CATEGORIZING A MOVING TARGET IN TERMS OF ITS SPEED ...

249

JOURNAL OF THE EXPERIMENTAL ANALYSIS OF BEHAVIOR 2002, 78, 249–270 NUMBER 3 (NOVEMBER)

CATEGORIZING A MOVING TARGET INTERMS OF ITS SPEED, DIRECTION, OR BOTH

WALTER T. HERBRANSON, THANE FREMOUW, AND CHARLES P. SHIMP

WHITMAN COLLEGE, UNIVERSITY OF WASHINGTON, ANDUNIVERSITY OF UTAH

Pigeons categorized a moving target in terms of its speed and direction in an adaptation of therandomization procedure used to study human categorization behavior (Ashby & Maddox, 1998).The target moved according to vectors that were sampled with equal probabilities from two slightlyoverlapping bivariate normal distributions with the dimensions of speed and direction. On the av-erage, pigeons categorized optimally in that they attended to either speed or direction alone, ordivided attention between them, as was required by different reinforcement contingencies. Decisionbounds were estimated for individual pigeons for each attentional task. Average slopes and y inter-cepts of these individually estimated decision bounds closely approximated the corresponding valuesfor optimal decision bounds. There is therefore at least one task in which pigeons, on the average,display flexibility and quantitative precision in allocating attention to speed and direction when theycategorize moving targets.

Key words: categorization, moving targets, speed, direction

The discrimination of motion can be criti-cal for survival. For many species, it is vital toknow if a predator is approaching or a preyis escaping. Correspondingly, motion percep-tion may have preceded the evolution of oth-er visual processes, such as color perceptionor visual acuity (Husband & Shimizu, 2001;Sekuler, 1975; Walls, 1942). An evolutionaryperspective led Walls to suggest that differentaspects of vision, such as acuity, sensitivity,and color, might be of chief importance todifferent human activities, such as watchmak-ing, night flying, and painting, respectively.Walls wrote, however, ‘‘to animals which in-vented the vertebrate eye, and hold patentson most of the features of the human model,the visual registration of movement was of thegreatest importance’’ (p. 342). Accordingly,one might expect that highly visual nonmam-malian vertebrates, such as birds, would haveevolved complex motion perception giventheir awe-inspiring flight characteristics that

This research was supported by grants from the Na-tional Science Foundation and the National Institute forMental Health. We thank David Wood for help in run-ning the experiments and in conducting some of the ini-tial data analyses. We also thank Sarah Creem, AlysonFroehlich, and David Strayer for helpful discussion andcomments on an earlier draft. We are grateful to GregAshby and Todd Maddox for graciously sending us usefulcomputer software for the randomization procedure.

Please address correspondence to Charles P. Shimp,Department of Psychology, 380 South 1530 East Room502, University of Utah, Salt Lake City, Utah 84112-0251(e-mail: [email protected]).

seem to require exquisitely precise abilities toprocess dynamic visual stimuli. This line ofreasoning has led researchers to look for, andto find, neuroanatomical, electrophysiologi-cal, and behavioral structures and processesthat underlie or are correlated with move-ment perception in several avian species(Frost, Wylie, & Wang, 1994; Husband & Shi-mizu, 2001; Lea & Dittrich, 2001; Martinoya,Rivaud, & Bloch, 1983; Potts, 1984; Watana-be, 1991).

Behavioral explorations of the pigeon’s re-sponses to motion included attempts to trainpigeons to discriminate different stimulus ve-locities (Hodos, Smith, & Bonbright, 1975)and to train them to track moving targets vi-sually (Skinner, 1960). More recently, Em-merton (1986) showed that pigeons integrat-ed a moving dot pattern to form anintegrated stimulus, Neiworth and Rilling(1987) showed that pigeons formed an inter-nal representation of a moving stimulus andextrapolated its movement during a periodwhen it was occluded (also see Rilling &LeClair, 1989; Rilling, LeClair, & Warner,1993), McVean and Davieson (1989) showedthat a pigeon intercepted targets moving ona conveyor belt, and Dittrich and Lea (1993)and Dittrich, Lea, Barrett, and Gurr (1998)showed that pigeons learned to form abstractconcepts of naturalistic patterns of movementsuch as walking or flying. Even this incom-plete list of behavioral research shows that pi-geons do indeed have at least some of the

250 WALTER T. HERBRANSON et al.

specific abilities to perceive and respond tomovement that one would expect of thembased on their formidable natural flight char-acteristics.

