`ARTICLE NO. 0002
`
`Just Say No: How Are Visual Searches Terminated
`When There Is No Target Present?
`MARVIN M. CHUN AND JEREMY M. WOLFE
`Massachusetts Institute of Technology and Harvard Medical School
`
`How should a visual search task be terminated when no target is found? Such
`searches could end after a serial search through all items, but blank trials in many
`tasks are terminated too quickly for that to be plausible. This paper proposes a solution
`based on Wolfe’s (1994) Guided Search model. The probability that each item is a
`target is computed in parallel based on items’ differences from each other and their
`similarity to the desired target. This probability is expressed as an activation. Activa-
`tions are examined in decreasing order until the target is found or until an activation
`threshold is reached. This threshold is set adaptively by the observer—more conserva-
`tive following misses, more liberal following successful trials. In addition, observers
`guess on some trials. The probability of a guess increases as trial duration increases.
`The model successfully explains blank trial performance. Specific predictions are
`tested by experiments. q 1996 Academic Press, Inc.
`
`Suppose that you have written an important phone number on a small piece
`of paper. You are searching for that piece of paper among a mess of various
`articles, journals, forms, and other miscellaneous paperwork on your desk.
`When should one stop searching? Certainly, one can stop when the phone
`number is found, but if it is not found, determining how long to keep looking
`depends on other factors. One could perform a serial, exhaustive search,
`checking every sheet of paper in the office until the phone number is found
`or no papers remain unexamined. More reasonably, one could stop searching
`when no likely papers remain unexamined. For instance, if the phone number
`was written on a small piece of paper, then a more efficient strategy would
`be to search just through items of similar size.
`
`Marvin M. Chun is from the Department of Brain and Cognitive Sciences; Jeremy M. Wolfe
`is from the Center for Ophthalmic Research, Brigham & Women’s Hospital and Department of
`Ophthalmology. Portions of this work were presented as a poster at the 1991 Annual Meeting
`of the Association for Research in Vision and Ophthalmology (ARVO), Sarasota, FL. This
`research was supported by grant EY05087, and in part by grants T32GM07484, and EY06592
`from the National Institutes of Health. We thank Kyle Cave, John Higgins, Lester Krueger,
`Gordon Logan, and Molly Potter for their helpful comments on earlier drafts of this work. We
`also gratefully acknowledge Stacia Friedman-Hill and Michelle Eng for their help in various
`phases of this project. Correspondence concerning this article should be addressed to Marvin M.
`Chun, Vision Sciences Laboratory, Department of Psychology, Harvard University, 33 Kirkland
`Street, Cambridge, MA 02138. E-mail: chun@isr.harvard.edu.
`39
`
`0010-0285/96 $18.00
`Copyright q 1996 by Academic Press, Inc.
`All rights of reproduction in any form reserved.
`
`a301$$0629
`
`04-12-96 18:29:09
`
`cogpa
`
`AP: Cog Psych
`
`IPR2020-00686
`Apple EX1020 Page 1
`
`
`
`40
`
`CHUN AND WOLFE
`This paper examines the termination of unsuccessful searches in a labora-
`tory task corresponding to the previous example, the visual search paradigm.
`Visual search has been a very useful tool in the analysis of visual processing
`(see Grossberg, Mingolla, & Ross, 1994; Treisman, 1988; and Wolfe, 1994,
`for reviews). In a visual search task, an observer looks for a designated target
`item among a variable number of distractor items. Typically, reaction time
`(RT) measures are taken to determine the time required for an observer to
`respond ‘‘yes’’ when a target is present or ‘‘no’’ when it is absent. The
`number of distractors (the ‘‘set size’’) is the independent variable. The shape
`and slope of RT 1 set size functions are the dependent measures that change
`with different search tasks. These changes have been used to infer two stages
`of visual processing. Some searches, notably those for targets defined by
`a single, basic feature (e.g., color, orientation, size), produce RTs that are
`independent of set size for target-present and for target-absent (blank) trials.
