{"title": "Top-Down Control of Visual Attention: A Rational Account", "book": "Advances in Neural Information Processing Systems", "page_first": 923, "page_last": 930, "abstract": null, "full_text": "Top-Down Control of Visual Attention:\n\nA Rational Account\n\nMichael C. Mozer\n\nDept. of Comp. Science &\nInstitute of Cog. Science\nUniversity of Colorado\nBoulder, CO 80309 USA\n\nMichael Shettel\n\nDept. of Comp. Science & \nInstitute of Cog. Science\nUniversity of Colorado\nBoulder, CO 80309 USA\n\nShaun Vecera\n\nDept. of Psychology\nUniversity of Iowa\n\nIowa City, IA 52242 USA\n\nAbstract\n\nTheories of visual attention commonly posit that early parallel processes extract con-\nspicuous features such as color contrast and motion from the visual field. These features\nare then combined into a saliency map, and attention is directed to the most salient\nregions first. Top-down attentional control is achieved by modulating the contribution of\ndifferent feature types to the saliency map. A key source of data concerning attentional\ncontrol comes from behavioral studies in which the effect of recent experience is exam-\nined as individuals repeatedly perform a perceptual discrimination task (e.g., \u201cwhat\nshape is the odd-colored object?\u201d). The robust finding is that repetition of features of\nrecent trials (e.g., target color) facilitates performance. We view this facilitation as an\nadaptation to the statistical structure of the environment. We propose a probabilistic\nmodel of the environment that is updated after each trial. Under the assumption that\nattentional control operates so as to make performance more efficient for more likely\nenvironmental states, we obtain parsimonious explanations for data from four different\nexperiments. Further, our model provides a rational explanation for why the influence of\npast experience on attentional control is short lived.\n\n1 INTRODUCTION\nThe brain does not have the computational capacity to fully process the massive quantity\nof information provided by the eyes. Selective attention operates to filter the spatiotempo-\nral stream to a manageable quantity. Key to understanding the nature of attention is dis-\ncovering the algorithm governing selection, i.e., understanding what information will be\nselected and what will be suppressed. Selection is influenced by attributes of the spa-\ntiotemporal stream, often referred to as bottom-up contributions to attention. For example,\nattention is drawn to abrupt onsets, motion, and regions of high contrast in brightness and\ncolor. Most theories of attention posit that some visual information processing is per-\nformed preattentively and in parallel across the visual field. This processing extracts prim-\nitive visual features such as color and motion, which provide the bottom-up cues for\nattentional guidance. However, attention is not driven willy nilly by these cues. The\ndeployment of attention can be modulated by task instructions, current goals, and domain\nknowledge, collectively referred to as top-down contributions to attention. \nHow do bottom-up and top-down contributions to attention interact? Most psychologi-\ncally and neurobiologically motivated models propose a very similar architecture in which\ninformation from bottom-up and top-down sources combines in a saliency (or activation)\nmap (e.g., Itti et al., 1998; Koch & Ullman, 1985; Mozer, 1991; Wolfe, 1994). The\nsaliency map indicates, for each location in the visual field, the relative importance of that\nlocation. Attention is drawn to the most salient locations first. \nFigure 1 sketches the basic architecture that incorporates bottom-up and top-down contri-\nbutions to the saliency map. The visual image is analyzed to extract maps of primitive fea-\ntures such as color and orientation. Associated with each location in a map is a scalar\n\n\fvisual image\n\nhorizontal\n\nvertical\n\nprimitive feature maps\n\ngreen\n\ntop-down gains\n\nred\n\nsaliency\n\nmap\n\nFIGURE 1. An attentional saliency\nmap constructed from bottom-up\nand top-down information\n\nbottom-up activations\n\nFIGURE 2. Sample display from\nExperiment 1 of Maljkovic and\nNakayama (1994)\n\nresponse or activation indicating the presence of a particular feature. Most models assume\nthat responses are stronger at locations with high local feature contrast, consistent with\nneurophysiological data, e.g., the response of a red feature detector to a red object