Journal of Experimental Psychology: Human ...

Journal of Experimental Psychology: Human Perception and Performance Visual Search for Singleton Targets Redundantly Defined in Two Feature Dimensions: Coactive Processing of Color-Motion Targets? Joseph Krummenacher and Hermann J. Müller Online First Publication, August 4, 2014. http://dx.doi.org/10.1037/a0037560

CITATION Krummenacher, J., & Müller, H. J. (2014, August 4). Visual Search for Singleton Targets Redundantly Defined in Two Feature Dimensions: Coactive Processing of Color-Motion Targets?. Journal of Experimental Psychology: Human Perception and Performance. Advance online publication. http://dx.doi.org/10.1037/a0037560

Journal of Experimental Psychology: Human Perception and Performance 2014, Vol. 40, No. 4, 000

© 2014 American Psychological Association 0096-1523/14/$12.00 http://dx.doi.org/10.1037/a0037560

This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

Visual Search for Singleton Targets Redundantly Defined in Two Feature Dimensions: Coactive Processing of Color-Motion Targets? Joseph Krummenacher

Hermann J. Müller

Ludwig-Maximilians-Universität, Munich

Ludwig-Maximilians-Universität, Munich, and Birkbeck College, University of London

In 2 visual search experiments, the role of feature contrast/saliency signals in generating detection responses to singleton feature targets in visual search was investigated using the redundant-target paradigm. Experiment 1 showed that coactive integration of dimensional signals is not restricted to targets defined on the color and orientation dimensions; rather, targets involving any of the combinations of color, orientation, and motion, are integrated coactively, as evidenced by violations of Miller’s (1982) race model inequality. Experiment 2 replicated the findings of Experiment 1 for color-motion targets, with the target items’ luminance adjusted, individually for each observer, to that of the distractors. The evidence for coactive processing of motion (saliency) with color and, respectively, orientation (saliency) signals suggests that, at variance with a recent suggestion by Li (2002; Koene & Zhaoping, 2007), signal integration in feature search tasks occurs at a stage following initial feature coding in primary visual cortex (V1), even though feature contrast computations in V1 may well contribute to saliency coding. In sum, the results suggest that detection responses were based on an integrated, overall-saliency representation indicating the presence of an odd-one-out item in the display, consistent with the dimensionweighting account of visual search (Müller et al., 1995, 2003). Keywords: redundancy gains, singleton search, dimension-based saliency processing, color-motion integration, dimension weighting account

Instead, selection may be based directly on the output of featurecoding cells in early visual areas, specifically in V1 (the “V1 hypothesis”; see below for details). The aim of the present study was to reexamine this issue, by reviewing the existing evidence for (the core assumptions made by) each of these opposing accounts and, based on the review, devising and carrying out a crucial experiment designed to decide between them.

The present study is concerned with the way in which the visual system selects conspicuous items (e.g., a conspicuous “target” presented among distractors in a search display) for further, focalattentional analysis and response-related processing. Saliencebased theories of visual selection assume that the allocation of focal attention is guided by a map of the display that represents the overall-conspicuity of each item within its surround: The item selected is that which achieves the highest salience. In essence, a whole class of theories of visual search reflect attempts to specify how salience is computed by the preattentive visual system (a prominent example being the guided search model; see next section for details). This—mainstream—view has recently been challenged by other approaches positing that no special saliency representation is necessary to account for attentional selection.

Saliency-Based Visual Selection Visual search for an object that differs from other objects by a conspicuous feature (e.g., a red apple among green apples) is efficient; phenomenally, such objects, or feature singletons, “pop out” of the display. Theories of visual search such as Wolfe’s (1994, 2007) guided search (GS) model explain efficient search for feature targets by a mechanism that allocates focal attention on the basis of a topographic saliency map (see also Itti & Koch, 2000; Koch & Ullman, 1985; Soltani & Koch, 2010). Saliency activations are generated in parallel in a set of dimension-specific modules (for color, motion, orientation, size, etc.; see Wolfe & Horowitz, 2004, for a taxonomy of dimensions) by contrasting feature values among nearby objects. The resulting “feature contrast” signals are integrated across dimensions into an overallsaliency map that controls the allocation of focal attention. Attention is directed to the location with the highest saliency activation. If the object found at this location is a distractor (i.e., a nontarget), this activation is suppressed, so that focal attention shifts to the next most salient location, and so on until the target is either detected or some criterion is reached for terminating the search. On this account, visual search for feature singleton targets is efficient

Joseph Krummenacher, Neuro-Cognitive Psychology, Department of Psychology, Ludwig-Maximilians-Universität, Munich; Hermann J. Müller, Department of General and Experimental Psychology, LudwigMaximilians-Universität, Munich, and Department of Psychological Sciences, Birkbeck College, University of London. We thank Zhaoping Li, one anonymous reviewer, and James T. Enns (editor) for their helpful comments on an earlier version of the article. The study was supported by Swiss National Science Foundation (Joseph Krummenacher) and German National Science Foundation (Hermann J. Müller, Joseph Krummenacher) grants. Correspondence concerning this article should be addressed to Joseph Krummenacher, Neuro-Cognitive Psychology, Department of Psychology, Ludwig-Maximilians-Universität, Munich, Leopoldstrasse 13, 80802 Munich. E-mail: [email protected] 1


2

¨ LLER KRUMMENACHER AND MU

because saliency coding gives rise to only one activation peak: at the location of the target, so that focal attention is immediately directed to it. Note, however, that saliency coding is not solely dependent on physical stimulus parameters (such as feature contrast). For example, Maljkovic and Nakayama (1994) showed that reaction times (RTs) to feature singletons are influenced by whether the definition of the target, and distractors, on a given experimental trial is the same as or different from that of the target, and distractors, on the previous trial. Repetition of the target (and distractor) features across trials generates a RT benefit, whereas swapping the target and distractor features incurs a cost. Maljkovic and Nakayama (1994) explained this pattern in terms of “priming of pop-out,” that is: rather than being purely stimulus-driven, saliency coding is subject to internal (short-term buffered) feature biases. [Note that GS also assumes a component of top-down feature-biasing in singleton search if the target-defining features are few and known in advance (see Wolfe, Butcher, Lee, & Hyle, 2003).] In a similar paradigm, with constant distractor features (e.g., green vertical bars) but variable target features (e.g., color targets: red or blue vertical bar; orientation target: left- or right-tilted green bar), Müller and colleagues (e.g., Found & Müller, 1996; Müller, Heller, & Ziegler, 1995) observed that target feature change (vs. repetition) relative to the previous trial produced a substantial and significant RT cost only when the new target was defined by a feature within a different dimension (e.g., color target preceded by orientation), not by a different feature within the same dimension (e.g., red target preceded by blue target). This and other effects led Müller and his colleagues to propose a “dimension-weighting” (DW) account of feature contrast signal integration into the overall-saliency map. On this account, the weight allocated to a particular dimension determines how rapidly a feature contrast signal in this dimension achieves (overall) saliency. If a target on a given trial is defined in a particular dimension, the weight assigned to signals in this dimension is increased, whereas that assigned to other dimensions is decreased. Because the weight distribution thus established persists across short periods of time, this will speed up saliency-based selection times for targets defined in a repeated dimension, and slow down selection times for targets in a changed dimension. Because cross-dimensional signal integration operates on dimensions-specific feature contrast signals (signaling that there is a feature difference in, say, color, but not what the exact color feature is that gave rise to this difference), this weighting dynamics will generate only dimension-based, but not feature-based, change effects. Importantly, the DW account assumes that weighting influences a stage where feature contrast information from different dimensions is integrated into a unitary (overall-) saliency map. This assumption of dimension-based signal integration was examined by Krummenacher, Müller, and Heller (2001, 2002) using a redundant-signals paradigm. Observers were presented with either one (nonredundant) target signal: a singleton target only defined by color or only by orientation, or with dual (redundant) target signals: a singleton target defined by both color and orientation. In more detail, distractors were always green vertically oriented bars; color targets were red or blue vertical bars, orientation targets were left- or right-tilted green bars; redundant targets differed from the distractors both by color and orientation (e.g., red and left-tilted bar). The task required responding positively to any target signal,

that is, detecting one signal was sufficient to respond “targetpresent.” In redundant-signals paradigms generally, and the specific variation used by Krummenacher et al., responses are typically faster to redundantly defined signals (color-and-orientation targets in Krummenacher et al.) than to singly defined signals (color-only or orientation-only targets), which is referred to as the “redundant-signals effect.” In principle, this effect may be generated by three types of mechanism. In a single-channel processing architecture, dimensional feature contrast signals (which are computed in parallel) are checked serially, one after the other (whether in random or nonrandom order). Nevertheless, RTs would be reduced on redundant-, relative to nonredundant-, signal trials, because a detection response can be given upon checking only one dimension; by contrast, on trials with a nonredundant signal, statistically 1.5 dimensions would need to checked to establish target presence (target detection would require one check in 50% of the trials, but two checks in the other 50%). In parallel-channel architectures, dimensional feature contrast signals computed in parallel may enter a race to trigger some output module (i.e., the module is activated by either one or the other signal). As the response is always triggered by the faster of two signals, average RTs are expected to be faster on redundant-signal trials than on trials with a nonredundant signal—an effect referred to as “statistical facilitation” (Raab, 1962). Alternatively, independent and parallel signals may jointly (co)activate an output module (i.e., the module sums or integrates the two signals), also producing RTs that are faster than the average RTs to nonredundant signals. To differentiate between the two latter processing architectures, Miller (1982) proposed analyzing the entire RT distributions and comparing the probabilities of responses on nonredundant versus redundantsignals trials having occurred by a given point in time according to the following “race model inequality” (RMI): P(RT⬍t|T1&T2) ⱕ P(RT⬍t|T1) ⫹ P(RT⬍t|T2), where t is the time since display onset and T1 and T2 are Target Signals 1 and 2; that is, the conditional probability of observing a RT shorter than time t given a redundant target (defined by signals T1 and T2), is less than or equal to the sum of the probabilities of observing a RT shorter than t given a nonredundant target (defined by signal T1 and, respectively, T2). Miller (1982) showed that if the RMI holds, multiple signals trigger the response in a parallel race (at least, there is no evidence to the contrary). By contrast, if the RMI is violated, the parallelrace assumption can be rejected, arguing in favor of independent signals coactivating a common output module. In order to investigate whether feature contrast signals computed in parallel in multiple dimensions trigger the detection response in a parallel race or a coactive manner, Krummenacher et al. (2001, 2002) applied the redundant-signals logic to a singleton feature search task, with color-only targets, orientation-only targets, and (redundantly defined) color-and-orientation targets. Krummenacher et al. found that mean RTs to redundant targets were significantly faster than RTs to (the “faster” of the) nonredundant targets. Importantly, Miller’s (1982) RMI was violated, providing evidence for coactive processing, or integration, of dimension-based feature contrast signals. Two other findings relating to the mechanism underlying the integration of dimension-based signals are noteworthy. Krummenacher et al. (2002) also used a condition in which two (separate) target items were presented (at different locations), with one dif-