`This is taken as evidence for a parallel stage that can process some aspects
`of visual input across large parts of the visual field at one time (Treisman &
`Souther, 1985). Other searches (e.g., a search for a ‘‘T’’ among ‘‘L’’s) pro-
`duce a linear increase in RT with set size. Characteristically, these searches
`yield RT 1 set size slopes of 20–30 ms/item on target trials and twice that,
`40–60 ms/item on blank trials (Treisman, 1988). Steep slopes and this 2:1
`ratio of blank to target trial slopes are indicative of a serial, self-terminating
`search (Kwak, Dagenbach, & Egeth, 1991; but see Townsend, 1990, for a
`discussion of the fact that limited-capacity parallel searches could underlie
`such search results).
`There are visual search tasks that produce results lying between the classic
`‘‘parallel’’ and ‘‘serial’’ patterns. In what we will call guided searches, infor-
`mation from the parallel stage of processing is used to guide the subsequent,
`serial deployment of visual attention from item to item (Cave & Wolfe, 1990;
`Hoffman, 1978, 1979; Wolfe, 1994; Wolfe & Cave, 1989; Wolfe, Cave, &
`Franzel, 1989). Thus, in a search for a red vertical target, parallel color
`processes can guide attention toward red items while parallel orientation
`processes guide attention toward vertical items producing quite shallow slopes
`(Cohen & Ivry, 1991; Egeth, Virzi, & Garbart, 1984; Treisman & Sato, 1990a;
`Wolfe et al., 1989) even though no single parallel process is sensitive to the
`conjunction of color and orientation (Wolfe, Chun, & Friedman-Hill, 1995).
`The existence of guided searches raises the theoretical problem that is at the
`heart of this paper. How and when should an observer terminate a search
`when no target is present? In principle, the answer is easy enough for strictly
`parallel and serial searches. In a parallel search, all items are processed in
`parallel, allowing for an efficient decision about target presence or absence.
`In a serial, self-terminating search, the observer searches until she finds the
`target or, on blank trials, until she has exhaustively examined the entire set
`of items, one by one. The endpoint of an unsuccessful guided search is less
`obvious. A guided search is a serial search through a subset of the items.
`
`a301$$0629
`
`04-12-96 18:29:09
`
`cogpa
`
`AP: Cog Psych
`
`IPR2020-00686
`Apple EX1020 Page 2
`
`
`
`41
`SEARCH TERMINATION
`How does the observer know that she has exhausted that subset? Why doesn’t
`the observer search all items on blank trials?
`DECISION PROCESSES IN VISUAL SEARCH
`We can imagine two different strategies for deciding when a target is not
`present without having to exhaustively search the entire display. People may
`simply search through the distractors that have a certain likelihood of being
`a target and ignore those items which are less similar to the target. For reasons
`that will be made clear below, we will refer to this as the activation threshold
`hypothesis. Another possibility is that as an observer runs in a visual search
`task, she may develop some internal estimate of how long it takes to find a
`target. With such an estimate, she may be able to make ‘‘educated guesses,’’
`since the probability of a guess being correct increases as evidence is accumu-
`lated as a search trial progresses. According to this timing hypothesis, observ-
`ers will terminate a trial when the duration of the trial exceeds some duration
`threshold, based on the assumption that the target should have been found
`by then.
`In this paper, we explore these two hypotheses in the context of the Guided
`Search model (Cave & Wolfe, 1990; Wolfe, 1994; Wolfe & Cave, 1989;
`Wolfe et al., 1989). Neither of the decision mechanisms we propose are
`entirely dependent on the specifics of the Guided Search model. However,
`Guided Search is sufficiently detailed to allow for specific, quantitative predic-
`tions to be tested. The Guided Search model holds that attention is guided
`toward candidate target items by parallel processes that activate items pos-
`sessing one or more target features. Activation is combined across feature
`types so that an item having two target features will receive more activation
`on average than an item having only one such feature. For example, consider
`a search for a red vertical item among red horizontal and green vertical
`distractors. In this guided search for a conjunction, the parallel feature proces-
`sor for color would activate all ‘‘red’’ items while the orientation processor
`would activate all ‘‘vertical’’ items. Information from these two feature mod-
`ules would be combined into an overall activation map. Attention is guided
`from one candidate target to the next in a serial manner in order of decreasing
`activation. If this combination of information from the parallel stage processes
`were perfect, then the target, if present, would always receive the highest
`activation and search for a conjunction target would be no less ‘‘parallel’’
`than search for a feature target. However, activations appear to be embedded
`in internal noise. The result is that an average target will have a higher
`activation than an average distractor but, on any one trial, some distractors
`may have higher activation than the target. As noted, in the Guided Search
`model, attention is deployed from item to item in decreasing order of activa-
`tion strength. Thus, ‘‘noisy,’’ high activation distractors will have to be
`checked and rejected by the serial stage first before the actual target is found.