is stron-\nger if the object is surrounded by green objects. The saliency map is obtained by taking a\nsum of bottom-up activations from the feature maps. The bottom-up activations are modu-\nlated by a top-down gain that specifies the contribution of a particular map to saliency in\nthe current task and environment. Wolfe (1994) describes a heuristic algorithm for deter-\nmining appropriate gains in a visual search task, where the goal is to detect a target object\namong distractor objects. Wolfe proposes that maps encoding features that discriminate\nbetween target and distractors have higher gains, and to be consistent with the data, he\nproposes limits on the magnitude of gain modulation and the number of gains that can be\nmodulated. More recently, Wolfe et al. (2003) have been explicit in proposing optimiza-\ntion as a principle for setting gains given the task definition and stimulus environment. \nOne aspect of optimizing attentional control involves configuring the attentional system to\nperform a given task; for example, in a visual search task for a red vertical target among\ngreen vertical and red horizontal distractors, the task definition should result in a higher\ngain for red and vertical feature maps than for other feature maps. However, there is a\nmore subtle form of gain modulation, which depends on the statistics of display environ-\nments. For example, if green vertical distractors predominate, then red is a better discrim-\ninative cue than vertical; and if red horizontal distractors predominate, then vertical is a\nbetter discriminative cue than red.\nIn this paper, we propose a model that encodes statistics of the environment in order to\nallow for optimization of attentional control to the structure of the environment. Our\nmodel is designed to address a key set of behavioral data, which we describe next.\n\n1.1 Attentional priming phenomena\nPsychological studies involve a sequence of experimental trials that begin with a stimulus\npresentation and end with a response from the human participant. Typically, trial order is\nrandomized, and the context preceding a trial is ignored. However, in sequential studies,\nperformance is examined on one trial contingent on the past history of trials. These\nsequential studies explore how experience influences future performance. Consider a the\nsequential attentional task of Maljkovic and Nakayama (1994). On each trial, the stimulus\ndisplay (Figure 2) consists of three notched diamonds, one a singleton in color\u2014either\ngreen among red or red among green. The task is to report whether the singleton diamond,\nreferred to as the target, is notched on the left or the right. The task is easy because the sin-\ngleton pops out, i.e., the time to locate the singleton does not depend on the number of dia-\nmonds in the display. Nonetheless, the response time significantly depends on the\nsequence of trials leading up to the current trial: If the target is the same color on the cur-\n\n\frent trial as on the previous trial, response time is roughly 100 ms faster than if the target is\na different color on the current trial. Considering that response times are on the order of\n700 ms, this effect, which we term attentional priming, is gigantic in the scheme of psy-\nchological phenomena. \n\n2 ATTENTIONAL CONTROL AS ADAPTATION TO THE \nSTATISTICS OF THE ENVIRONMENT\nWe interpret the phenomenon of attentional priming via a particular perspective on atten-\ntional control, which can be summarized in two bullets. \n\n\u2022 The perceptual system dynamically constructs a probabilistic model of the environ-\n\nment based on its past experience. \n\n\u2022 Control parameters of the attentional system are tuned so as to optimize performance\n\nunder the current environmental model. \n\nThe primary focus of this paper is the environmental model, but we first discuss the nature\nof performance optimization.\nThe role of attention is to make processing of some stimuli more efficient, and conse-\nquently, the processing of other stimuli less efficient. For example, if the gain on the red\nfeature map is turned up, processing will be efficient for red items, but competition from\nred items will reduce the efficiency for green items. Thus, optimal control should tune the\nsystem for the most likely states of the world by minimizing an objective function such as:\n\nJ g( )\n\n\u2211=\ne\n\nP e( )RTg e( )\n\n(1)\n\nwhere g is a vector of top-down gains, e is an index over environmental states, P(.) is the\nprobability of an environmental state, and RTg(.) is the expected response time\u2014assuming\na constant error rate\u2014to the environmental state under gains g. Determining the optimal\ngains is a challenge because every gain setting will result in facilitation of responses to\nsome environmental states but hindrance of responses to other states.