, H., Bishop, P. O., & Orban, G. A. (1978). Hypercomplex andsimple/complex cell classification in cat striate cortex. Journal ofNeurophysiology, 41, 1071-2095.

Klein, R. M., & Briand, K. A. (1986, May). Allocation of attentionin visual space. Paper presented at the Banff Annual Seminar inCognitive Science, Banff, Alberta, Canada.

Krose, B. J. A., & Julesz, B. (1989). The control and speed of shiftsof attention. Vision Research, 29, 1607-1619.


Larsen, A., & Bundesen, C. (1992). The efficiency of holistic tem-plate matching in recognition of unconstrained handwritten dig-its. Psychological Research/Psychologische Forschung, 54, 187-193.

McKee, S. P., & Westheimer, G. (1978). Improvement in vernieracuity with practice. Perception & Psychophysics, 24, 258-262.

McLeod, P., Driver, J., & Crisp, J. (1988). Visual search for con-junctions of movement and form in parallel. Nature, 332, 154-155.

Mewhort, D. J. K. (1987). Information stores and mechanisms:Early stages in visual processing. In H. Heuer & A. F. Sanders(Eds.), Perspectives on perception and action (pp. 335-357).Hillsdale, NJ: Erlbaum.

Michael, J. (1978). Color-sensitive complex cells in monkey striatecortex. Journal of Neurophysiology, 41, 1250-1266.

Miller, J. O. (1982). Divided attention: Evidence for coactivationwith redundant signals. Cognitive Psychology, 14, 247-279.

Mordkoff, J. T, Yantis, S., & Egeth, D. E. (1990). Detecting con-junctions of color and form in parallel. Perception & Psycho-physics, 48, 157-168.

Morrone, M. C., Burr, D. C., & Spinelli, D. (1989). Discriminationof spatial phase in central and peripheral vision. Vision Research,29, 433^45.

Myers, G. L., & Frisk, A. D. (1987). Training consistent task com-ponents: Application of automatic and controlled processing the-ory in industrial task training. Human Factors, 29, 255-268.

Nakayama, K., & Silverman, G. H. (1986). Serial and parallel pro-cessing of visual feature conjunctions. Nature, 320, 264-265.

Neisser, U. (1967). Cognitive psychology. New York: Appleton-Century-Crofts.

Nelson, J. J., & Frost, B. J. (1985). Intracortical facilitation amongco-oriented, co-axially aligned simple cells in cat striate cortex.Experimental Brain Research, 61, 54-61.

Northdurft, H. C. (1985). Sensitivity for structure gradient in texturediscrimination tasks. Vision Research, 25, 1957-1968.

Pachella, R. G. (1974). The interpretation of reaction time ininformation-processing research. In B. Kantowitz (Ed.), Humaninformation processing: Tutorials in performance and cognition(pp. 41-82). Hillsdale, NJ: Erlbaum.

Parker, A., & Hawken, M. J. (1988). Two-dimensional spatial struc-ture of receptive fields in monkey striate cortex. Journal of theOptical Society of America A, 5, 598-605.

Pashler, H. (1987). Target-distractor discriminability in visualsearch. Perception & Psychophysics, 41, 285-292.

Pashler, H. (1988). Cross dimensional interaction and texture seg-regation. Perception & Psychophysics, 43, 307-318.

Prinzmetal, W. (1990). Neon colors illuminate reading units. Jour-nal of Experimental Psychology: Human Perception and Per-formance, 16, 584-597.

Prinzmetal, W., & Millis-Wright, M. (1984). Cognitive and lin-guistic factors affect visual feature integration. Cognitive Psy-chology, 16, 305-340.

Raab, D. (1962). Statistical facilitation of simple reaction time.Transactions of the New York Academy of Sciences, 24, 574-590.

Rentschler, I. (1985). Symmetry-coded cells in visual cortex? Na-ture, 317, 581-582.

Rentschler, I., Hubner, M., & Caelli, T. (1988). On the discrimi-nation of compound Gabor signals and textures. Vision Research,28, 279-291.

Rentschler, I., & Treutwein, B. (1985). Loss of spatial phase rela-tionships in extrafoveal vision. Nature, 313, 308-310.

Sagi, D., & Julesz, B. (1985). Fast noninertial shifts of attention.Spatial Vision, 1, 141-149.

Sagi, D., & Julesz, B. (1987). Short-range limitation on detectionof feature differences. Spatial Vision, 2,

Schneider, W. (1985). Toward a model of attention and the devel-opment of automatic processing. In M. Posner & O. S. Marin(Eds.), Attention and performance XI (pp. 475^92). Hillsdale,NJ: Erlbaum.

Schneider, W., & Frisk, A. D. (1982a). Concurrent automatic andcontrolled visual search: Can processing occur without resourcecost. Journal of Experimental Psychology: Learning, Memory,and Cognition, 8, 261-278.

Schneider, W., & Frisk, A. D. (1982b). Degree of consistent train-ing: Improvements in search performance and automatic processdevelopment. Perception & Psychophysics, 31, 160-168.

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

Shiffrin, R. M. (1975). The locus and role of attention in memorysystems. In P. M. A. Rabbitt & S. Dornic (Eds.), Attention andperformance V (pp. 168-193). London: Academic Press.

Shiffrin, R. M., & Schneider, W. (1977). Controlled and automatichuman information processing: II. Perceptual learning, automaticattending, and a general theory. Psychological Review, 84, 127-190.