REDUNDANCY GAINS IN COLOR-MOTION SINGLETON SEARCH

fering from distractors on the color and the other on the orientation dimension or with both targets differing from distractors on the same dimension (e.g., a red and a blue target). In addition, the distance between dual target items was varied in three steps: dual items presented next to each other, or separated by one or, respectively, two distractors. The results showed that mean RTs to dual redundant targets were significantly faster than those to (the “faster” of the) nonredundant targets, whether dual targets were defined in different dimensions (one color, one orientation) or the same dimension (either two color or two orientation). Importantly, however, RMI violations were observed only for dual targets defined in different dimensions (not for dual targets defined in the same dimension), and, with different-dimension dual targets, the magnitude of the violations was modulated by the intertarget spatial distance: the closer the two target locations the larger the probability of coactive processing. Krummenacher et al. (2002) took this pattern to indicate that (a) the signals that are integrated are dimension-, rather than feature-, specific in nature (there was no coactivation by multiple features defined on the same dimension); and (b) the integration is location-specific (or spatially scaled). Krummenacher and colleagues interpreted this set of findings within the theoretical frameworks of GS and DW: Saliency coding is based on dimension-based feature contrast signals that are then integrated, across dimensions, into an overall-saliency map guiding the allocation of focal attention. Importantly, this representation is situated at a stage independent of and subsequent to the initial representation of features.

The Challenge: The V1 Hypothesis Recently, however, Zhaoping Li (2002; Koene & Zhaoping, 2007) proposed an alternative account that calls into question critical assumptions of the DW (and GS) models: Namely, that dimension-based feature contrast signals are integrated, in a weighted manner, at a postfeatural stage of processing: an overallsaliency map. In more detail, Koene and Zhaoping (2007) argued that saliency coding and redundancy gains are associated with the activation of neurons in visual area V1 (the “V1 hypothesis;” Li, 2002; see also Koene & Zhaoping, 2007). It is well established that local visual features, such as color and orientation, are initially represented by V1 cells tuned to particular feature values. On the V1 hypothesis, saliency signals are generated automatically as a result of “iso-feature inhibition” among cells coding features that are shared by multiple (nontarget) items within a circumscribed display region. For example, in visual search for singleton feature targets with, say, a red vertical target item presented within an array of green vertical distractor items (see Figure 1), cells responding to vertical orientation would inhibit each other, as all items are vertical (iso-feature inhibition). Likewise, cells coding green (items) would inhibit each other. By contrast, the cell coding the (only) red item remains the only one highly active (as it does not receive inhibitory input), thus acting as a saliency “pointer” to the target’s location. The (saliency) activation of the nonsuppressed cell is then transmitted to a (winner-take-all) mechanism controlling the allocation of focal attention; essentially, this mechanism selects the most active cell, or cluster of cells, in a “winnertake-all” process—that is, in the manner of a parallel race.1 Importantly in the present context, the V1 hypothesis provides an alternative explanation, to that proposed by Krummenacher et

a

b

3

c

d

Figure 1. Illustrations of the search displays presented in Experiment 1. Shown is the color and motion condition with (a) singly defined color (red), (b) motion (sinusoidal motion indicated by the arrow), and (c) redundantly defined color and motion targets; distractors in target-present and targetabsent trials were green vertical bars; a target-absent display is shown in (d). Note that the stimulus size and distances are not to scale.

al. (2001, 2002), of the “coactivation” RT gains for redundantly defined relative to singly defined targets—at least targets redundantly defined by color and orientation (as in Krummenacher et al., 2002). The V1 hypothesis explains this finding by assuming that the location of a target item differing from distractors by both color and orientation—for example, a red left-tilted (target) bar among green vertical (distractor) bars—is signaled by activation of three types of cell (at the same location): cells tuned to red, cells tuned to right-ward tilt, and cells tuned to both (i.e., a conjunction) of red and right-ward tilt. Crucially, on the V1 account, it is the latter type of (“conjunction”) cell that is assumed to explain “coactivation” gains. Indeed, there is physiological evidence for the existence of such cells tuned, for example, to a conjunction of orientation with color (see Horwitz & Albright, 2005; Livingstone & Hubel, 1984; Ts’o & Gilbert, 1988). Although the underlying logic with regard to RMI tests is not elaborated in detail in Zhaoping Li’s articles (Li, 2002; Koene & Zhaoping, 2007), it may be illustrated using the above example in which the redundant target was a red left-tilted bar, and the nonredundant targets either a red vertical bar (color target) or a right-tilted green bar (orientation target), among green vertical (distractor) bars. According to Li, in the two single-feature conditions, the target would be signaled by either a “red” cell or a “right-tilt” cell, and a redundant target by three types of cell: a “red” cell, a “right-tilt” cell, and a “red &right-tilt” cell. In contrast, according to Krummenacher et al., the redundant target would be signaled by only two types of cell: a “red” cell and a “right-tilt” cell. Now, if one assumes a parallel race among the various detectors for focal-attentional selection, and assuming that (with matched signal strengths in the two dimensions) the distribution of race completion times is the same for all types of cell, then having a third competitor in the race, as in Zhaoping Li’s model, would produce statistical gains—potentially even at the fast end of the completion time distribution (but see below)— compared with a race with only two competitors, as assumed by Krummenacher et al.: If one has three, on average equally fast horses in a race, then the likelihood that one will complete the race rapidly is greater than if one enters only two horses in the race! Thus, what Krummenacher et al. took as evidence for coactivation (namely, RMI violations under the assumption of a race between 1 Note that, because saliency is signaled by the activation of a featuretuned cell (in the example, a color-sensitive cell tuned to “red”), the saliency signal as such is not “feature-less,” that is, selection would immediately make featural information about the target available.


4


two competitors) may in fact just reflect the outcome of a parallel race among three competitors. Note that real coactivation gains, violating race-model assumptions (Miller, 1982), would be seen in such a model only if the third type of cell, the feature conjunction cell, were to complete the race faster than the single-feature cells. In this case, one should see a shift in the completion time distribution toward the fast end on the time axis, that is, the “very fastest” completion times should be faster than those achieved by single-feature cells. It is not clear whether there is a physiological evidence for this. Also, it is not clear how feature conjunctions are actually coded. Koene and Zhaoping (2007) consider two scenarios: (a) conjunctions units being activated directly by double features (see their Figure 1b), that is, conjunction units are situated at the same coding level as single-feature units; and (b) alternatively, conjunction units are coactivated by (activated) single-feature units, that is, they are situated at a subsequent coding level to the single-feature units (see their Figure 8). Although Koene and Zhaoping (2007) advocate scenario one, arguably, Scenario 2 remains feasible—importantly, with both single features and feature conjunctions being coded in the same area: V1. In more detail, in this scenario, cells coding, say, color-orientation conjunctions are coactivated by color-only and orientation-only cells—that is, their activation would require one further processing step (involving at least one further synaptic connection). Logically then, such a conjunction cell would not be activated (above threshold) prior to a single-feature cell having been triggered. Further, assuming that a conjunction cell is triggered only if it receives “conjunctive” input from two singlefeature cells (e.g., one color and one orientation cell), its activation would be dependent on the slower coded feature.2 Given this, a conjunction cell would enter the race for selection relatively late. Crucially, on the V1 hypothesis, there should be no (real or spurious) coactivation gains when redundant targets are defined by a conjunction of color or, respectively, orientation with motion. The reason is that at the level of V1, color and orientation on the one hand and motion on the other are coded in neurophysiologically segregated processing streams. Lateral geniculate nucleus (LGN) magno-cellular layers project to layer 4C␣, LGN parvocellular layers project to layer 4C␤ of V1. Within V1, magno- and parvo-cellular projections remain segregated. Thus, while color and orientation are represented within the same (V1) streams and layers, motion and color or, respectively, motion and orientation information is not. In other words, there would be no cells responding to, say, both a particular color and a particular (direction of) motion, or a particular orientation and a particular motion. Concerning the latter, however, a cell tuned to, say, upwarddirected motion would be activated less by a vertical bar moving upward through its receptive field compared to a horizontal bar (in fact, the cell will not respond at all to a vertical bar if the bar’s ends remain outside its receptive field, RF, during the movement, as it may happen given the small RFs of V1 cells). In this sense, a motion-direction-tuned cell would also exhibit orientation sensitivity, that is, act like a conjunction cell. Therefore, the critical test of the V1 hypothesis would involve combinations of color with motion.3,4 Thus, if there were significant violations of the RMI only for redundant targets defined by color and orientation (and, maybe, orientation and motion), but not for targets defined by color and motion, this would provide strong support for Zhaoping Li’s V1