\nThe optimal control problem could be solved via direct reinforcement learning, but the\nrapidity of human learning makes this possibility unlikely: In a variety of experimental\ntasks, evidence suggests that adaptation to a new task or environment can occur in just one\nor two trials (e.g., Rogers & Monsell, 1996). Model-based reinforcement learning is an\nattractive alternative, because given a model, optimization can occur without further expe-\nrience in the real world. Although the number of real-world trials necessary to achieve a\ngiven level of performance is comparable for direct and model-based reinforcement learn-\ning in stationary environments (Kearns & Singh, 1999), naturalistic environments can be\nviewed as highly nonstationary. In such a situation, the framework we suggest is well\nmotivated: After each experience, the environment model is updated. The updated envi-\nronmental model is then used to retune the attentional system.\nIn this paper, we propose a particular model of the environment suitable for visual search\ntasks. Rather than explicitly modeling the optimization of attentional control by setting\ngains, we assume that the optimization process will serve to minimize Equation 1.\nBecause any gain adjustment will facilitate performance in some environmental states and\nhinder performance in others, an optimized control system should obtain faster reaction\ntimes for more probable environmental states. This assumption allows us to explain exper-\nimental results in a minimal, parsimonious framework.\n\n3 MODELING THE ENVIRONMENT\nFocusing on the domain of visual search, we characterize the environment in terms of a\n\n\fprobability distribution over configurations of target and distractor features. We distin-\nguish three classes of features: defining, reported, and irrelevant. To explain these terms,\nconsider the task of searching a display of size varying, colored, notched diamonds (Fig-\nure 2), with the task of detecting the singleton in color and judging the notch location.\nColor is the defining feature, notch location is the reported feature, and size is an irrele-\nvant feature. To simplify the exposition, we treat all features as having discrete values, an\nassumption which is true of the experimental tasks we model. We begin by considering\ndisplays containing a single target and a single distractor, and shortly generalize to multid-\nistractor displays.\nWe use the framework of Bayesian networks to characterize the environment. Each fea-\nture of the target and distractor is a discrete random variable, e.g., Tcolor for target color\nand Dnotch for the location of the notch on the distractor. The Bayes net encodes the prob-\nability distribution over environmental states; in our working example, this distribution is \n\nP(Tcolor, Tsize, Tnotch, Dcolor, Dsize, Dnotch).\n\nThe structure of the Bayes net specifies the relationships among the features. The simplest\nmodel one could consider would be to treat the features as independent, illustrated in Fig-\nure 3a for singleton-color search task. The opposite extreme would be the full joint distri-\nbution, which could be represented by a look up table indexed by the six features, or by\nthe cascading Bayes net architecture in Figure 3b. The architecture we propose, which\nwe\u2019ll refer to as the dominance model (Figure 3c), has an intermediate dependency struc-\nture, and expresses the joint distribution as:\nP(Tcolor)P(Dcolor|Tcolor)P(Tsize|Tcolor)P(Tnotch|Tcolor)P(Dsize|Dcolor)P(Dnotch|Tcolor).\nThe structured model is constructed based on three rules.\n\n1. The defining feature of the target is at the root of the tree.\n2. The defining feature of the distractor is conditionally dependent on the defining fea-\nture of the target. We refer to this rule as dominance of the target over the distractor.\n3. The reported and irrelevant features of target (distractor) are conditionally dependent\non the defining feature of the target (distractor). We refer to this rule as dominance of\nthe defining feature over nondefining features.