What other abilities might pigeons have toprocess visual movement information? Onemight expect highly visual nonmammalianvertebrates, such as pigeons, to have evolvedboth to perceive motion and so that the spe-cific motion of an object facilitates the rec-ognition of what has moved. Specifically, wespeculated that a pigeon might categorizemoving targets purely on the basis of theirmovements, even when the targets provideno other basis for categorization, there aremany different target movements, and somemovements are highly ambiguous as to howthey should be categorized. Consider thatnatural moving stimuli presumably often be-long to multidimensional fuzzy categorieswithout either necessary or sufficient combi-nations of component features (Huber, 2001;Rosch & Mervis, 1975; Shimp, 1973). An ob-ject blowing about in the wind, for example,can move in many ways, with continuouslyvarying speeds and directions, and any localcombination of speed and direction onlyprobabilistically diagnoses its future move-ment. Also, the very same local combinationof speed and direction sometimes might oc-cur in the context of one overall pattern ofmovement, if, say, the object is a humming-bird, and sometimes in the context of a dif-ferent overall pattern, if the object is a drag-onfly. A pigeon presumably discriminatesdifferent categories of moving objects (e.g.,pigeons, hawks, leaves, insects, and pieces ofpaper) in part by their dynamic properties,including their speeds and directions. Al-though any one brief estimate of a movingobject’s speed and direction might not un-ambiguously diagnose whether it is a pigeonor a hawk, or a hummingbird or a dragonfly,several such estimates might diagnose with atleast better than chance levels which it is. Infact, it seems possible that if a flying object issufficiently far away and in poor light, its dy-namic flight characteristics could identify itbetter than its visual features like the size andshape of its wings. An especially interestingcase therefore obtains if there are no visualidentifying features at all between two objectsexcept for their patterns of movement. Insuch a case, no visual scrutiny of the static

object itself could specify its nature: Only itsmovement would identify it (Dittrich et al.,1998; Emmerton, 1990).

A method used in the literature on humancategorization can be adapted to facilitate theunderstanding of attentional processing ofthe kinds of ambiguous stimuli just described,in which there are two dimensions like speedand direction, each variable varies continu-ously over a great many values, and anyspeed–direction pair might diagnose eitherone or the other of two categories of movingobjects. The procedure with humans is calledthe randomization procedure and has been usedextensively to study the role of attention andmemory in categorization of multidimension-al stimuli (Ashby & Maddox, 1998). We willpresent this procedure in terms of a versionwe previously developed for use with pigeons.In that research (Herbranson, Fremouw, &Shimp, 1999), pigeons viewed two-dimension-al stimuli defined as rectangles varying inwidth and height. Stimuli were drawn fromtwo categories that overlapped, in the sensethat any rectangle could be sampled from ei-ther category. Most rectangles, however, weremore likely to be sampled from one categorythan from the other: Rectangles were usuallydiagnostic of the category from which theywere sampled. For example, in one condi-tion, rectangles that were taller than theywere wide were more likely to be sampledfrom one category, and rectangles wider thantall were more likely to be sampled from theother category, but any given rectangle couldbe sampled from either category. An on-linereal-time demonstration of this task is avail-able in Shimp, Herbranson, and Fremouw(2001).

Figure 1 is taken from Herbranson et al.(1999) and summarizes such a task. The leftpanel shows two ill-defined categories, A andB, in the form of two overlapping normal bi-variate distributions. The space over whichthe distributions are defined is typically re-ferred to as the stimulus space, in which eachpoint represents a particular two-dimensionalrectangle with width x and height y. The thirdcoordinate, z, is the likelihood with which therectangle will occur given a particular cate-gory. In Figure 1, as in the experiments de-scribed below, each variable in each distri-bution has the same variance, and thecovariance between variables in each distri-

251CATEGORIZING MOVING TARGETS

Fig. 1. Left: Bivariate normal distributions represent likelihoods with which rectangles in the experiment byHerbranson, Fremouw, and Shimp (1999) were sampled from either of two ill-defined (overlapping) categories, Aand B. A rectangle is represented in the stimulus space as a point with coordinates equal to the corresponding widthand height. A pigeon successively categorized individual two-dimensional stimuli, and a choice was reinforced if itcorresponded to the category, either A (left key) or B (right key), from which a stimulus was sampled. One arbitrarycontour of equal likelihood is shown for each category. Each contour consists of all points corresponding to rectanglesequally likely to be sampled from a category (taken from Figure 1 in Herbranson et al., 1999, p. 114). Right: arbitrarycontours of equal likelihood for each category and the corresponding linear optimal decision bound, x 5 y, accordingto which a rectangle should be categorized as an A or as a B, depending on whether the rectangle was taller thanwide or wider than tall, respectively. A real-time interactive demonstration of this task is available atwww.pigeon.psy.tufts.edu/avc/shimp.