`We do not have direct access to the thresholds and decision rules proposed
`
`a301$$0629
`
`04-12-96 18:29:09
`
`cogpa
`
`AP: Cog Psych
`
`IPR2020-00686
`Apple EX1020 Page 3
`
`
`
`42
`CHUN AND WOLFE
`here but we can predict their impact on RTs and error rates. The paper is
`organized in the following manner: First we describe the proposed activation
`threshold mechanism in more detail and present some simulation results based
`on theory that the activation threshold alone accounts for blank trial perfor-
`mance. These simulation results deviate from the data in the literature, so we
`next incorporate the timing/guessing mechanism into the model. This gener-
`ates better simulation results and a set of concrete predictions. Finally, these
`predictions are tested empirically in a series of three visual search experi-
`ments. The results are consistent with the model.
`The Activation Threshold Mechanism
`What happens on blank trials in a guided search? Clearly, it would be
`inefficient to search exhaustively through the entire display, since the average
`activation of the target will lie above the average activation of the distractors
`(otherwise there is no guidance). It should be possible to safely reject some
`distractors in parallel on the basis of their low activations. We propose that
`an internal activation threshold is used as a cutoff criterion for terminating
`search on blank trials. Setting the correct value for this activation threshold
`is essentially a signal detection problem (see Pavel, Econopouly, & Landy,
`1992). This is illustrated in Fig. 1 which shows hypothetical average activation
`distributions for distractors and for targets compiled across trials for two
`visual search tasks. The activation of the distractors is modeled as a normal
`distribution with some mean and standard deviation. The distribution from
`which the target activation will be drawn on any given trial is simply this
`distractor distribution with an activation ‘‘signal’’ added to it. RTs for target
`and blank trials can be derived graphically from Fig. 1. On a target trial,
`observers will have to examine all distractors with activations above the target
`activation value. On average, the proportion of distractors examined will be
`those falling in region ‘‘C,’’ the region of the distractor distribution above the
`average target activation. On blank trials, the activation threshold hypothesis
`simply states that observers examine all distractors above an activation thresh-
`old. Thus, the proportion of distractors examined on an average blank trial
`will correspond to the area of the distractor distribution in regions B and C
`in Fig. 1. If the activation threshold is set to a level higher than the lowest
`target activation, then on some target trials, search will be terminated before
`the target is found. The target will be missed and a miss error is generated. The
`proportion of target activations that gives rise to miss errors will correspond to
`region A. If the activation threshold is set higher, fewer distractors will be
`checked on blank trials, reducing blank trial RTs. However more misses will
`be produced; a classic speed-accuracy trade-off. If a different task produces
`a greater signal (Fig. 1b), fewer distractors will lie above either the average
`target activation or the activation threshold and, thus, target and blank RTs
`will decline.
`The distributions in Fig. 1 are fine theoretical constructs but it is implausible
`
`a301$$0629
`
`04-12-96 18:29:09
`
`cogpa
`
`AP: Cog Psych
`
`IPR2020-00686
`Apple EX1020 Page 4
`
`
`
`SEARCH TERMINATION
`
`43
`
`FIG. 1. Visual search can be conceived of as a signal detection problem where each distractor
`item has an activation drawn from some noise distribution and where the target has an activation
`drawn from that distribution with a signal added to it. Average target-present RTs correspond
`to the proportion of distractors with activations above the average target activation (area C). In
`the present model, blank trial RTs correspond to the proportion of distractors above the activation
`threshold (areas B / C) and miss rate corresponds to the proportion of targets with activations
`below the activation threshold (area A).
`
`to assume that observers have any conscious or unconscious access to the
`precise shape of the distributions or to the magnitudes of activations. How,
`then, is the observer to set an activation threshold that will produce an accept-
`able error rate? We have modeled this setting of the threshold as an internal
`‘‘staircase’’ procedure. In visual search tasks, the observer is usually in-
`structed to respond as fast as possible while making few errors (5–10%).