account (according to which real or spurious coactivation effects are due to color-orientation conjunction cells coming into play). Comparing redundancy gains for color-and-orientation targets on the one hand with those for color-and-motion and, respectively, orientation-and-motion targets on the other hand, Koene and Zhaoping (2007) indeed found evidence for the dissociation expected on the V1 hypothesis: The RT distribution analyses revealed redundancy gains for color and orientation signals (and for orientation and motion signals), but not for color and motion signals (see also Poom, 2009, who reported a dissociation between orientation-motion targets, which produced RMI violations, and color-motion targets, which did not; note, however, that Poom found no evidence of coactivation for orientation-color targets, at variance with both Krummenacher et al., 2001, 2002, and Koene & Zhaoping, 2007). Koene and Zhaoping took this to strongly argue against the integration account of Krummenacher et al. (2001, 2002), according to which coactivation gains should be general, that is, evident for all combinations of dimensions. In fact, Koene and Zhaoping (2007) proposed that “. . . there is no longer a need for a separate master saliency map in addition to the feature maps. This is because the attentional selection system could simply ignore (or be blind to) the separation between the single- and double-feature maps, treat all units (or neurons) from these feature maps as if they were from a single neural population, find the most active unit or neuron among them, and direct attention to its RF [receptive field]” (p. 10). Note, however, that Koene and Zhaoping’s (2007) results may not be as clear-cut as they appear at first sight. In fact, there are several issues pertaining to the stimuli used as well as the analyses conducted. To start with the latter, Koene and Zhaoping did not formally test for violations of Miller’s (1982) RMI. Rather, they compared the entire distribution of observed redundant-target RTs with the distribution derived by Monte Carlo simulation randomly drawing from the sets of the two (observed) single-feature RTs and then selecting the shorter of the two drawn RTs (to simulate a race between the two respective features). As can be seen from their Figure 5, compared with the predicted RTs, the observed distributions of redundant-target RTs were relatively more skewed toward 2 Arguably, for the V1 hypothesis to have explanatory power, the critical “conjunction” cells would have to be logical AND detectors, rather than an INCLUSIVE-OR detectors. An AND detector would be activated only if both features are (simultaneously) present; an INCLUSIVE-OR detector, by contrast, would be activated whether (only) one or the other or both features are present. If the cells were INCLUSIVE-OR detectors, they would equally come into play whether the target is defined by only one of the features or by both features. From the selection rule described in Zhaoping and Snowden (2006), it becomes clear that Zhaoping Li conceives of the critical cells as AND detectors. 3 Interactions between the different pathways may occur only from V2 onward. In V2, there are connections between thick stripes (receiving input from layer 4B [magno-] of V1) and thin stripes (receiving input from [parvo-] blob and interblob regions of layers 2 and 3 of V1). Further, in V2 there are back-projections from V4 to the thick stripes. Further upstream, there are direct connections between area MT and V4 and V3. 4 Note that according to Li (personal communication, June 6, 2011), coactivation gains would actually be expected for conjunctions of orientation with motion, based on evidence of V1 cells responding to both orientation and motion (although the conclusions to be drawn from this evidence may be subject to debate). However, for the reasons outlined above, we agree that the critical test would involve combinations of color with motion.



the fast end of response times, in particular for color-orientation (CO) and orientation-motion (OM) targets. However, there were in fact no discernible differences between the observed and the (on the race model) expected RT frequencies for the “very fastest” time bins, in any of the three conditions (CO, CM, OM). From this it is clear that, had Koene and Zhaoping tested for RMI violations, they would have found none at the fast end of the RT distributions—and it remains doubtful whether they would have found any with longer response times (this is so because the RMI sets the probability of a response to a redundant target having occurred by time t in relation to the summed probabilities of a response having occurred to the respective single-feature targets). Given this, and contrary to their claim, Koene and Zhaoping did not, in fact, replicate the findings of Krummenacher et al. for color-andorientation targets. On the other hand, Koene and Zhaoping’s data show deviations of the observed from the expected RT probabilities for the RT bins just before and at the peak of the distributions [p(observed) ⬎ p(expected)]. This would be consistent with the idea that another competitor (or actor) had entered the race (or come into play), albeit with some delay, that is, too late to influence the “very fastest” response times (whether by statistical facilitation or coactivation). In terms of an alternative take of the V1 hypothesis, this could indeed be the conjunction cells, whose activation would be secondary to those of the single-feature cells (see above and Figure 8 in Koene & Zhaoping, 2007). The reason why the Krummenacher et al.’s (2001, 2002; see also Zehetleitner et al., 2008) results for color-and-orientation targets were not replicated by Koene and Zhaoping (2007) may have to do with their stimulus material. In contrast to Krummenacher et al., who used maximum (color and orientation) feature contrast, Koene and Zhaoping used feature contrast levels for singling out their target stimuli from among the distractors that were chosen to be nonoptimal (so as to leave room for improvement by redundant target coding), yielding relatively low response speeds: average RTs close to 700 ms in a two-alternative forcedchoice (2AFC), left/right hemi-field localization decision for single-feature targets, which compares with some 400 ms in a 2AFC target-present/-absent detection in the studies of Krummenacher et al. Thus, their target stimuli were relatively nonsalient (with target saliency being reduced, e.g., as a result of iso-feature suppression: the closer the target is featurally to the distractors, the more it would itself be affected by iso-feature suppression), and it was actually not ensured that they were true “pop-out” targets in terms of search-rate criteria (i.e., detected in a spatially parallel manner, as evidenced by a near-zero slope of the function relating RT performance to the number of display elements). It is known that, in terms of search rate, target detection becomes less efficient as feature contrast decreases (see, e.g., Wolfe, 1998). In this case, detection is thought to increasingly rely on serial processing of limited regions of the display (in the extreme, of a single item), because there is an increasing likelihood that, due to noise in the coding process, one or more distractors will achieve a higher saliency than the target in the first pass (even a redundant target, whose saliency may be affected by increasing iso-feature suppression in two dimensions), delaying target selection. Given this, the stimuli used by Koene and Zhaoping (2007) might not have been optimal for examining the question at issue (i.e., testing for RMI violations). Assuming that saliency computations are completed

5

within a certain amount of time following display onset, selecting and checking a wrong (distractor) item prior to the target would diminish the advantage that could be derived from redundant target coding—the reason being that by the time search progresses to the target, even a single-feature target would have achieved abovethreshold saliency. That is, redundant target coding would influence only the speed of the first selection decision, but not that of any subsequent decisions that would be made at a time by which all types of target, whether redundantly or nonredundantly defined, would be singled out by above-threshold activity— even though the activation peak would still be higher for redundant targets; this is predicated on the (reasonable) assumption that the threshold would be set low enough for a single-feature target (presented on one half of the trials in the study of Koene and Zhaoping and two thirds of the trials in the studies of Krummenacher et al.) to be able to cross it.5 In summary then, the failure of Koene and Zhaoping (2007) to replicate the findings of Krummenacher et al. (2001, 2002; Zehetleitner et al., 2011) may not be surprising. Given this, the present study was designed to reexamine— using a better suited task and procedure—whether violations of the RMI would be found not only for color-orientation targets, but also for colormotion (as well as orientation-motion) targets. If RMI violations were found for all types of redundant targets, this would strengthen saliency summation accounts, with signal integration occurring at a relatively late level, such as that of an overall-saliency map.

Experiment 1 Experiment 1 was designed to test the prediction of the DW account that detection of any redundant combination of dimensional signals in singleton feature tasks expedites search RTs compared with RTs to nonredundant signals. In Experiment 1, RTs to redundant targets defined by color and orientation, color and motion, and orientation and motion were compared with the respective RTs to nonredundant targets. Note that Experiment 1 used the same, stationary color and orientation stimuli (CO condition) as used by Krummenacher et al. (2001, 2002, where both the distractors and the target were stationary), with maximum color and maximum orientation feature contrast. To keep the overall display parameters similar in the color-motion (CM) and orientation-motion (OM) conditions, the target was the only stimulus that moved: It oscillated sinusoidally, on the horizontal axis. This created a target motion signal, that is: Sinusoidally varying motion contrast (as a result of diminishing/increasing separation) relative to the nearby, stationary distractors. The movement parameters were set (in preexperimental testing) such as to produce pop-out as rapidly as the (maximum contrast) color and orientation 5 On the other hand, with relatively nonsalient targets, conjunction cells, which may come into play only relatively “late” (see above), may have a greater impact on selection decisions than with highly salient targets— accounting for the pattern of findings of Koene and Zhaoping (2007). Assuming a fixed activation threshold for making a selection decision (set in accordance with the activation level achievable by a single-feature target), then, in a coactivation model, the single-feature inputs would produce a rise in the activation level prior to any conjunction input contributing to evidence accumulation. By contrast, if the target is highly salient, the conjunction input may not come “on-stream” before the decision is made (i.e., it may come so late as to have little influence on the decision).

6



features. Note that, in terms of iso-feature suppression, this motion stimulus set-up should be similar to the situation realized by Koene and Zhaoping (where the target moved horizontally in one direction and the distractors in counterdirection): Stationary distractors may be only weakly (if at all) represented in the motion system, with the target being the only item achieving a high activation. Furthermore, in the present study, all combinations of (redundant) features (CO, CM, OM) were blocked. These design features were chosen to ensure effective target pop-out in all conditions and optimal conditions for coactivation effects (i.e., violations of the RMI at the “very fast” end of the RT distributions) to be observable.