bution is zero. The right panel in Figure 1shows two equal-likelihood contours, each ofwhich efficiently summarizes a bivariate nor-mal distribution by showing points corre-sponding to stimuli that are equally likely tooccur given a particular category. The rightpanel also shows the optimal decision bound,the line formed by the points correspondingto rectangles that are equally likely to occurgiven either category. In the task representedin Figure 1, as in all of the tasks describedbelow, the optimal decision bound is astraight line, according to which stimuli arecategorized optimally when a stimulus on oneside is categorized as belonging to CategoryA and a stimulus on the other side is cate-gorized as belonging to Category B (Ashby &Gott, 1988).

A pigeon categorizing stimuli over trialsproduces an empirical equivalent to the stim-ulus space shown in Figure 1. Each point insuch a stimulus space shows how an individ-ual pigeon categorized a particular stimulus.An empirical decision bound can be estimat-ed from the data points to summarize a pi-geon’s categorizations, and that estimated

bound can be compared to the optimalbound. See the Method section below for fur-ther details.

The present experiment adapted the ran-domization procedure to study how pigeonscategorize a target moving in any of manypossible directions at any of many possiblespeeds. We used a procedure virtually iden-tical to that of Herbranson et al. (1999), ex-cept that we replaced the two dimensions ofheight and width of static rectangles withspeed and direction of moving targets. Onemay assume that a pigeon can ‘‘selectively at-tend’’ to either or both dimensions, speedand direction, of an object’s movement. Wetherefore asked whether a pigeon can cate-gorize a target moving in any of various di-rections and at any of various speeds in termsof both its speed and direction combined, itsdirection alone, or its speed alone. The firsttask was arranged to determine the extent towhich a pigeon could make nearly optimalcategorizations of a moving target when op-timality required divided attention to bothspeed and direction and suitable integrationof information from both. Subsequent tasks


determined the extent to which a pigeoncould make nearly optimal categorzationswhen optimality required the pigeon to at-tend selectively to just one relevant dimen-sion and ignore the other, irrelevant, dimen-sion of a moving target.

METHOD

Subjects

Four experimentally naive male White Car-neau pigeons (Columba livia) were obtainedfrom the Palmetto Pigeon Plant (Sumter,SC). Each was maintained at approximately80% of its free-feeding weight, with supple-mental grain provided as needed in homecages after daily experimental sessions. Eachpigeon was housed individually in a standardpigeon cage with free access to water and grit,in a colony room with a 14:10 hr light/darkcycle. All experimental sessions took placeduring the light cycle at approximately thesame time 5 or 6 days per week. Pigeon 2died during Condition 2.

Apparatus

The four experimental chambers had in-ternal dimensions of 38 cm long by 34.5 cmwide by 50 cm high. Each had three clearplastic response keys (3.5 cm by 3.5 cm)mounted in a horizontal row within a clearPlexiglas viewing window (17 cm wide by 7cm high). The viewing window itself wasmounted in the front wall of the chamber,with its base 20 cm above the chamber floor.A computer monitor with a 15-in. screen waslocated 8 cm behind this front wall. Eachchamber was interfaced via digital input-out-put cards to a 90- or 100-mHz personal com-puter that controlled all experimental contin-gencies, presented stimuli, and recordeddata. A fan and white noise helped to maskextraneous sounds. A digital sound meterheld at approximately the position of a pi-geon’s head gave a reading (C scale) that var-ied across chambers from approximately 88to 93 dB, which was sufficient to mask most,if not all, sounds from outside the chamber.

Procedure

The general procedure involved displayinga small illuminated moving region (the ‘‘tar-get’’) on a computer screen for a brief period

and then requiring the pigeon to categorizethe target’s movement as belonging to one orthe other of two bivariate normal distribu-tions. (An interactive demonstration of allfour categorization tasks described below isavailable in Malloy et al., 1991.)