`According to the model, the observer keeps the RTs and the error rate at the
`desired levels as follows: When the observer correctly terminates a blank
`trial search, he raises his activation threshold in order to terminate the next
`blank trial more quickly and improve his overall speed. When the observer
`commits an error, he lowers his activation threshold in order not to miss
`future targets with low activation, thus preserving an acceptable error rate.
`The specific error rate can be controlled by varying the relative size of the
`‘‘up’’ and ‘‘down’’ staircase ‘‘steps’’ (Levitt, 1971). For example, if the step
`‘‘down’’ following a miss is 20 times the size of the step up following a
`successful blank trial, then the staircase converges on an error rate of approxi-
`mately 4% (see simulation below).
`
`a301$$0629
`
`04-12-96 18:29:09
`
`cogpa
`
`AP: Cog Psych
`
`IPR2020-00686
`Apple EX1020 Page 5
`
`
`
`44
`CHUN AND WOLFE
`Blank-Target Slope Ratios: A Retrospective Analysis
`In the signal detection scheme outlined above and illustrated in Fig. 1, the
`ratio of blank trial slopes to target trial slopes is determined by the ratio of
`areas B / C to area C. For standard serial, self-terminating slopes, that ratio
`would be 2:1. Since this paper is intended to model the full range of visual
`search tasks from parallel through guided to serial searches, it is important
`to know the normative slope ratio for a range of search tasks. To obtain this
`information, we retrospectively examined slope ratios for a wide range of
`search tasks. To do this we took 733 pairs of target and blank trial slopes
`from 65 different search experiments run in this laboratory over a period of
`4 years. Each pair represents 300 trials run on one observer. Most observers
`are represented several times because they were tested on several search tasks
`but at least 100 individuals are represented. Search tasks include simple
`feature searches, complex (‘‘serial’’) feature search, easy and hard conjunction
`searches, and ‘‘serial’’ searches (e.g., ‘‘T’’ among ‘‘L’’s). The slowest
`searches are those for conjunctions of two colors or two orientations (Wolfe
`et al., 1990) and some complex orientation searches (Wolfe, Friedman-Hill,
`Stewart, & O’Connell, 1992). Most of the data has been reported in previous
`work from this laboratory (cf., Wolfe, 1994).
`The results of this analysis are shown in Fig. 2. The main figure shows
`all data with target slopes between 0 and 60 ms/item and blank slopes
`between 0 and 150 ms/item. The inset shows all slopes. The solid line is
`the 2:1 ratio line and not the regression line. It is clear that a 2:1 ratio is a
`reasonable description of the main trend in the data. For the entire data set,
`the regression slope is 1.99 with an R2 value of 0.77. Removing the highest
`slopes produces a slope of 1.70. The deviation from 2.0 seems to reflect
`the influence of points near 0.0 where the ratio becomes meaningless. If,
`for example, we examine the 168 points with target slopes between 5 and
`12 ms/item (a reasonable definition of ‘‘guided’’ searches), the regression
`line has a slope of 2.04, though R2 is much reduced (0.12). The purpose of
`this analysis is to show that a 2:1 slope ratio is the most reasonable fit to
`the full range of various search tasks that have been tested in this laboratory.
`This is not meant to imply that every search task will show 2:1 slope ratios.
`Indeed, as shown in Fig. 2, there is considerable variation of slope ratios.
`Breaking down the analysis by search type (e.g., feature search vs conjunc-
`tion search) does not lead to any systematic variation between search task
`and slope ratio. In sum, for general modeling purposes, we choose to simu-
`late the search task to approximate the normative 2:1 slope ratio that best
`fits the entire range of data.
`The 2:1 ratio raises a problem. Returning to Fig. 1, we obtain a 2:1 slope
`ratio when area B equals area C. This may be the case for one signal strength,
`but as the distribution of target activations slides left and right with changes
`in signal strength, the ratio of area B to area C will change. Specifically, as
`
`a301$$0629
`
`04-12-96 18:29:09
`
`cogpa
`
`AP: Cog Psych
`
`IPR2020-00686
`Apple EX1020 Page 6
`
`
`
`SEARCH TERMINATION
`
`45
`
`FIG. 2. Blank trial slope as a function of target trial slope. Each of the 733 points represents
`one observer performing 300 trials on one visual search task. Over 70 different search tasks are
`represented. Though there is substantial scatter, it is clear that a 2:1 slope ratio is a good
`approximation to the data over the full range of search tasks.