Method Participants. Ten observers took part in Experiment 1 (age range 20 to 35 years, M ⫽ 24.6 years; eight female). All observers had normal or corrected-to-normal vision, including color vision. They provided written informed consent, and were paid at a rate of €8.00 (about $11.00) per hour or received course credit for participating. Apparatus. Observers were seated at a distance of approximately 57 cm from the display with eye-screen distance maintained through the use of a chin rest. The experiment was carried out in a darkened room with dim background illumination to prevent screen reflections. Stimuli were presented on a Sony Trinitron 17” color monitor (at a frame rate of 60 Hz) controlled by a Dell Pentium computer. Observers responded by pressing the left or right button of a purpose-built response keypad with the index finger of the left or right hand, respectively. RTs and errors were recorded by the computer, and a computer-generated error feedback (a short “bleep”) was given when an erroneous response was made; no feedback was given after correct responses. Stimuli and timing. The search display consisted of 36 bar elements, each about 1° of visual angle in height and 0.25° in width, arranged in the cells of a virtual 6 ⫻ 6 matrix (for examples, see Figure 1). Search items were slightly jittered, with a maximum displacement of 0.4° horizontally and vertically relative to the exact center of the cell. The horizontal and the vertical distances between the centers of display objects ranged between 1.6° and 2.0° of visual angle. All nontarget items were green vertical bars, targets were either defined by color (red vertical bar), orientation (green 45° right-tilted), motion (sinusoidal, in horizontal direction, with a maximum amplitude of about ⫾0.6° of visual angle and an average velocity of about 2°/s), color and orientation (red and right-tilted bar), color and motion (red and moving bar), and orientation and motion (right-tilted and moving bar). The set of stimuli was separately presented in three conditions consisting of two singly defined targets and the target produced by combining the two single target dimensions: (a) color, orientation, color and orientation, in the following referred to as the CO (color and orientation) condition; (b) color, motion, color and motion, CM; and (c) orientation, motion, orientation and motion, OM. Targets were presented at a randomly chosen location within the inner 4 ⫻ 4 cells of the display matrix so as to ensure that target items were surrounded by the same number of neighboring distractors (i.e., top avoid edge effects). Observes were not informed of this restriction.

The stimuli (targets and distractors) were physically isoluminant at 4.3 cd/m2, with luminance adjusted by means of changes of the RGB values of the colors red and green. A trial started with the simultaneous onset of all display elements and the display stayed on the screen until the observer responded. The intertrial interval was 1,000 ms. Design and procedure. Targets differed from the distractors in either a single dimension (color: red; orientation: 45° right-tilted; motion: sinusoidal horizontal movement) or in two dimensions (color and motion: red-horizontal movement; color and orientation: red-45° right-tilted; orientation and motion: 45° right-tilted-horizontal movement). Color, orientation, and motion targets were presented in three separate sessions dedicated to the three possible dual combinations of target dimensions: (Condition I) targets defined by color (T1), orientation (T2), and by color and orientation T1&T2; (Condition II) targets defined by color, motion, and by color & motion; and (Condition III) targets defined by orientation, motion, and by orientation & motion. Within blocks of trials, single targets T1, T2, and redundant targets T1&T2 were presented with equal probabilities, in order to permit optimal testing for violations of Miller’s (1982) RMI. All observers completed all three conditions, and care was taken that the order of conditions was counterbalanced as best possible across observers in order to control for potential sequence effects. In each of the sessions, targets were presented in a randomized order within blocks, with blocks containing 50% target-present and 50% target-absent trials. Each block consisted of 72 trials (36 target-present, 36 target-absent trials) and started with three unrecorded “warm-up” trials. Of the target-present trials, 12 contained targets defined in one of the two dimensions, 12 in the respective other dimension, and 12 targets were defined by a redundant combination of the two dimensions. Each of the three sessions comprised 10 blocks with a session total of 720 trials (consequently, the total number of trials in Experiment 1 was 2160). Experimental sessions were completed on three successive days. Observers were instructed to respond as rapidly and as accurately as possible to the presence (vs. absence) of a target within the trial display. A target-present response was to be given whenever there was an object in the display that differed from the other objects, irrespective of its definition. All observers were familiarized with the task and the experimental procedure by performing at least two blocks of 36 trials prior to the experiment. Observers were free to take short breaks between blocks.

Results and Discussion Trials with RTs faster than 200 ms and slower than 2,000 ms were excluded from analysis. Next, statistical moments were calculated separately for each observer and condition in order to exclude RTs more than ⫾3 standard deviations from the mean from analysis. Further, “twins” of false-alarm RTs were removed from the correct-response RT distributions, correcting for fast correct guesses. (Overall, less than 2% of trials were removed.) Error rates. Error rates were low overall (2.40%), with 3.06% target misses and 1.75% false alarms. Miss and false-alarm rates were analyzed in a repeated-measure ANOVA with the terms error type (miss, false alarm) and condition (CO, CM, OM). Misses were significantly more frequent overall than false alarms, F(1, 9) ⫽ 45.204, MSe ⫽ .568, p ⬍ .001, ␩p2 ⫽ .834. Importantly,

error rates did not differ between conditions (CO: 2.33%; CM: 2.40%; OM 2.47: %; F(2, 18) ⬍ 1, ns). The Error Type ⫻ Condition interaction was not significant, F(2, 18) ⫽ 1.738, p ⬎ .20. Miss rates were further analyzed in a separate ANOVA with the terms condition (CO, CM, OM) and target type (single, redundant); for this analysis, error rates for the two singly defined targets within each condition were pooled, the comparison of interest being single versus redundant target type. Although miss rates did not differ between conditions (CO: 2.33%; CM: 2.62%; OM: 2.75%; F(2, 18) ⬍ 1, ns), there was a significant main effect of target type (F(1, 9) ⫽ 171.005, MSe ⫽ .748, p ⬍ .001, ␩p2 ⫽ .950): Miss rates were lower for redundant targets than for singly defined targets, in all conditions (1.11% vs. 4.03%). The Condition ⫻ Target Type interaction was not significant, F(2, 18) ⬍ 1, ns. Reaction times. Mean target-present and -absent RTs were subjected to a repeated-measures ANOVA with the terms condition (CO, CM, OM) and trial (present, absent). Although overall RTs differed little among conditions (CO: 434.4 ms; CM: 440.8 ms; OM: 428.6 ms; F(2, 18) ⬍ 1, ns), target-absent RTs (447.5, 451.6, and 424.0 ms, respectively for the CO, CM, and OM conditions) were marginally slower than -present RTs (441.0 vs. 428.1 ms; F(1, 9) ⫽ 4.013, MSe ⫽ 623.141, p ⫽ .076, ␩p2 ⫽ .308). Slower target-absent than -present RTs in search for color and orientation feature singletons have been reported before by Krummenacher et al. (2002), who took the time difference to reflect a waiting period during which a potentially still developing (saliency) signal is allowed to reach the threshold for triggering a

7

target-present response (cf. Chun & Wolfe, 1996). That is, slower target-absent RTs reflect a hedging strategy employed to avoid missing targets on trials on which accumulation of the saliency evidence occurs relatively slowly. Subsequent analyses focused on target-present RTs. In a first step, RTs were analyzed in an (omnibus) ANOVA comparing the three experimental conditions (CO, CM, OM) of Experiment 1. As the interpretation of the results of this omnibus ANOVA cannot go beyond the level of effects common to all conditions (overall RTs, overall redundancy gains), in the subsequent analyses, RTs for each condition were analyzed separately in order to describe effects of specific combinations of dimensions in detail. Target-present RTs of the three conditions of Experiment 1 (see Figure 2) were examined in a repeated-measures ANOVA with the terms condition (CO, CM, OM) and target type (single 1, single 2, redundant; single 1 refers to singly defined color targets in the CO and CM conditions and singly defined orientation targets in the OM condition; single 2 refers to singly defined motion targets in the CM and OM conditions and singly defined orientation target in the CO condition). Mean target-present RTs did not differ significantly among the three experimental conditions (CO ⫽ 421.3, CM ⫽ 430.0, and OM ⫽ 433.1 ms; F(2, 18) ⬍ 1, ns). The main effect of target type was significant, F(2, 18) ⫽ 61.544, MSe ⫽ 316.279, p ⬍ .001, ␩p2 ⫽ .872: Redundant targets were responded to faster than single-feature targets (planned contrasts, single 1 vs. redundant: F(1, 9) ⫽ 50.544, MSe ⫽ 123.856, p ⬍ .001, ␩p2 ⫽ .849; single 2 vs. redundant: F(1, 9) ⫽ 121.740, MSe ⫽ 212.740,

450 40

400

30 350

20

300 250

10 CO

CM Experiment 1

Redundancy gain [ms]

Mean RT [ms]

500

0

OM

C O M S R G

600 550 500

40 30

450

20

400

10

350

CMa

Redundancy gain [ms]

650 Mean RT [ms]



0

Experiment 2

Figure 2. Mean RT and standard errors of the mean of the three search conditions (CO: color and orientation, CM: color and motion, and OM: orientation and motion) of Experiment 1 (top panel) and the luminance-adjusted color and motion condition (CMa) of Experiment 2 (bottom panel). Light gray bars (C) represent RTs to single color targets, dark gray bars (O) represent RTs to single orientation targets, gray bars (M) RTs to single motion targets, off-white bars (S) represent average single target RTs calculated according to the Miller and Lopes (1988; see Experiment 1, Results), black bars (R) represent RTs to redundant targets of the respective condition, and dashed bars (G; right-hand scale) represent redundancy gains calculated as the RT difference between average single (S) and redundant target (R) RTs of the respective conditions.