\n\nAs we will demonstrate, the dominance model produces a parsimonious account of a wide\nrange of experimental data.\n\n3.1 Updating the environment model\nThe model\u2019s parameters are the conditional distributions embodied in the links. In the\nexample of Figure 3c with binary random variables, the model has 11 parameters. How-\never, these parameters are determined by the environment: To be adaptive in nonstationary\nenvironments, the model must be updated following each experienced state. We propose a\nsimple exponentially weighted averaging approach. For two variables V and W with\nobserved values v and w on trial t, a conditional distribution, \n, is\n\nPt V u= W w=\n\n\u03b4uv\n\n(\n\n)\n\n=\n\n(a)\n\nTcolor\n\nDcolor\n\n(b)\n\nTcolor\n\nDcolor\n\n(c)\n\nTcolor\n\nDcolor\n\nTsize\n\nDsize\n\nTsize\n\nDsize\n\nTsize\n\nDsize\n\nTnotch\n\nDnotch\n\nTnotch\n\nDnotch\n\nTnotch\n\nDnotch\n\nFIGURE 3. Three models of a visual-search environment with colored, notched, size-varying diamonds. (a)\nfeature-independence model; (b) full-joint model; (c) dominance model.\n\n\f is the Kronecker delta. The distribution representing the environment\n\n\u03b4\n\ndefined, where \nfollowing trial t, denoted \nE V u= W w=\nPt\n(\n\n)\n\nE\nPt\n=\n\n, is then updated as follows:\n\u03b1Pt\n\nE V u= W w=\n1\u2013\n\n+\n\n)\n\n(\n\n)\n\n\u03b1\n\nE\nP0\n\n1 \u03b1\u2013(\n\n)Pt V u= W w=\n\n(\n\n(2)\n is a memory constant. Note that no update is performed for values of W\n\nfor all u, where \nother than w. An analogous update is performed for unconditional distributions. \nHow the model is initialized\u2014i.e., specifying \u2014is irrelevant, because all experimental\ntasks that we model, participants begin the experiment with many dozens of practice trials.\nE\nP0\n do\nData is not collected during practice trials. Consequently, any transient effects of \nE\nP0\nnot impact the results. In our simulations, we begin with a uniform distribution for \n,\nand include practice trials as in the human studies.\nThus far, we\u2019ve assumed a single target and a single distractor. The experiments that we\nmodel involve multiple distractors. The simple extension we require to handle multiple\ndistractors is to define a frequentist probability for each distractor feature V,\nPt V v= W w=\n, where \n is the count of co-occurrences of feature val-\nues v and w among the distractors, and \nOur model is extremely simple. Given a description of the visual search task and environ-\nment, the model has only a single degree of freedom, \n. In all simulations, we fix\n\u03b1\n\n does not qualitatively impact any result.\n\n; however, the choice of \n\n is the count of w.\n\nCvw Cw\u2044\n\nCvw\nCw\n\n0.75\n\n)\n\n=\n\n(\n\n=\n\n\u03b1\n\n\u03b1\n\n4 SIMULATIONS\nIn this section, we show that the model can explain a range of data from four different\nexperiments examining attentional priming. All experiments measure response times of\nparticipants. On each trial, the model can be used to obtain a probability of the display\nconfiguration (the environmental state) on that trial, given the history of trials to that\npoint. Our critical assumption\u2014as motivated earlier\u2014is that response times monotoni-\ncally decrease with increasing probability, indicating that visual information processing is\nbetter configured for more likely environmental states. The particular relationship we\nassume is that response times are linear in log probability. This assumption yields long\nresponse time tails, as are observed in all human studies. \n\n4.1 Maljkovic and Nakayama (1994, Experiment 5)\nIn this experiment, participants were asked to search for a singleton in color in a display of\nthree red or green diamonds. Each diamond was notched on either the left or right side,\nand the task was to report the side of the notch on the color singleton. The well-practiced\nparticipants made very few errors. Reaction time (RT) was examined as a function of\nwhether the target on a given trial is the same or different color as the target on trial n steps\nback or ahead. Figure 4 shows the results, with the human RTs in the left panel and the\nsimulation log probabilities in the right panel. The horizontal axis represents n. Both\ngraphs show the same outcome: repetition of target color facilitates performance. This\ninfluence lasts only for a half dozen trials, with an exponentially decreasing influence fur-\nther into the past. In the model, this decreasing influence is due to the exponential decay of\nrecent history (Equation 2). Figure 4 also shows that\u2014as expected\u2014the future has no\ninfluence on the current trial.