Pretraining. Pigeons were pretrained in ses-sions consisting successively of habituation tothe chamber, magazine training, and auto-shaping to peck consistently on each key.Stimuli used for autoshaping were blocks ofvarious colors (2.4 cm square; green, red, orblue) appearing one at a time behind thethree response keys.

Stimuli. Stimuli consisted of a small whitetarget, approximately circular (ø1 cm diam-eter), moving at a particular speed in a par-ticular direction. The white target appearedagainst the background of an otherwise blankscreen. At the beginning of a trial, the targetwas initially displayed at a random locationwithin an area (3.8 cm by 3.8 cm square) be-hind the center response key. A single peckto the center key after the target was pre-sented initiated movement in a straight lineand at a constant speed. The direction andspeed of movement varied across trials. Speedvaried from a low of about 0.10 cm/s to ahigh of about 1.0 cm/s. Direction of move-ment varied approximately 89.58 from verticalin both directions. That is, targets couldmove almost directly to the left or to theright, or in any intermediate upwards direc-tion. The range of each variable was madeabout as large as the apparatus would permit,and no effort was made to scale the two di-mensions so that numerically equal varianceswere psychologically equal. Accordingly, anyresulting consequences of one dimension’sbeing ‘‘easier’’ than the other are unknown.

Both dimensions varied almost continuous-ly: Possible speeds and directions in these ex-periments were limited chiefly by the pixels,the size of the monitors, processor speed, andthe software used (Turbo Pascal 7.0). (Unlikethe experiments in Herbranson et al., 1999,in which stimuli were presented in charactermode, here we used graphics mode.) In ad-dition, the position of the transparent re-sponse keys imposed limitations on possibleor appropriate target locations, because it wasnot desirable for targets to overrun the re-sponse keys. Accordingly, the target size andvector properties described above were cho-


sen so that targets could not be producedthat would overrun response keys or travel offthe edge of the screen.

On each trial, a stimulus was randomly cho-sen from one of the two equally likely cate-gories. Each category had a correspondingtwo-dimensional bivariate normal distribu-tion, with the two dimensions being speedand direction. The bivariate normal distri-butions were approximate, in the sense thatspeeds and directions were drawn frombounded stimulus spaces, as shown below inFigures 2 to 10. If a speed or direction wassampled that did not fall within the stimulusspace, it was rejected, and sampling contin-ued until an appropriate stimulus was sam-pled. Any possible speed or direction withina specified stimulus space could be selectedfrom either category, so that the categories inthis sense were ‘‘ill defined’’ and the task wasa probabilistic discrimination (Shimp, 1973).Most speed–direction pairs, however, weremore likely to be generated by one categorythan the other so that for most vectors, oneresponse was more likely to be reinforcedthan the other.

Once a target’s speed and direction wereselected for a particular trial, an additionalcontrol procedure was implemented for thefollowing reason. Speed is the ratio of dis-tance to time, so categorization responsesbased on distance or time, rather than speed,could lead to better than chance accuracy: Ifthe target were always presented for the sameduration, given a selected speed, then thelength of the line over which the target trav-eled could serve as a cue for better thanchance accuracy. Alternatively, if the targetwere always presented for the same length ofline, again given a particular selected speed,then time of presentation could serve simi-larly. Therefore, after the target’s speed anddirection were selected, the computer ran-domly chose to present the stimulus for ei-ther a specified duration or a specified dis-tance. Finally, depending on whether thiscontrol procedure selected duration orlength of line, the computer randomly choseeither a time varying from 2 to 3 s or a lengthvarying from 1.9 to 3.8 cm. In summary, thiscontrol procedure was designed to encouragecategorization of stimuli based on speed anddirection, rather than on either the distancethe target moved or the duration it was visible

(see McKee, 1981, for further discussion re-lated to these control issues). It is importantin the face of all these procedural details,however, not to lose sight of the fact that, re-gardless of whether control was by distance,speed, time, or some combination of thesevariables, all these possibilities are indepen-dent of the other dimension, direction.

Trial organization. Each of a session’s 80 tri-als consisted sequentially of an orienting cue,presentation of the moving target, a catego-rization response, either reinforcement or acorrection procedure, and an intertrial inter-val.