8


p ⬍ .001, ␩p2 ⫽ .931), in all conditions (Condition ⫻ Target Type interaction: F(4, 36) ⫽ 1.236, p ⬎ .30). The effects on RTs of particular dimensional combinations were examined in detail in follow-on ANOVAs comparing single with redundant targets for the separate conditions (CO, CM, OM). These ANOVAs revealed mean RT redundancy gains in all three experimental conditions. Mean RT redundancy gains and violations of the RMI. One of the main aims of Experiment 1 was to examine whether targets redundantly defined by the possible combinations of color and, respectively, orientation with motion (CM and OM) would yield RT redundancy gains and violations of Miller’s (1982) RMI. To quantify mean RT redundancy gains and to compare the gain pattern in the CM and OM conditions with that in the CO condition, RTs of the faster of the two single targets were compared, separately for each condition, to the respective redundant target condition (this is a standard, conservative way of examining for mean RT redundancy gains; see, e.g., Zehetleitner et al., 2011, for an overview of different types of comparison). The faster of the two single dimensions was identified following a procedure proposed by Miller and Lopes (1988). For each of the three conditions, and individually for each observer, RTs to the relevant single targets were compared using a paired-samples t test with a relatively liberal significance level of .10. It is interesting to identify the (single) dimension to which observers responded comparatively faster in the different combinations of dimensions: In the CO condition, eight out of 10 observers responded faster to color than to orientation targets; the two remaining participants showed no difference. Similarly, in the CM condition, seven observers showed faster RTs to color than to motion targets, with the other three responding comparably fast. In the OM condition, by contrast, there was little systematic preference for a particular type of target: Four observers responded faster to motion targets, one faster to orientation targets, and five equally fast to orientation and motion targets. Thus, color as a target dimension tended to be generally processed faster than the respective other dimension— consistent with a special role for color in the guidance of search (e.g., Rock, Linnett, Grant & Mack, 1992; Theeuwes, 1992). Subsequently, RTs to redundant targets were compared with the faster of the two single-target RTs whenever the t test comparing the two relevant dimensions showed a significant difference; the average of the mean RTs was used instead if the RTs did not differ reliably between the two single dimensions. The resulting redundancy gains were statistically examined in an ANOVA with the terms condition (CO, CM, OM) and target type (single, redundant). Although the main effect of condition was nonsignificant (CO ⫽ 408.4 ms, CM ⫽ 415.8 ms, and OM ⫽ 420.7 ms; F(2, 18)⬍1, ns), redundant targets were responded to faster than (the faster of the two) single targets (402.8 vs. 427.2 ms), with near equivalent mean RT redundancy gains in the three conditions (24.3, 23.7, and 25.1 ms for the CO, CM, and OM conditions, respectively; Target ⫻ Condition interaction: F(2, 18) ⬍ 1, ns). Thus, combinations of color and motion and orientation and motion targets yield mean RT redundancy gains of a magnitude comparable with those with color and orientation combinations. However, as such, the finding of mean RT redundancy gains does not tell how the gains are generated in the various conditions: whether as a result of serial checking of dimensions for

the presence of a feature contrast signal, parallel race between signals in separate dimensions, or coactivation of an overallsaliency unit by signals coded in the respective dimensions. Examining for violations of Miller’s (1982) RMI permits inferences to be drawn about the underlying processing architecture: significant violations would provide evidence in favor of a (parallel) coactivation model, as opposed to a parallel-race or serial-checking model. On saliency summation models (such as the GS and DW accounts), one would expect “coactivation” gains (i.e., RMI violations) for all combinations of dimensions. By contrast, on Zhaoping Li’s V1 hypothesis (Koene & Zhaoping, 2007; Li, 2002), such gains should be evident at most for color-orientation (and orientation-motion; see Footnote 4 above) targets (as such targets would be coded by V1 cells tuned to a conjunction of color and orientation features), but definitely not for color-motion targets (as there are no V1 cells to a conjunction of color with motion). To distinguish between parallel-race and (parallel) coactive processing in the CO, CM, and, respectively, OM conditions of Experiment 1, the fast parts (quantiles) of the RT distributions for redundantly and singly defined targets were tested for violations of Miller’s (1982) RMI. A Vincentizing procedure (Vincent, 1912) was used to discretize RT distributions into 5% quantiles (separately for each condition), and, following Kiesel et al. (2007), testing for RMI violations was restricted to the first five quantiles of the Vincentized cumulative distribution functions.6 The results of the RMI tests are presented in Table 1 (see also Figure 2, top row). Numerically, the summed probabilities of responding to nonredundant targets are smaller than the criterion value, indicative of a parallel race of signals for all tested quantiles in the CO and CM conditions [see p(c&o) and p(c&m)] and in four out of five quantiles in the OM condition [see p(o&m)]. That is, numerically, the RMI was violated in all but one of the tested quantiles of the three redundant-signals combinations of Experiment 1, two of which involve targets defined on the motion dimension. Statistical significance of the violations was assessed using paired-samples t tests. The results show that in the CO condition, the difference between the predicted (parallel race) and observed values was statistically (highly) significant for quantiles 1 to 4. The finding replicates earlier reports of violations of the RMI in CO singleton feature search conditions (Krummenacher et al., 2001, 2002) and provides further support for parallel coactive processing of color and orientation signals. More importantly, significant violations of the RMI were also revealed for the 5% and 10% quantiles of the CM and the 5% quantile of the OM condition, with a marginally significant violation of the 10% quantile of the OM condition (see Table 1 and Figure 2, top row). This finding clearly shows that dimensionally redundantly defined targets are 6 Prior to conducting the tests, “twins” of false-alarm RTs were removed from the correct-response RT distributions, effectively correcting for fast correct guesses (Eriksen, 1988; Grice et al., 1977). An iterative “kill-thetwin” procedure was used that eliminated potential fast guesses within ⫾3 ms of the incorrect response RT. The procedure searched first for the exact numerical equivalent of the incorrect response RT; if none was found, the procedure searched for an RT within the range ⫾1 ms of the incorrect response RT, and so forth up to the maximum range of ⫾3 ms. The twin was not killed when it was outside this range (there were hardly any cases where a twin was not found within the maximum range).


9

Table 1 Summary of the Results of Tests for Violations of the RMI in the Three Conditions: Color and Orientation (CO), Color and Motion (CM), and Orientation and Motion (OM) of Experiment 1 and the Luminance-Adjusted Color and Motion (CMa) Condition of Experiment 2


Experiment 1 Q

CO p(c&o)

t(9)

p

CM p(c&m)

t(9)

p

OM p(o&m)

t(9)

p

5 10 15 20 25

3.35 7.40 11.56 16.75 22.70

2.947 3.455 5.929 2.397 0.947

⬍.001 ⬍.001 ⬍.001 ⬍.05 ⬎.35

3.23 7.52 13.60 17.37 22.90

2.406 1.891 0.746 1.348 1.147

⬍.05 ⬍.05 ⬎.20 ⬎.10 ⬎.10

3.06 8.36 13.60 18.66 25.12

3.177 1.386 1.110 0.996 ⫺0.062

⬍.001 ⬎.05 ⬍ .10 ⬎.10 ⬍ .15 ⬎.15 ⬍ .20 ⬎.45

Experiment 2 Q

CMa p(c&m)

t(7)

p

5 10 15 20 25

3.19 5.92 8.41 14.59 17.56

2.314 3.927 6.477 2.061 2.903

⬍.05 ⬍.01 ⬍.001 ⬍.05 ⬍.05

Note. The leftmost column presents the redundant target cumulative RT probability quantiles (Q). Column 2 gives the sum of the probabilities that a response was given at the time corresponding to the respective quantile of the redundant target distribution for the CO condition (p(c&o)). Columns 3 and 4 give the paired-samples T- and p-values from inference testing, respectively. Columns 2 to 4 of the CO condition are repeated for the CM (p(c&m)) and OM, and (p(o&m)) conditions.

processed in a parallel coactive fashion in all experimental conditions, including those in which motion is one of the target-defining dimensions.

Experiment 2 One potential problem with the motion (CM, OM) conditions of Experiment 1 may be that the observed violations of the RMI were owing to the fact that the moving (target) item’s luminance may have been subjectively reduced relative to the stationary distractors (as a result of the target gradually changing position on its movement path), even though the items were physically isoluminant under static measurement conditions. In other words, in addition to color or, respectively, orientation and motion contrast of the target relative to the distractors, luminance contrast might have provided another source of information as to target presence, and this source might have been critical for the violations of RMI in the CM and OM conditions. (That is, the RMI violations observed in Experiment 1 may be attributable to color and luminance and, respectively, orientation and luminance producing coactivation effects, rather than color and motion and, respectively orientation and motion.) Experiment 2 was designed to avoid any such problems by having observers adjust the luminance of the moving target item to that of the stationary items in a preexperimental part of the experiment. Note that only the—theoretically critical (see above)— color-motion condition was examined in Experiment 2.