\n\n4.2 Maljkovic and Nakayama (1994, Experiment 8)\nIn the previous experiment, it is impossible to determine whether facilitation is due to rep-\netition of the target\u2019s color or the distractor\u2019s color, because the display contains only two\ncolors, and therefore repetition of target color implies repetition of distractor color. To\nunconfound these two potential factors, an experiment like the previous one was con-\n\n\fducted using four distinct colors, allowing one to examine the effect of repeating the target\ncolor while varying the distractor color, and vice versa. The sequence of trials was com-\nposed of subsequences of up-to-six consecutive trials with either the target or distractor\ncolor held constant while the other color was varied trial to trial. Following each subse-\nquence, both target and distractors were changed. Figure 5 shows that for both humans and\nthe simulation, performance improves toward an asymptote as the number of target and\ndistractor repetitions increases; in the model, the asymptote is due to the probability of the\nrepeated color in the environment model approaching 1.0. The performance improvement\nis greater for target than distractor repetition; in the model, this difference is due to the\ndominance of the defining feature of the target over the defining feature of the distractor.\n\n4.3 Huang, Holcombe, and Pashler (2004, Experiment 1)\nHuang et al. (2004) and Hillstrom (2000) conducted studies to determine whether repeti-\ntions of one feature facilitate performance independently of repetitions of another feature.\nIn the Huang et al. study, participants searched for a singleton in size in a display consist-\ning of lines that were short and long, slanted left or right, and colored white or black. The\nreported feature was target slant. Slant, size, and color were uncorrelated. Huang et al. dis-\ncovered that repeating an irrelevant feature (color or orientation) facilitated performance,\nbut only when the defining feature (size) was repeated. As shown in Figure 6, the model\nreplicates human performance, due to the dominance of the defining feature over the\nreported and irrelevant features.\n\n4.4 Wolfe, Butcher, Lee, and Hyde (2003, Experiment 1)\nIn an empirical tour-de-force, Wolfe et al. (2003) explored singleton search over a range of\nenvironments. The task is to detect the presence or absence of a singleton in displays con-\n\n)\nc\ne\ns\nm\n\n \n\ni\n\n(\n \ne\nm\nT\nn\no\ni\nt\nc\na\ne\nR\n\n610\n\n600\n\n590\n\n580\n\n570\n\n560\n\n550\n\nHuman data\n\nDifferent Color \n\nSame Color \n\n(cid:358)15\n\n(cid:358)13\n\n(cid:358)11\n\n(cid:358)9\n\n(cid:358)7\n\n(cid:358)5\n\n(cid:358)3\n\n(cid:358)1\n\n+1\n\nPast\n\nRelative Trial Number \n\n+5\n\n+3\nFuture\n\n+7\n\n)\n)\nl\na\ni\nr\nt\n(\nP\n(\ng\no\nl\n(cid:358)\n\n3.4\n\n3.2\n\n3\n\n2.8\n\n2.6\n\nSimulation\n\nDifferent Color \n\nSame Color \n\n(cid:358)15\n\n(cid:358)13\n\n(cid:358)11\n\n(cid:358)9\n\n(cid:358)7\n\n(cid:358)5\n\n(cid:358)3\n\n(cid:358)1\n\n+1\n\nPast\n\nRelative Trial Number \n\n+5\n\n+3\nFuture\n\n+7\n\nFIGURE 4. Experiment 5 of Maljkovic and Nakayama (1994): performance on a given trial conditional on the\ncolor of the target on a previous or subsequent trial. Human data is from subject KN.\n\nFIGURE 5. Experiment 8 of\nMaljkovic \nNakayama\n(1994). (left panel) human data,\naverage of subjects KN and\nSS; (right panel) simulation\n\nand \n\nFIGURE 6. Experiment 1 of\nHuang, Holcombe, & Pashler\n(2004). (left panel) human data;\n(right panel) simulation\n\n)\nc\ne\ns\nm\n\n \n\ni\n\n(\n \ne\nm\nT\nn\no\ni\nt\nc\na\ne\nR\n\n650\n\n640\n\n630\n\n620\n\n610\n\n600\n\n590\n\n580\n\n1\n\nDistractors\nSame\n\nTarget\nSame\n\n)\n)\nl\na\ni\nr\nt\n(\nP\n(\ng\no\nl\n(cid:358)\n\n6\n\n5.5\n\n5\n\n4.5\n\n4\n\n3.5\n\n3\n\nDistractors\nSame\n\nTarget\nSame\n\n2\nOrder in Sequence\n\n4\n\n3\n\n5\n\n6\n\n1\n\n2\nOrder in Sequence\n\n3\n\n4\n\n5\n\n6\n\n1050\n\n)\nc\ne\ns\nm\n\n1000\n\n \n\ni\n\n(\n \ne\nm\nT\nn\no\ni\nt\nc\na\ne\nR\n\n950\n\n900\n\n850\n\n Size Alternate\n\nSize Repeat \n\n)\n)\nl\na\ni\nr\nt\n(\nP\n(\ng\no\nl\n(cid:358)\n\n4.2\n\n4\n\n3.8\n\n3.6\n\n3.4\n\n3.2\n\n3\n\n Size Alternate\n\nSize Repeat \n\nColor Repeat \n\nColor Alternate \n\nColor