Each trial began with an orienting cue (a2.4-cm green square) presented directly be-hind the center key. The first center-key peckto occur after 1 s turned off this orienting cueand turned on a white circular target (1 cmdiameter) presented randomly within a cir-cumscribed region (see above) behind thecenter key. The next peck to the center keystarted the target moving with a speed anddirection chosen as described above. The tar-get stopped moving after it either traveled arandomly chosen distance or had traveled fora randomly chosen length of time, as de-scribed above. Once the moving targetreached its terminal distance or duration, itdisappeared from the screen and coloredsquares (2.4 cm; red and blue on left andright, respectively) appeared directly behindthe two side keys. A pigeon then categorizedthe moving target that had been presented asan exemplar of Category A or Category B. Aleft response was reinforced if the stimuluswas generated by Category A, and a right re-sponse was reinforced if the stimulus was gen-erated by Category B.

Reinforcement. If a choice corresponded tothe category from which the stimulus hadbeen sampled, mixed grain was presented ina hopper located directly beneath the win-dow. Hopper presentation time varied acrosspigeons from 1.7 to 2.3 s to maintain individ-ual deprivation levels accurately. Following re-inforcement, there was a 5-s intertrial inter-val, during which the monitor was blank.

Correction procedure. If a choice did not cor-respond to the category from which a stim-ulus was sampled, a 10-s correction intervalbegan. This interval was signaled by thehouselight flashing on and off every 0.5 s. Atrial was then repeated, with the same moving


target. Any subsequent errors caused the cor-rection procedure to continue to recycle,with the same stimulus, until the correct re-sponse occurred. Only the initial choice wasrecorded and included in data analysis.

Experimental Conditions

Conditions varied in terms of the stimulusdimensions upon which reinforcement de-pended: both speed and direction, speedalone while direction varied randomly, or di-rection alone while speed varied randomly.Conditions lasted 40 to 48 days and were end-ed mostly due to a combination of informalexamination of the stability of overall accu-racy of categorization and experimenter con-venience. Best fitting decision bounds wereestimated only after the completion of the en-tire experiment. In all five conditions, therewas a 5% overlap between Categories A andB, so that optimal categorization would leadto reinforcement on an average of 95% of thetrials.

Speed and direction were varied acrossconditions in terms of arbitrary units in theTurbo Pascal programming language, with afraction of a degree corresponding to direc-tion and repetitions of a fixed delay corre-sponding to speed. An experimenter estimat-ed the actual speeds and directions, andthereby estimated how to translate Pascal pa-rameters to actual speeds and directions, byrepeatedly measuring the target’s movementon a pigeon’s monitor for a given speed anddirection, then for a different speed–direc-tion pair, and so on. These estimates were dif-ficult to make, however, because the monitorsurface was not completely flat, and the targetwas not a perfect point. A pigeon, of course,presumably experienced similar difficulties inestimating speed and direction.

For the sake of completeness, experimentalconditions are described in terms of both thearbitrary software values and the correspond-ing estimated speeds and directions. Smallersoftware numbers for speeds and directionsin the figures below refer to faster speeds andto movement to the right, respectively. An ap-proximate correspondence between speedand arbitrary software units is provided by thefollowing relation: Target speed (in centime-ters per second) was estimated to equal 1.052 0.000685 3 s, where s was the speed, insoftware units, drawn from a normal distri-

bution. Direction varied in increments of0.56928, and target direction in degrees wasequal to 2447.567 1 0.570149 3 0.5692d,where d was the direction, in software units,drawn from a normal distribution.

Condition 1: Divided attention across both speedand direction. In Condition 1, the speed of aCategory A stimulus was drawn from a nor-mal distribution with a mean of 1,050 (0.33cm/s) and a standard deviation of 180 (0.12cm/s). For Category B, speed was drawn froma normal distribution with a mean of 550(0.67 cm/s) and standard deviation of 180(0.12 cm/s). Direction for a Category A stim-ulus was drawn from a normal distributionhaving a mean of 733 (29.68 right of vertical)and a standard deviation of 60 (34.208), andfor a Category B stimulus from a normal dis-tribution with a mean of 837 (229.68 left ofvertical) and standard deviation of 60(34.208). The top panel of Figure 2 shows thelocations of the two prototypes (categorymeans) in the stimulus space, equal-likeli-hood contours at one and two standard de-viations, and the optimal decision bound.The optimal decision bound in Figure 2 rep-resents the intersection of the two stimulusdistributions, and indicates the optimal strat-egy for responding, which would result in95% of choice responses being reinforced.Optimal performance on this task required apigeon to attend to both speed and direction,and to integrate them in a specific way. Fail-ure to attend to both dimensions or to com-bine the information from each correctlywould result in suboptimal performance.