Method Participants. Eight new observers (six female; age range 20 to 25 years, M ⫽ 22.3 years; all with normal or corrected-to-

normal vision, including color vision, and all naïve with regard to the purpose of the study) completed Experiment 2. Procedure and stimuli. The procedure and stimuli were essentially the same as in (the CM condition of) Experiment 1, with the following exceptions. Stimuli were presented on a 19” CRT monitor (Philips Brilliance P202) with a screen resolution of 1,280 ⫻ 1,024 pixels run at a screen refresh rate of 100 Hz. Before the search experiment proper, participants completed three blocks of 20 trials each in which the perceived luminance of each possible target items (red vertical, green moving, red vertical moving bar) was adjusted to the luminance of the surrounding nonmoving (green vertical) distractor items. Using Watson and Pelli’s (1993) QUEST procedure, the luminance of a particular target item was adjusted by calculating, after each trial, a Bayesian estimate of the participant’s psychometric function and equiluminance threshold of the RGB values used for color target and distractor items. In more detail, in each of the three adjustment blocks, participants were presented, for 200 ms, with a 3 ⫻ 3 array of items, with the target always appearing in the center surrounded by eight distractors (green vertical bars). Display items subtended 1° of visual angle in height and .25° in width, they were separated by an average distance of 1.75° of visual angle, with the position of individual items jittered by minimum distance of 0° and a maximum distance of ⫾0.3° of visual angle horizontally and vertically relative to the center of the cell of a virtual grid underlying the array. Luminance was adjusted by decreasing or increasing the saturation (i.e., the relevant RGB value) of the deviant central item relative to the surrounding items. Luminance (saturation) of distractor items was fixed at RGB values of 0, 135, 0, that is, a setting well below saturation. The initial saturation of the deviant (red) item was set to the (maximum) value of 255, well above those of the distractors; participants were informed about this fact (so as to

10


provide additional information as to which button they had to press to indicate that the deviant item’s luminance was higher compared with that of the surrounding items). Luminance was decreased or increased according the psychometric function estimated on the basis of the participant’s judgments. Individual luminance values were saved for later use in the search experiment.7 The search experiment proper comprised of a total of 720 trials, 360 target-present and 360 target-absent trials, presented in 10 blocks of 72 trials each.


Results and Discussion Overall, the results of Experiment 2 replicate the findings of Experiment 1. Error rates. Error rates were low overall, with 3.6% misses and 1.4% false alarms; the miss rate was significantly higher than the false-alarm rate (paired-samples t test: t(7) ⫽ 3.118, two-tailed p ⫽ .017). A repeated-measures ANOVA of the miss rates with the single factor target type (color, motion, color and motion) revealed the main effect to be significant, F(2, 14) ⫽ 6.229, MSe ⫽ 2.351, p ⫽ .012: miss rates were significantly lower for color and motion trials (0.1%) than for color (0.7%; F(1, 7) ⫽ 11.160, MSe ⫽ .303, ␩p2 ⫽ .615, p ⫽ .012) and motion trials (2.8%; F(1, 7) ⫽ 7.429, Mse ⫽ 7.280, ␩p2 ⫽ .515, p ⫽ .030). Mean RTs and redundancy gains. Mean RTs (see Figure 2, bottom panel) in the individually luminance-adjusted color and motion (CMa) search condition of Experiment 2 were, on average, 523.2 ms, 588.2 ms, and 498.9 ms for color, motion, and color and motion targets, respectively. Target-absent and -present RTs did not differ statistically, 597.7 vs. 536.8 ms; t(7) ⫽ 1.092, two-tailed p ⫽ .311. RTs in the redundant-target condition were faster, by 24.3 ms on average, than in the faster of the two nonredundant conditions; note that targets defined by color were responded to faster than orientation-defined targets by all observers in Experiment 2. A repeated-measures ANOVA with the factor target definition (color, motion, color and motion) yielded a significant main effect (F(2, 14) ⫽ 83.727, MSe ⫽ 203.927, ␩p2 ⫽ .923, p ⫽ .000); planned simple contrasts revealed RTs to redundant (color and motion) targets to be significantly faster than RTs to, respectively, color targets (F(1, 7) ⫽ 24.207, Mse ⫽ 194.349, ␩p2 ⫽ .776, p ⫽ .002) and motion targets (F(1, 7) ⫽ 108.336, Mse ⫽ 589.366, ␩p2 ⫽ .939, p ⫽ .000). As in Experiment 1, mean RT redundancy gains were calculated as the difference between redundant target trial RTs (498.9 ms) and the single target RT (536.5) obained according to Miller and Lopes (1988). (Note that four out of eight participants of Experiment 2 responded faster to the color than the motion target, in the other four, participants RTs to single target did not differ.) The RT benefit for redundant relative to single targets of 37.6 ms was shown to be statistically significant, t(7) ⫽ 4.893, two-tailed p ⫽ .002. Violations of the RMI. The results of the RMI tests are presented in Table 1 (see also Figure 3). As can be seen from the table, the summed probabilities of responding to nonredundant targets are numerically smaller than the probability of responding to a redundant target for all quantiles tested (Quantiles 1 to 5), and in fact tests of Miller’s (1982) RMI revealed the violations to be significant for all five quantiles. This result argues against an uncontrolled luminance contrast between the moving target and

the stationary distractors having been responsible for the RMI violations obtained in Experiment 1. Thus, the findings of Experiments 1 and 2 taken together strongly support the assumption that motion signals are processed in the same way as color and orientation signals. The GS and DW models assume that local feature differences give rise to dimension-specific feature contrast signals that “point to” or “index” locations exhibiting discontinuities relative to the surrounding context. The dimension-based signals are integrated, in a spatially specific manner, into an overall-saliency representation of the display guiding the allocation of focal attention. Given that the dimensional combinations of motion with color (and motion with orientation) targets produced both mean RT redundancy gains and violations of the RMI, the most parsimonious explanation of the present result is that motion signals (here representing motion of the target item relative to the stationary distractor items) are integrated with color (and with orientation) signals at the overallsaliency representation stage, in the same way as color and orientation signals are integrated. Certainly, given that color and motion are coded in separate pathways (so that there are no color-motion conjunction cells), the integration must occur later than V1—at variance with Zhaoping Li’s (2002; Koene & Zhaoping, 2007) V1 hypothesis. In summary, the finding of coactivation effects in the CM condition (as well as the OM condition of Experiment 1) would argue in favor of an overall-saliency processing architecture, while of course not ruling out (or rather while expressly acknowledging) a role of V1 mechanisms in feature contrast and saliency computations.

General Discussion The present study was designed to test predictions deriving from saliency-based models of visual search, such as guided search (e.g., Wolfe, 1994, 2007) and dimension weighting (e.g., Müller, Heller & Ziegler, 1995; Müller, Reimann, & Krummenacher, 2003), against those of Zhaoping Li’s (2002; Koene & Zhaoping, 2007) V1 hypothesis. Saliency-based models assume that target detection involves the computation of dimension-based feature contrast signals that are then integrated into a supradimensional (overall-) saliency map—which, in turn, guides the allocation of focal attention to particular locations in the visual field. Importantly, in saliency-based models, the overall-saliency representation itself is assumed to be a “featureless” (it just signals that a location differs in some feature(s) from its surround) and relatively late representation (e.g., in the LIP and/or FEF; Gottlieb, Kusunoki, & Goldberg, 1998; Goldberg, Bisley, Powell, Gottlieb, & Kusunoki, 2002; Thompson & Bichot, 2005), rather than corresponding to the activation of feature-coding neurons (feature detectors) in early visual areas. The present study used variants of the redundant-signals paradigm, which permits the assumptions of both classes of model to be tested against each other. Expedited mean RTs (mean RT redundancy gains) to dimensionally redun7 The average RGB values (and associated standard deviations) were, respectively, for the red nonmoving target 148.5 (19.3), the green moving target 120.5 (12.8), and the red moving target 157.5 (33.5). That is, the adjusted values were slightly above the RGB (i.e., the saturation/luminance) value of the green nonmoving distractors for red targets, irrespective of whether they were stationary or moving, and slightly below for green moving targets.

Cumulative probability

REDUNDANCY GAINS IN COLOR-MOTION SINGLETON SEARCH 1

1

1

0.8

0.8

0.8

0.6

0.6

0.6

0.4

0.4

0.4

0.2

0.2

0.2

0 200

450 CO

0 200

700

450 700 CM RT distribution

0 200

450 OM

11

700

Cumulative probability


1 0.8 0.6 0.4 0.2 0 200

500 CMa

800

RT distribution

Figure 3. Cumulative probability density functions (CDFs) of the color and orientation (CO), color and motion (CM), and motion and orientation (OM) conditions of Experiments 1 (top panel) and the luminance-adjusted color and motion (CMa) condition of Experiment 2 (bottom panel). CDFs were obtained by fitting normal distributions to the overall (grand) mean of individual mean RTs and standard deviations of the Vincentized (see Experiment 1, Results) RT distributions of the two (nonredundant) single-dimension and the redundant target trail conditions. Dashed lines represent color targets, dotted lines orientation, and dashed and dotted lines motion targets and solid lines redundant targets of the respective condition.

dantly, compared with singly, defined targets would be consistent with a parallel race of redundant-target signals to trigger a detection response. Beyond this, violations of Miller’s (1982) race model inequality (RMI) in analyses of the entire redundant- and single-target RT distributions would be indicative of redundant signals coactivating, in parallel, an integrative mechanism that mediates response selection and execution. Both mean RT redundancy gains and violations of the RMI were previously demonstrated in search for singleton feature targets redundantly defined by color and orientation (Krummenacher et al., 2001, 2002). Zhaoping Li’s (2002; Koene & Zhaoping, 2007) V1 hypothesis challenges a core assumption of saliency-based models: that saliency representations are independent of feature representations. Instead, on the V1 hypothesis, saliency corresponds to the (level of) activation of neurons signaling the presence of specific features and feature combinations in V1. Experiment 1 examined one specific prediction of the V1 hypothesis, namely, that there would not be any redundancy gains involving motion signals (see Koene & Zhaoping, 2007), because the coding of motion occurs in a separable processing stream to that of color and orientation (although, for the reasons elaborated earlier in the article and as acknowledged in Footnote 4, the critical test involves combinations of motion with color, rather than with orientation). Experiment 1 revealed significant mean RT redundancy gains and violations of the RMI for combinations for redundant color and orientation signals, replicating earlier findings by Krummenacher et al. (2001, 2002). More importantly, Experiments 1 and 2 dem-

onstrated these effects also for targets defined by combinations of motion with color (and motion with orientation). These findings are at variance with the V1 hypothesis, but they are in line with saliency-based models, which assume that the locus of the saliency representation is independent of and subsequent to the representation of particular features or feature combinations. In summary then, the present results would argue in favor of a relatively late (later than V1), but preattentive representation of visual saliency (at variance with the V1 hypothesis).