Sperling, G. (1960). The information available in brief visual pre-sentations. Psychological Monographs, 74 (11, Whole No. 498).

Steinman, S. B. (1987). Serial and parallel search in pattern vision.Perception, 16, 389-398.

Teichner, W. H., & Krebs, M. J. (1974). Visual search for simpletargets. Psychological Bulletin, 81, 15-41.

Townsend, J. T. (1971). A note on the identifiability of parallel andserial processes. Perception & Psychophysics, 10, 161-163.

Townsend, J. T. (1976). Serial and within-stage independent parallelmodel equivalence on the minimum completion time. Journal ofMathematical Psychology, 14, 219-238.

Townsend, J. T. (1990). Serial vs. parallel processing: Sometimesthey look like Tweedledum and Tweedledee but they can (andshould) be distinguished. Psychological Science, I, 46-54.

Townsend, J. T., & Ashby, F. G. (1983). The stochastic modelingof elementary psychological processes. New York: CambridgeUniversity Press.

Treisman, A. (1960). Contextual cues in selective listening. Quar-terly Journal of Experimental Psychology, 12, 242-248.

Treisman, A. (1969). Strategies and models of selective attention.Psychological Review, 76, 282-299.

Treisman, A. (1982). Perceptual grouping and attention in visualsearch for features and for objects. Journal of Experimental Psy-chology: Human Perception and Performance, 8, 194-214.

Treisman, A. (1988). Features and objects: The fourteenth Bartlettmemorial lecture. The Quarterly Journal of Experimental Psy-chology, 40A, 201-237.

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

Treisman, A., & Gormican, S. (1988). Feature analysis in earlyvision: Evidence from search asymmetries. Psychological Re-view, 95, 15-48.

Treisman, A., & Patterson, R. (1984). Emergent features, attention,and object perception. Journal of Experimental Psychology: Hu-man Perception and Performance, 10, 12-31.

Treisman, A., & Sato, S. (1990). Conjunction search revisited. Jour-nal of Experimental Psychology: Human Perception and Per-formance, 16, 459^78.

Treisman, A., & Souther, J. (1985). Search asymmetry: A searchdiagnostic for preattentive processing of separable features. Jour-nal of Experimental Psychology: General, 114, 285-310.

Treisman, A., Sykes, M., & Gelade, G. (1977). Selective attentionand stimulus integration. In S. Dornic (Ed.), Attention and per-formance VI (pp. 333-361). Hillsdale, NJ: Erlbaum.


Tsal, Y. (1983). Movements of attention across the visual field.Journal of Experimental Psychology: Human Perception andPerformance, 9, 523-530.

Tsotsos, J. K. (1990). Analyzing vision at the complexity level.Brain and Behavioral Sciences, 13, 423-^469.

van der Heijden, A. H. C. (1987). Central selection in vision. In H.Heuer & A. F. Sanders (Eds.), Perspectives on perception andaction (pp. 421-446). Hillsdale, NJ: Erlbaum.

von der Heydt, R., Peterhaus, E., & Baumgartner, G. (1984). Illu-sory contours and cortical neuron responses. Science, 224, 1260-1262.

Vorberg, D. (1977). On the equivalence of parallel and serial mod-els of information processing. Paper presented at the Tenth Math-

ematical Psychology Meetings, Los Angeles, CA.Westheimer, G. (1982). The spatial grain of the perifoveal visual

field. Vision Research. 22, 157-162.Wiesel, T. N. (1960). Receptive fields of ganglion cells in the cat's

retina. Journal of Physiology, 153, 583-594.Wolfe, J. M, Cave, K. R.. & Franzel, S. J. (1989). Guided search:

An alternative to the feature integration model of visual search.Journal of Experimental Psychology: Human Perception andPerformance. 15. 419-433.

Received July 16, 1991Revision received July 13, 1992

Accepted August 12, 1992 •

AMERICAN PSYCHOLOGICAL ASSOCIATIONSUBSCRIPTION CLAIMS INFORMATION Today'sDate:_

We provide this form to assist members, institutions, and nonmember individuals with any subscription problems. With theappropriate information we can begin a resolution. If you use the services of an agent, please do NOT duplicate claims throughthem and directly to us. PLEASE PRINT CLEARLY AND IN INK IF POSSIBLE.

PRINT FULL NAME OR KEY NAME OF INSTITUTION MEMBER OR CUSTOMER NUMBER (MAY BE FOUND ON ANY PAST ISSUE LABEL)

DATE YOUR ORDER WAS MAILED (OR PHONED)

CHECK/CARD CLEARED DATE:.STATE/COUNTRY

YOUR NAME AND PHONE NUMBER

TITLE

(If possible, send a copy, front and back, of your cancelled check to help us in our researchof your claim.)

ISSUES: MISSING DAMAGED

VOLUME OR YEAR NUMBER OR MONTH

Thank you. Once a claim is received and resolved, delivery of replacement issues routinely takes 4-6 weeks.

———————— (TO BE FILLED OUT BY APA STAFF) —————————

DATE RECEIVED:.ACTION TAKEN: _STAFF NAME:

DATE OF ACTION: _INV. NO. & DATE:LABEL NO. & DATE:

Send this form to APA Subscription Claims, 750 First Street, NE, Washington, DC 20002-4242

PLEASE DO NOT REMOVE. A PHOTOCOPY MAY BE USED.

Representation and selection of relative position.

Documents

Transcript of Representation and selection of relative position.