The bottom panel of Figure 2 shows thetwo prototypes corresponding to CategoriesA and B, with direction indicated by the di-rection of the arrows and speed indicated bythe length of the arrows, with length directlyproportional to speed. The bottom panel ofFigure 2 shows that targets moving more slow-ly to the upper right were usually exemplarsof Category A, and targets moving morequickly to the upper left were usually exem-plars of Category B. Optimal behavior re-quired a pigeon to categorize most targetsmoving to the right as belonging to CategoryA. Because Category A was associated with theleft key, pigeons should have pecked the leftkey after viewing targets moving to the right.

Condition 2: Divided attention across both speedand direction. In Condition 2, the association


Fig. 2. Left: the stimulus space, two equal likelihoodcontours at one and two standard deviations from theaverage of each category, and the optimal bound forCondition 1. Optimal performance on the task requireda pigeon to divide attention between both dimensions,speed and direction, and to integrate the informationfrom both. Scale units are arbitrary Turbo Pascal softwarenumbers according to which smaller numbers refer tofaster speeds and to movement to the right (see text fordetails). Right: the prototypes corresponding to Catego-ries A and B, in which direction of the target is repre-sented by the direction of an arrow and speed is repre-sented by the length of the line (faster targetscorrespond to longer lines).

between categories and vectors was changed,but integration of information from both di-mensions was still required for optimal per-formance. The speed of a Category A stimu-lus was drawn from a normal distributionwith a mean of 1,050 (0.33 cm/s) and a stan-dard deviation of 180. For Category B, speedwas drawn from a normal distribution with amean of 550 (0.67 cm/s) and standard devi-ation of 180. Thus, the speed dimension ofvectors for Categories A and B remained thesame. Direction, however, was reversed. Thedirection of a Category A stimulus was drawnfrom a normal distribution having a mean of837 (229.68 left of vertical) and a standarddeviation of 60, and the direction of a Cate-gory B stimulus was drawn from a normal dis-tribution with a mean of 733 (29.68 right ofvertical) and standard deviation of 60. Figure3 corresponds to Figure 2. The top panelagain shows the locations of the two proto-types in the stimulus space, two equal-likeli-hood contours at one and two standard de-viations, and the optimal bound. The bottompanel again shows the two prototypes, withtargets corresponding to Category A now usu-ally moving more slowly to the upper left in-stead of to the upper right and targets cor-responding to Category B usually movingmore quickly to the upper right instead of tothe upper left. The optimal decision boundin Figure 3 represents the intersection of thetwo stimulus distributions, and indicates theoptimal strategy for responding. To performoptimally, a pigeon had to continue to attendto both speed and direction, but now had tocombine the corresponding information dif-ferently. Integration alone, however, was notsufficient: Continuing to integrate speed anddirection as in Condition 1 would have re-sulted in chance performance.

Condition 3: Selective attention to direction, withspeed irrelevant. In Condition 3, the speed ofboth a Category A stimulus and a Category Bstimulus was drawn from a normal distribu-tion with a mean of 800 (0.50 cm/s) and astandard deviation of 180. Speed was there-fore irrelevant. Direction was drawn from anormal distribution having a mean of 687(55.98 right of vertical) and a standard devi-ation of 60 for Category A stimuli, and direc-tions of Category B stimuli were drawn froma distribution with a mean of 883 (255.98 leftof vertical) and a standard deviation of 60.


Fig. 3. Top: the stimulus space, two equal likelihoodcontours at one and two standard deviations from theaverage of each category, and the optimal bound forCondition 2. Optimal performance on the task requireda pigeon to divide attention between both dimensionsand to integrate the information from both. Scale unitsare arbitrary numbers according to which smaller num-bers refer to faster speeds and to movement to the right(see text for details). Bottom: the prototypes correspond-ing to Categories A and B, in which direction of the tar-get is represented by the direction of the arrow andspeed is represented by the length of the line (faster tar-gets correspond to longer lines).

Figure 4 parallels Figures 2 and 3 and showsthe prototypes, stimulus space, and optimaldecision bound. The optimal decision boundin Figure 4 indicates the optimal strategy forresponding, which again would result in 95%of the choice responses being reinforced.The critical feature of Condition 3 was thatspeed was no longer diagnostic of categorymembership. To perform optimally on thistask, a pigeon had to attend to directionalone.