`
`signal strength increases, area C will decrease faster than area B as can be
`seen in Fig. 1. If the activation distributions for targets and distractors were
`normal and had the same standard deviation, the slope ratio should increase
`systematically with increasing signal strength. As discussed above, in the
`actual data, there is no evidence for any such systematic deviation change in
`slope ratios.
`There are, no doubt, a host of possible solutions to this problem. The
`solution used in Guided Search 2.0 (Wolfe, 1994) is illustrated in Fig. 3. If
`the targets are drawn from a distribution whose standard deviation declines
`as the signal strength increases, then the distance, in activation units, between
`the average target activation and the activation threshold decreases with in-
`creasing signal strength and an average 2:1 ratio can be maintained.
`Though we do not know the neural correlates of target and distractor
`activations, the idea that target activations may become more precise as signal
`strength increases seems plausible. One scenario that we have simulated could
`be labeled ‘‘neuron recruitment.’’ Suppose that each ‘‘neuron’’ in the system
`has the same variability. The number of neurons activated by an item is
`directly related to the signal strength of that item: more signal, more neurons.
`The overall activation for an item is the average of the activated neurons. By
`
`a301$$0629
`
`04-12-96 18:29:09
`
`cogpa
`
`AP: Cog Psych
`
`IPR2020-00686
`Apple EX1020 Page 7
`
`
`
`46
`
`CHUN AND WOLFE
`
`FIG. 3. Target and blank trial slopes are in a 2:1 slope ratio when areas B and C of Fig. 1 are
`equal. This relationship will not be maintained as signal strength changes if distractor and target
`distributions are normal and have the same variance. One way to preserve the relationship is
`shown here. Reduce the variance of the target distribution as signal strength increases.
`
`the law of large numbers, an increased number of activated units with unit
`variance results in smaller variability in their averaged output. Thus, stronger
`signals have smaller variance (see Zohary, Shadlen, & Newsome, 1994). In
`the full Guided Search 2.0 simulation of Wolfe (1994), this scheme produced
`results that were in good qualitative agreement with human data. In the more
`limited simulation reported in the remainder of this paper, we manipulate the
`variance directly with a function that reduces the variance as signal strength
`increases and that yields the approximate 2:1 slope ratio seen in the human
`data. Details are discussed below.
`Simulating the Activation Threshold
`In the simulation, each trial was modeled as a list of activations. Distractor
`activations were generated from a normal population with a mean of 300 and
`a standard deviation of 100 arbitrary units. On half the trials, one distractor
`was replaced by a target item. For any block of trials, this target was drawn
`from a distribution with a mean defined by the size of the signal and a standard
`deviation calculated to satisfy two constraints: First, the activation threshold
`must be set so that the proportion of distractor activations lying between the
`activation threshold and the target average will be equal to the proportion of
`
`a301$$0629
`
`04-12-96 18:29:09
`
`cogpa
`
`AP: Cog Psych
`
`IPR2020-00686
`Apple EX1020 Page 8
`
`
`
`47
`SEARCH TERMINATION
`distractor activations above the target average (areas B and C in Fig. 1). This
`will yield a 2:1 slope ratio. Second, the activation threshold must lie about
`1.7 standard deviations below the target average on the target distribution.
`This will yield an approximate 4% error rate (area A in Fig. 1). Different
`error rates can be produced by changing the distance from target average to
`activation threshold. Given these constraints (and a tabulated normal distribu-
`tion), the required standard deviation of the target distribution can be deter-
`mined for any signal strength. For computing purposes, relationship of the
`standard deviation of the target distribution to the signal can be well approxi-
`mated by
`
`1
`(1)
`SDtarget Å SDdistractor 1
`1 / (2 1 Zsignal)
`Where Zsignal is the Z-score of the signal in the distractor distribution. When
`there is no signal, SDtarget equals SDdistractor. As the signal grows, SDtarget de-
`creases, going to zero in the limit. (Note that this equation has no theoretical
`implications of its own. It represents pure curve fitting.)
`For each trial in the simulation, set size was randomly set to 4, 10, or 16.