Implication of the Present Results for the V1 Hypothesis The main aim of the present study was to reexamine the V1 hypothesis, in particular, to test whether combined-feature coding (by conjunction) cells at V1 level would be able to explain the full pattern of redundant-signals effects, that is, not only mean RT redundancy gains but also violations of the RMI at the “very fast” end of the RT distribution. The answer is no: targets defined, in particular, by both motion and color do produce similar RMI violations to those seen with color-orientation targets, while (according to neurophysiological research) there is no neural “machinery” at the V1 level that could give rise to these effects. This would argue that (at least) the results observed in the color-motion conditions of the present experiments cannot have arisen at the level of V1—instead, they must be attributable to a higher processing stage.


12


What about search for color-orientation targets? For these, given that there are neurons tuned to a combination of color and orientation, a V1 account would appear feasible in principle. However, even here there are doubts that the V1 hypothesis can account fully for the observed RT redundancy gains. As argued in the beginning of the article, what Koene and Zhaoping (2007) found were not violations of the RMI at the very fast end of the RT distributions (in fact, they did not formally test for RMI violations), but rather faster redundant-target RTs than predicted from the (fastest) single-target RTs only for “slower” time bins of the RT distributions. This could indeed suggest that conjunction cells contributed to the RT gains in this (somewhat later) time range. Given that such color-orientation conjunction cells (AND cells; see Footnote 2) are coactivated by signals from both color-only and orientationonly cells (as considered by Koene & Zhaoping, 2007, as an alternative to the V1 hypothesis; compare their Figure 4 with their Figure 1b), it follows that the output of such cells could contribute to performance only later, after the activation of the color-only and orientation-only cells (which themselves may not be activated without a relative time delay). If conjunction cells are (co-) activated in this way, then it follows that the output of such cells could not be contributing to RMI violations at the very fastest end of the RT distributions (which would be solely driven by the outputs of the color-only or orientation-only cells, which become available earlier). Given this, we argue that even with redundantly defined color-and-orientation targets, the RMI violations revealed in the present study point to another integration stage outside of V1, which starts to accumulate information from dimension-specific (feature contrast) coding mechanisms as soon as this information becomes available. We suggest that this is the same stage as with motion-color and motion-orientation (redundant) targets. Plausibly, this stage is located in brain areas that make up the oculomotor network (Fecteau & Munoz, 2006) and include the superior colliculus (McPeek & Keller, 2002), the pulvinar thalamus (Robinson & Petersen, 1992), the frontal eye fields (FEF; Sato & Schall, 2003), and the lateral intraparietal sulcus (LIP; Bisley & Goldberg, 2003)—where neurons are known to signal saliency in a largely featureless fashion.8 There are other arguments, based on the results of previous studies (investigating the positional “priming of pop-out” effect; see, e.g., Geyer, Müller, & Krummenacher, 2007; Kumada, 2001; Maljkovic & Nakayama, 1996), against a strong version of the V1 hypothesis. This hypothesis entails that saliency signaling is strictly limited to a particular location and determined by a particular combination of target and distractor features. Accordingly, if examining for effects of intertrial history, identical targetdistractor feature combinations presented at the same target and distractor locations should yield benefits due to residual neural activation: increased starting activity of the cell coding the targetdefining feature at the previous target location and reduced activity of cells coding nondefining features (shared with distractors) at the previous target and distractor locations. In contrast, if the same target-distractor feature combination is presented at a different location, or if a target with a different feature value but defined on the same dimension is presented at the same (or a different) location as on the preceding trial, the benefits should be abolished. However, the assumption (entailed in the V1 hypothesis) of a feature-based mechanism underlying intertrial effects is called into question by, for example, the findings of Found and Müller (1996):

They showed that there is a dimension-repetition benefit, which is seen in terms of expedited search times when the target is repeatedly defined in the same dimension (whether it is defined by the same or a different feature); furthermore, the dimension repetition effect is nonlocal, that is, it is seen whether the target appears at the same or a different location (even a location in the opposite hemifield). The latter was systematically explored by Krummenacher, Müller, Zehetleitner, and Geyer (2009) who found that, with a simple target detection task, the dimension repetition benefit was additive to a weak effect of “location” repetition (essentially, detection times increased slightly with increasing distance of the current from the previous target, within a distance range of up to 16° of visual angle). It is difficult to see how the V1 hypothesis could explain both the dimension specificity and the location independence of intertrial effects in singleton detection tasks. In this context, it is interesting to note that Krummenacher et al. (2001) found the strength of RMI violations to be modulated, too, by intertrial target history: Violations were more marked for the second compared with the first of two repeated redundant (colororientation) targets, relative to the respective (i.e., first vs. second of two repeated) singly defined targets. Krummenacher et al. took this to mean that the first or two redundant targets adjusts the dimensional weight set (the relative weights for color and orientation signals) such as to permit optimal signal integration to occur. That is, a balanced weight set would permit both (redundant) target signals to activate the integrating neuron at the same time; by contrast, with a weight set biased toward, say, color, the integrating neuron—and thus detection—would be driven mainly by the color signal of a redundant target. Finally, Krummenacher et al. (2002) found dual (i.e., two separate) singleton targets at spatially separate positions to also produce violations of the RMI, provided that these were (a) defined in separate dimensions (e.g., one defined by color and one by orientation; there were no violations, when redundant targets were defined in the same dimension, e.g., one red and one blue target); and (b) the spatial distance between them less than 5° of visual angle (reliable coactivation effects were evident for distances up to 2.5°). Although small, this integration range (somewhere between 2.5°–5°) is larger than the typical receptive fields of V1 cells (smaller than 1°) in central vision (eccentricities ⱕ2.5°)—suggesting that the integration stage is located beyond V1. Another finding of Krummenacher et al. (2002) at odds with the V1 hypothesis was that, in contrast to nearby dual targets defined on different dimensions, dual targets defined on the same dimension (e.g., a red and a green singleton) failed to produce violations of the RMI. On the V1 hypothesis, V1 cells responding to different features would not be suppressed by iso-feature suppression, so that when located in close proximity they would both more strongly signal “saliency” within the target area compared with a single signal. If the speed of visual selection (attentional deployment) depends on the strength of localized saliency signals, this 8 Some psychophysical evidence for featureless saliency coding has also been provided by Müller et al. (1995, 2004), who showed that it takes longer to explicitly encode specific features values for response as compared with just the signals permitting detection decisions, and when observers have just to detect a pop-out target, they tend to be unable to report the target-defining features above chance. For related ERP evidence, see, for example, Töllner, Rangelov, and Müller (2012).



should have translated into RMI violation. By contrast, if only the outputs of dimension-specific feature contrast computations are integrated at the overall-saliency stage (as assumed by the DW account; see Zehetleitner et al., 2008, for review), the lack of RMI violations with dual targets defined on the same dimension is not surprising. In summary, although we take our results to argue against a strong V1 hypothesis9, we do not wish to claim that V1 is not involved in saliency coding. On the contrary, as rightly pointed out by Zhaoping Li and her colleagues, V1 (and also later featurespecific visual areas) are likely to contribute to feature contrast computations (e.g., via iso-feature suppression) that provide the input to a higher-level integration stage.

9 Of course, lack of CM cells in V1 may not be definite. Although, according to Horwitz and Albright (2005), V1 has as few, if any, CM cells, there are some reports of such cells in the literature (e.g., Michael, 1978; Tamura, Sato, Katsuyama, Hata, & Tsumoto, 1996). If their numbers were significant, and if they were directly activated by dual color and motion features (rather than being secondary coactivated by already activated color and motion detectors), the present finding of RMI violations at the fast end of the RT distribution for the CM condition (which are comparable in magnitude with those in the CO and OM conditions) would not critically challenge the V1 hypothesis. The implication is that more neurophysiological research would be required to reveal to which extent and exactly how color-motion conjunctions are computed in V1. However, while not denying the possibility of V1 conjunction cells contributing to coactivation gains (see above), we take the very robust RMI violations produced by CM targets in the present study to argue against the strong claim of the V1 hypothesis that these gains are exclusively produced by conjunction coding in V1.