Repeat \n\nColor Alternate \n\n\fsisting of colored (red or green), oriented (horizontal or vertical) lines. Target-absent trials\nwere used primarily to ensure participants were searching the display. The experiment\nexamined seven experimental conditions, which varied in the amount of uncertainty as to\nthe target identity. The essential conditions, from least to most uncertainty, are: blocked\n(e.g., target always red vertical among green horizontals), mixed feature (e.g., target\nalways a color singleton), mixed dimension (e.g., target either red or vertical), and fully\nmixed (target could be red, green, vertical, or horizontal). With this design, one can ascer-\ntain how uncertainty in the environment and in the target definition influence task diffi-\nculty. Because the defining feature in this experiment could be either color or orientation,\nwe modeled the environment with two Bayes nets\u2014one color dominant and one orienta-\ntion dominant\u2014and performed model averaging. A comparison of Figures 7a and 7b\nshow a correspondence between human RTs and model predictions. Less uncertainty in\nthe environment leads to more efficient performance. One interesting result from the\nmodel is its prediction that the mixed-feature condition is easier than the fully-mixed con-\ndition; that is, search is more efficient when the dimension (i.e., color vs. orientation) of\nthe singleton is known, even though the model has no abstract representation of feature\ndimensions, only feature values.\n\n\u03b1\n\n4.5 Optimal adaptation constant\nIn all simulations so far, we fixed the memory constant. From the human data, it is clear\nthat memory for recent experience is relatively short lived, on the order of a half dozen tri-\nals (e.g., left panel of Figure 4). In this section we provide a rational argument for the short\nduration of memory in attentional control.\nFigure 7c shows mean negative log probability in each condition of the Wolfe et al. (2003)\nexperiment, as a function of \n. To assess these probabilities, for each experimental condi-\ntion, the model was initialized so that all of the conditional distributions were uniform,\nand then a block of trials was run. Log probability for all trials in the block was averaged.\nThe negative log probability (y axis of the Figure) is a measure of the model\u2019s mispredic-\ntion of the next trial in the sequence.\nFor complex environments, such as the fully-mixed condition, a small memory constant is\ndetrimental: With rapid memory decay, the effective history of trials is a high-variance\nsample of the distribution of environmental states. For simple environments, a large mem-\nory constant is detrimental: With slow memory decay, the model does not transition\nquickly from the initial environmental model to one that reflects the statistics of a new\nenvironment. Thus, the memory constant is constrained by being large enough that the\nenvironment model can hold on to sufficient history to represent complex environments,\nand by being small enough that the model adapts quickly to novel environments. If the\nconditions in Wolfe et al. give some indication of the range of naturalistic environments an\nagent encounters, we have a rational account of why attentional priming is so short lived.\nWhether priming lasts 2 trials or 20, the surprising empirical result is that it does not last\n200 or 2000 trials. Our rational argument provides a rough insight into this finding.\n\n(a)\n\nHuman Data\n\n(b)\n\nSimulation\n\n(c)\n\n)\nc\ne\ns\nm\n\n(\n \ne\nm\n\ni\nt\n \n\nn\no\n\ni\nt\nc\na\ne\nr\n\n480\n\n460\n\n440\n\n420\n\n400\n\n380\n\n360\n\nfully mixed\nmixed feature\nmixed dimension\nblocked\n\nfully mixed\nmixed feature\nmixed dimension\nblocked\n\n4\n\n3\n\n2\n\n1\n\n0\n\n)\n)\nl\na\ni\nr\nt\n(\n\nP\n(\ng\no\n(cid:358)\n\nl\n\nred or vert \n\ntarget type\n\nred and vert\n\nred or vert \n\ntarget type\n\nred and vert\n\n)\n)\nl\na\ni\nr\nt\n(\nP\n(\ng\no\nl\n(cid:358)\n\n5\n\n4\n\n3\n\n2\n\n1\n\n0\n0\n\n0.5\n\nBlocked Red or Vertical\nBlocked Red and Vertical\nMixed Feature\nMixed Dimension\nFully Mixed\n\n0.8\n\nMemory Constant\n\n0.9\n\n0.95\n\n0.98\n\nFIGURE 7. (a) Human data for Wolfe et al. (2003), Experiment 1; (b) simulation; (c) misprediction of model\n(i.e., lower y value = better) as a function of \n\n for five experimental condition\n\n\u03b1\n\n\f5 DISCUSSION\nThe psychological literature contains two opposing accounts of attentional priming and its\nrelation to attentional control. Huang et al. (2004) and Hillstrom (2000) propose an epi-\nsodic account in which a distinct memory trace\u2014representing the complete configuration\nof features in the display\u2014is laid down for each trial, and priming depends on configural\nsimilarity of the current trial to previous trials. Alternatively, Maljkovic and Nakayama\n(1994) and Wolfe et al. (2003) propose a feature-strengthening account in which detection\nof a feature on one trial increases its ability to attract attention on subsequent trials, and\npriming is proportional to the number of overlapping features from one trial to the next.