`RT was determined by counting the number of distractors with activations
`above the target activation for target trials and by counting the number of
`distractors with activations above the activation threshold for blank and miss
`(error) trials. These counts were converted to RTs in milliseconds by multi-
`plying by 50 ms, an estimate of the amount of time required to process a
`single item. Four hundred fifty milliseconds were added to each simulated
`trial RT as a constant to reflect the time required to make a response.
`The activation threshold was initially set to the mean of the distractor
`distribution. It was adjusted from trial to trial in a staircase manner. If the
`simulated observer correctly terminated a blank trial, the threshold was in-
`creased by one step (5 arbitrary units of activation). If the simulated observer
`made an error, the threshold was decreased by k steps, where ‘‘k’’ is the
`parameter that determines the error rate. For instance, a ‘‘k’’ of 20 steps (100
`units) yields an error rate of about 4%. The threshold was not changed on
`successful target-present trials. Thus, for any run of the simulation for the
`activation threshold, the relevant parameters were signal strength and the
`staircase parameter. We are not proposing that real observers have any specific
`notion about these staircase steps. The step size is merely a way to express
`the real observer’s automatic effort to set an appropriate activation threshold.
`Our implementation of this is a computationally simple approximation to the
`idea of incrementing the threshold in little steps with each successful ‘‘no’’
`response and of decrementing the threshold in big jumps with each ‘‘miss’’
`error. Successful blank trials should suggest to the observer that the search
`could be terminated more quickly and that the threshold could rise. Misses
`demonstrate that a search was terminated too quickly and that the threshold
`should be lowered. The size of the drop in threshold is determined by the
`
`a301$$0629
`
`04-12-96 18:29:09
`
`cogpa
`
`AP: Cog Psych
`
`IPR2020-00686
`Apple EX1020 Page 9
`
`
`
`48
`
`CHUN AND WOLFE
`
`FIG. 4. Simulated data for a visual search with a signal strength of 100 and an error rate of
`4%.
`
`observer’s tolerance for errors. Lower tolerance should yield bigger step sizes.
`Note that the observer does not need to keep track of percent error. He merely
`reacts to the feedback from each trial with a lesser or greater degree of
`caution.
`We first present the simulation results for the activation threshold mecha-
`nism alone. Fig. 4 shows average RT as a function of set size for 600 trials
`on one simulated observer. The staircase parameter is 20 and the resulting
`miss error rate is 4.3%. These results would be typical for a ‘‘guided’’ search
`(e.g., conjunction of color and orientation (Treisman & Sato, 1990b; Wolfe
`et al., 1989). Note also that the variability increases with set size and is
`greater for blank trials than for target trials, an attribute of real search data
`(Ward & McClelland, 1989) not accounted for in the earlier versions of the
`Guided Search model.
`The way in which the simulation achieves the results in Fig. 4 is illustrated
`in Fig. 5. Figure 5a shows the activation distributions for targets and dis-
`tractors that underlie the performance shown in Fig. 4a. Average simulated
`target activation is 396 with a standard deviation of 31. Average distractor
`activation is 292; the standard deviation equals 99. 15% of the distractors are
`above the average target activation and will need to be checked on an average
`target trial. If a full serial examination of all items yields 50 ms/item, then a
`search through 15% of the items would yield a slope of 7.5 close to the
`obtained regression slope of 9.5 ms/item, as in Fig. 4a. Forty-two percent of
`distractors are above the average activation threshold of 313 yielding a blank
`trial slope of about 21 as in Fig. 4a. The average activation threshold for
`trials in which the target was missed was 366. This was close to the theoretical
`threshold value of 343 which would lead to a predicted error rate of 4%.
`Figure 5b simply illustrates the continuous adjustment of the threshold over
`
`a301$$0629
`
`04-12-96 18:29:09
`
`cogpa
`
`AP: Cog Psych
`
`IPR2020-00686
`Apple EX1020 Page 10
`
`
`
`SEARCH TERMINATION
`
`49
`
`FIG. 5. (a) shows the activations for targets and distractors underlying the simulated data
`plotted in Fig. 4a. (b) gives a record of the changes in the activation threshold over 200 trials.
`Miss error trials are accompanied by sharp decreases in the activation threshold. The threshold
`gradually increases with each succesful blank trial.
`
`200 of the 500 trials and its relationship to the trial type. The threshold goes
`up a little when blank trials are correct, down a lot when target trials are
`missed.