References Bisley, J. W., & Goldberg, M. E. (2003). Neuronal activity in the lateral intraparietal area and spatial attention. Science, 299, 81– 86. doi: 10.1126/science.1077395 Chun, M. M., & Wolfe, J. M. (1996). Just say no: How are visual search trials terminated when there is no target present? Cognitive Psychology, 30, 39 –78. doi:10.1006/cogp.1996.0002 Eriksen, C. W. (1988). A source of error in attempts to distinguish coactivation from separate activation in the perception of redundant targets. Perception & Psychophysics, 44, 191–193. doi:10.3758/BF03208712 Fecteau, J. H., & Munoz, D. P. (2006). Salience, relevance, and firing: A priority map for target selection. Trends in Cognitive Sciences, 10, 382–390. doi:10.1016/j.tics.2006.06.011 Found, A., & Müller, H. J. (1996). Searching for unknown feature targets on more than one dimension: Investigating a “dimension weighting” account. Perception and Psychophysics, 58, 88 –101. doi:10.3758/BF03205479 Geyer, T., Müller, H. J., & Krummenacher, J. (2007). Cross-trial priming of element positions in visual pop- out search is dependent on stimulus arrangement. Journal of Experimental Psychology: Human Perception and Performance, 33, 788 –797. doi:10.1037/0096-1523.33.4.788 Goldberg, M. E., Bisley, J., Powell, K. D., Gottlieb, J., & Kusunoki, M. (2002). The role of the lateral intraparietal area of the monkey in the generation of saccades and visuospatial attention. Annals of the New York Academy of Sciences, 956, 205–215. doi:10.1111/j.1749-6632 .2002.tb02820.x Gottlieb, J., Kusunoki, M., & Goldberg, M. E. (1998). The representation of visual salience in monkey parietal cortex. Nature, 391, 481– 484. doi:10.1038/35135 Grice, G. R., Nullmeyer, R., & Spiker, V. A. (1977). Application of variable criterion theory to choice reaction time. Perception & Psychophysics, 22, 431– 449. doi:10.3758/BF03199509

13

Horwitz, G. D., & Albright, T. D. (2005). Paucity of chromatic linear motion detectors in macaque V1. Journal of Vision, 5, 525–533. doi: 10.1167/5.6.4 Itti, L., & Koch, C. (2000). A saliency-based search mechanism for overt and covert shifts of visual attention. Vision Research, 40, 1489 –1506. doi:10.1016/S0042-6989(99)00163-7 Kiesel, A., Miller, J., & Ulrich, R. (2007). Systematic biases and Type I error accumulation in tests of the race model inequality. Behavior Research Methods, 39, 539 –551. doi:10.3758/BF03193024 Koch, C., & Ullman, S. (1985). Shifts in selective visual attention: Towards the underlying neural circuitry. Human Neurobiology, 4, 219 –227. Koene, A. R., & Zhaoping, L. (2007). Feature-specific interactions in salience from combined feature contrasts: Evidence for a bottom-up saliency map in V1. Journal of Vision, 7(7), 1–14. Krummenacher, J., Müller, H. J., & Heller, D. (2001). Visual search for dimensionally redundant pop-out targets: Evidence for parallel-coactive processing of dimensions. Perception & Psychophysics, 63, 901–917. doi:10.3758/BF03194446 Krummenacher, J., Müller, H. J., & Heller, D. (2002). Visual search for dimensionally redundant pop-out targets: Parallel-coactive processing of dimensions is location-specific. Journal of Experimental Psychology: Human Perception and Performance, 28, 1303–1322. doi:10.1037/ 0096-1523.28.6.1303 Krummenacher, J., Müller, H. J., Zehetleitner, M., & Geyer, T. (2009). Dimension- and space-based intertrial effects in visual pop-out search: Modulation by task demands for focal-attentional processing. Psychological Research, 73, 186 –197. doi:10.1007/s00426-008-0206-y Kumada, T. (2001). Feature-based control of attention: Evidence for two forms of dimension weighting. Perception & Psychophysics, 63, 698 – 708. doi:10.3758/BF03194430 Li, Z. (2002). A saliency map in primary visual cortex. Trends in Cognitive Sciences, 6, 9 –16. doi:10.1016/S1364-6613(00)01817-9 Livingstone, M. S., & Hubel, D. H. (1984). Anatomy and physiology of a color system in the primate visual cortex. Journal of Neuroscience, 4, 309 –356. Maljkovic, V., & Nakayama, K. (1994). The priming of pop-out I: Role of features. Memory & Cognition, 22, 657– 672. doi:10.3758/BF03209251 Maljkovic, V., & Nakayama, K. (1996). Priming of pop-out: II. The role of position. Perception & Psychophysics, 58, 977–991. doi:10.3758/ BF03206826 McPeek, R. M., & Keller, E. L. (2002). Saccade target selection in the superior colliculus during a visual search task. Journal of Neurophysiology, 88, 2019 –2034. Michael, C. R. (1978). Color vision mechanisms in monkey striate cortex: Dual-opponent cells with concentric receptive fields. Journal of Neurophysiology, 41, 572–588. Miller, J. (1982). Divided attention: Evidence for coactivation with redundant signals. Cognitive Psychology, 14, 247–279. doi:10.1016/00100285(82)90010-X Miller, J., & Lopes, A. (1988). Testing race models by estimating the smaller of two true mean or true median reaction times: An analysis of estimation bias. Perception & Psychophysics, 44, 513–524. doi:10.3758/ BF03207485 Müller, H. J., Heller, D., & Ziegler, J. (1995). Visual search for singleton feature targets within and across feature dimensions. Perception and Psychophysics, 57, 1–17. doi:10.3758/BF03211845 Müller, H. J., Krummenacher, J., & Heller, D. (2004). Dimension-specific inter-trial facilitation in visual search for pop-out targets: Evidence for a top-down modulable visual short-term memory effect. Visual Cognition, 11, 577– 602. doi:10.1080/13506280344000419 Müller, H. J., Reimann, B., & Krummenacher, J. (2003). Visual search for singleton feature targets across dimensions: Stimulus- and expectancydriven effects in dimensional weighting. Journal of Experimental Psy-


14


chology: Human Perception and Performance, 29, 1021–1035. doi: 10.1037/0096-1523.29.5.1021 Poom, L. (2009). Integration of colour, motion, orientation, and spatial frequency in visual search. Perception, 38, 708 –718. doi:10.1068/p6072 Raab, D. H. (1962). Statistical facilitation of simple reaction time. Transactions of the New York Academy of Sciences, 24, 574 –590. doi: 10.1111/j.2164-0947.1962.tb01433.x Robinson, D. L., & Petersen, S. (1992). The pulvinar and visual salience. Trends in Neurosciences, 15, 127–132. doi:10.1016/01662236(92)90354-B Rock, I., Linnett, C., Grant, P., & Mack, A. (1992). Perception without attention: Results of a new method. Cognitive Psychology, 24, 502–534. doi:10.1016/0010-0285(92)90017-V Sato, T. R., & Schall, J. D. (2003). Effects of stimulus-response compatibility on neural selection in frontal eye field. Neuron, 38, 637– 648. doi:10.1016/S0896-6273(03)00237-X Soltani, A., & Koch, C. (2010). Visual saliency computations: Mechanisms, constraints, and the effect of feedback. The Journal of Neuroscience, 30, 12831–12843. doi:10.1523/JNEUROSCI.1517-10.2010 Tamura, H., Sato, H., Katsuyama, N., Hata, Y., & Tsumoto, T. (1996). Less segregated processing of visual information in V2 than in V1 of the monkey visual cortex. European Journal of Neuroscience, 8, 300 –309. doi:10.1111/j.1460-9568.1996.tb01214.x Theeuwes, J. (1992). Perceptual selectivity for color and form. Perception and Psychophysics, 51, 599 – 606. doi:10.3758/BF03211656 Thompson, K. G., & Bichot, N. P. (2005). A visual salience map in the primate frontal eye field. Progress in Brain Research, 147, 249 –262. doi:10.1016/S0079-6123(04)47019-8 Töllner, T., Rangelov, D., & Müller, H. J. (2012). How the speed of motor-response decisions, but not focal-attentional selection, differs as a function of task set and target prevalence. Proceedings of the National Academy of Sciences of the United States of America, 109, e1990-9. doi:10.1073/pnas.1206382109 Ts’o, D. Y., & Gilbert, C. D. (1988). The organization of chromatic and spatial interactions in the primate striate cortex. Journal of Neuroscience, 8, 1712–1727.

Vincent, S. B. (1912). The function of the vibrissae in the behavior of the white rat. Behavior Monographs, 1, 7– 85. Watson, A. B., & Pelli, D. G. (1983). QUEST: A Bayesian adaptive psychometric method. Perception & Psychophysics, 33, 113–120. doi: 10.3758/BF03202828 Wolfe, J. M. (1994). Guided Search 2.0: A revised model of visual search. Psychonomic Bulletin & Review, 1, 202–238. doi:10.3758/BF03200774 Wolfe, J. M. (1998). What can 1,000,000 trials tell us about visual search? Psychological Science, 9, 33–39. doi:10.1111/1467-9280.00006 Wolfe, J. M. (2007). Guided search 4.0: Current progress with a model of visual search. In W. Gray (Ed.), Integrated models of cognitive systems (pp. 99 –119). New York, NY: Oxford. Wolfe, J. M., Butcher, S., Lee, C., & Hyle, M. (2003). Changing your mind: On the contributions of top-down and bottom-up guidance in visual search for feature singletons. Journal of Experimental Psychology: Human Perception and Performance, 29, 483–502. doi:10.1037/ 0096-1523.29.2.483 Wolfe, J. M., & Horowitz, T. S. (2004). What attributes guide the deployment of visual attention and how do they do it? Nature Reviews Neuroscience, 5 495–501. doi:10.1038/nrn1411 Zehetleitner, M., Krummenacher, J., Geyer, T., Hegenloh, M., & Müller, H. J. (2011). Intention and trial history modulate dimensional weights in localization of pop-out targets. Attention, Perception, & Psychophysics, 73, 349 –363. doi:10.3758/s13414-010-0035-0 Zehetleitner, M., Müller, H. J., & Krummenacher, J. (2008). The redundant-signals paradigm and preattentive visual processing. Frontiers in Bioscience, May, 5279 –5293. Zhaoping, L., & Snowden, R. B. (2006). A theory of a saliency map in primary visual cortex (V1) tested by psychophysics of colour-orientation interference in texture segregation. Visual Cognition, 14, 911–933. doi: 10.1080/13506280500196035

Received March 11, 2011 Revision received May 21, 2014 Accepted May 28, 2014 䡲