\nThe episodic account corresponds roughly to the full joint model (Figure 3b), and the fea-\nture-strengthening account corresponds roughly to the independence model (Figure 3a).\nNeither account is adequate to explain the range of data we presented. However, an inter-\nmediate account, the dominance model (Figure 3c), is not only sufficient, but it offers a\nparsimonious, rational explanation. Beyond the model\u2019s basic assumptions, it has only one\nfree parameter, and can explain results from diverse experimental paradigms.\nThe model makes a further theoretical contribution. Wolfe et al. distinguish the environ-\nments in their experiment in terms of the amount of top-down control available, implying\nthat different mechanisms might be operating in different environments. However, in our\naccount, top-down control is not some substance distributed in different amounts depend-\ning on the nature of the environment. Our account treats all environments uniformly, rely-\ning on attentional control to adapt to the environment at hand. \nWe conclude with two limitations of the present work. First, our account presumes a par-\nticular network architecture, instead of a more elegant Bayesian approach that specifies\npriors over architectures, and performs automatic model selection via the sequence of tri-\nals. We did explore such a Bayesian approach, but it was unable to explain the data. Sec-\nond, at least one finding in the literature is problematic for the model. Hillstrom (2000)\noccasionally finds that RTs slow when an irrelevant target feature is repeated but the defin-\ning target feature is not. However, because this effect is observed only in some experi-\nments, it is likely that any model would require elaboration to explain the variability.\n\nACKNOWLEDGEMENTS\nWe thank Jeremy Wolfe for providing the raw data from his experiment for reanalysis. This research was funded\nby NSF BCS Award 0339103.\nREFERENCES\nHuang, L, Holcombe, A. O., & Pashler, H. (2004). Repetition priming in visual search: Episodic retrieval, not\n\nfeature priming. Memory & Cognition, 32, 12\u201320.\n\nHillstrom, A. P. (2000). Repetition effects in visual search. Perception & Psychophysics, 62, 800-817.\nItti, L., Koch, C., & Niebur, E. (1998). A model of saliency-based visual attention for rapid scene analysis. IEEE\n\nTrans. Pattern Analysis & Machine Intelligence, 20, 1254\u20131259.\n\nKearns, M., & Singh, S. (1999). Finite-sample convergence rates for Q-learning and indirect algorithms. In\n\nAdvances in Neural Information Processing Systems 11 (pp. 996\u20131002). Cambridge, MA: MIT Press.\n\nKoch, C. and Ullman, S. (1985). Shifts in selective visual attention: towards the underlying neural circuitry.\n\nHuman Neurobiology, 4, 219\u2013227.\n\nMaljkovic, V., & Nakayama, K. (1994). Priming of pop-out: I. Role of features. Mem. & Cognition, 22, 657-672.\nMozer, M. C. (1991). The perception of multiple objects: A connectionist approach. Cambridge, MA: MIT Press.\nRogers, R. D., & Monsell, S. (1995). The cost of a predictable switch between simple cognitive tasks. Journal of\n\nExperimental Psychology: General, 124, 207\u2013231.\n\nWolfe, J.M. (1994). Guided Search 2.0: A Revised Model of Visual Search. Psych. Bull. & Rev., 1, 202\u2013238.\nWolfe, J. S., Butcher, S. J., Lee, C., & Hyde, M. (2003). Changing your mind: on the contributions of top-down\nand bottom-up guidance in visual search for feature singletons. Journal of Exptl. Psychology: Human Percep-\ntion & Performance, 29, 483-502.\n\n\f", "award": [], "sourceid": 2875, "authors": [{"given_name": "Michael", "family_name": "Shettel", "institution": null}, {"given_name": "Shaun", "family_name": "Vecera", "institution": null}, {"given_name": "Michael", "family_name": "Mozer", "institution": null}]}