`The threshold setting staircase parameter determines how blank trials are
`terminated. Simulated results for systematic variation of this parameter are
`
`a301$$0629
`
`04-12-96 18:29:09
`
`cogpa
`
`AP: Cog Psych
`
`IPR2020-00686
`Apple EX1020 Page 11
`
`
`
`50
`
`CHUN AND WOLFE
`
`FIG. 6. Simulation results showing the effect of the step size parameter on Blank trial RT and
`miss error rate.
`
`shown in Fig. 6. Increasing the staircase ‘‘step size’’ decreases miss errors
`and increases blank trial reaction times—a classic speed-accuracy tradeoff.
`The simulation shows the results for six ‘‘observers’’ averaged over runs at
`each of three different signal strengths (0, 100, 200). A more liberal criterion
`(smaller step size) means fewer distractors are checked. Blank trial searches
`terminate more quickly at a cost of more miss errors.
`Interrupt (Timing) Mechanism: The Problem of False Alarms
`The activation threshold hypothesis provides a concise explanation for how
`blank trials are terminated. It makes adequate predictions regarding blank
`trial RTs and miss error rates. However, the activation threshold can say
`nothing about false alarms. While false alarms typically occur less frequently
`than miss errors, the presence of false alarms require some explanation within
`the context of any model of visual search. We propose that false alarm errors
`are typically produced by an ‘‘educated guessing’’ strategy based on timing
`estimates. That is, due to miscellaneous factors such as boredom, fatigue,
`frustration, anticipation, etc., there may be a small proportion of trials where
`observers simply terminate a trial with a guess. These guess responses will
`be correct half of the time probabilistically, allowing the observer to terminate
`a trial more quickly. Moreover the probability of a guess being correct in-
`creases as time elapses within a search trial. Suppose an observer has deter-
`mined through practice that he usually takes less than 1000 ms to find a
`target. If more than a second has elapsed on a certain trial, the observer may
`respond ‘‘no’’, guessing that it was likely to have been a blank trial since
`the target had not been found by then. Observers may also incorrectly guess
`‘‘yes’’ on a proportion of trials, producing a handful of false alarms.
`
`a301$$0629
`
`04-12-96 18:29:09
`
`cogpa
`
`AP: Cog Psych
`
`IPR2020-00686
`Apple EX1020 Page 12
`
`
`
`51
`SEARCH TERMINATION
`A version of this timing hypothesis was previously offered as the primary
`mechanism for terminating blank trials in an earlier version of Guided Search
`(Wolfe & Cave, 1989). According to this account, an observer could use an
`internal timing threshold to terminate a search trial. It would be based on
`some implicit knowledge of the average time it takes to find the target and
`the variance of these ‘‘yes’’ responses. This timing threshold can be estimated
`from the distribution of an individual’s response time distributions for target-
`present responses. The logic is analogous to that shown in Fig. 1 of the
`activation threshold. However, pilot analyses of reaction time distributions
`from data obtained in our lab suggested that observers do not employ a timing
`threshold as the primary mechanism for terminating blank trials. In our present
`model, we incorporate a version of the timing hypothesis into our interrupt
`‘‘guessing’’ mechanism so that the probability of making a guess increases
`as the trial duration increases. This is a plausible assumption given that the
`probability of a guess being correct increases as evidence accumulates during
`a search trial.
`We model the interrupt mechanism as a simple random guessing strategy
`in the simulation. The probability of the simulation for making a guessing
`response is controlled by a single guessing parameter, g. On any given trial,
`for any given ‘‘g,’’ the likelihood of making a guess response increases with
`reaction time. On trials where a guess response was generated, we imple-
`mented the probability of guessing ‘‘no’’ to be 80% and guessing ‘‘yes’’ to
`be 20%.
`A schematic processing flow diagram of the full model is illustrated in
`Fig. 7.
`Full Simulation
`A complete simulation having both activation threshold and interrupt com-
`ponents was run to generate joint predictions for reaction time and error rate
`across a range of visual search tasks. Thus, we varied signal strength, the
`stepsize parameter, and the guessing rule parameter. We tested signal strengths
`0, 100, and 200, to model the full range of visual search tasks ranging from
`serial to guided (conjunction) to feature search. Staircase step sizes were 10