Condition 4: Selective attention to speed, withdirection irrelevant. In Condition 4, the speedof a Category A stimulus was drawn from anormal distribution with a mean of 1,094(0.31 cm/s) and a standard deviation of 180.For Category B, speed was drawn from a nor-mal distribution with a mean of 506 (0.70cm/s) and a standard deviation of 180. Di-rection was drawn for both categories from anormal distribution having a mean of 785(vertical) and a standard deviation of 60. Fig-ure 5 illustrates the task in the same manneras did Figures 2 to 4 for Conditions 1 to 3.The critical feature of Condition 4 was thatdirection was no longer diagnostic of cate-gory membership. Thus, to perform optimal-ly on this task, a pigeon had to attend tospeed alone and to ignore direction.

Condition 5: Replication of integration task ofCondition 1. Condition 5 replicated the taskin Condition 1, so Figure 2 summarizes Con-dition 5 as well as Condition 1.

Method of estimating decision bounds. Best fit-ting estimated decision bounds were ob-tained in the following way. A decision boundwas estimated for each individual pigeon. Foreach condition, each of a pigeon’s categori-zations over the last 10 days of that conditionwas compared successively to the predictionsmade by a series of possible linear decisionbounds passing through the stimulus space.A comparison was made for each of thesepossible decision bounds. Specifically, foreach possible bound, the number of trials onwhich the bound predicted the same catego-rization as was observed from an individualpigeon was calculated. The exhaustive searchused a step size equal to 0.005 for slope and0.4 for intercept. The best fitting bound wasdefined as that which accounted for more in-dividual categorizations than any otherbound. It was, that is, the bound that maxi-mized the number of individual categoriza-


Fig. 4. Top: The stimulus space, two equal likelihoodcontours at one and two standard deviations from theaverage of each category, and the optimal bound forCondition 3. Optimal performance on the task requireda pigeon to attend selectively to direction and to ignorespeed. Scale units are arbitrary numbers according towhich smaller numbers refer to faster speeds and tomovement to the right (see text for details). Bottom: theprototypes corresponding to Categories A and B, inwhich direction is represented by the direction of thearrow and speed is represented by the length of the ar-row.

Fig. 5. Top: the stimulus space, two equal likelihoodcontours at one and two standard deviations from theaverage of each category, and the optimal bound forCondition 4. Optimal performance on the task requireda pigeon to attend selectively to speed and to ignore di-rection. Scale units are arbitrary numbers according towhich smaller numbers refer to faster speeds and tomovement to the right (see text for details). Bottom: theprototypes corresponding to Categories A and B, inwhich direction is represented by the direction of thearrow and speed is represented by the length of the ar-row.


tions of stimuli that were the same as a pi-geon’s categorizations of those same stimuli.Most searches produced more than one bestfitting solution: The estimated decisionbound was usually not unique. Table 1 showsthe most extreme values of the slope and in-tercept for the estimates tied for the largestnumber of categorizations accounted for (seealso Herbranson et al., 1999).

A decision bound estimated in this way,even if it approximates the optimal decisionbound, does not necessarily either maximizethe number of categorizations accounted foror accurately reflect all possible decisionbounds. Consider the following hypotheticalexample of performance in an integrationtask in which optimal performance requiresattention to, and appropriate integration of,information from both dimensions. A pigeonmight, on a random trial-by-trial basis, selec-tively attend to either speed or direction, butnot both simultaneously. The resulting esti-mated decision bound might approximate anoptimal decision bound and thereby incor-rectly imply integration of information fromboth dimensions. The estimated boundwould not accurately reflect the actual deci-sion processes, and in general, estimatedbounds do not, by themselves, diagnose un-derlying decision processes. To diagnose ac-tual processes more accurately, examinationof other features of behavior, such as theoverall response pattern in the stimulus spaceand overall accuracy, is essential. For exam-ple, in the case just described, the estimateddecision bound would be flanked in the stim-ulus space by a different pattern of responsesfrom that which corresponds to respondingbased on information integration, and overallaccuracy could not consistently outperform asuboptimal selective attention strategy. In thefollowing, all three forms of results—the es-timated decision bound, the stimulus space,and accuracy—are therefore presented.

RESULTS

Condition 1: Divided AttentionAcross Both Speed and Direction

Table 1 gives the critical individual-pigeonand group-average numerical data averagedover the last 10 days of the condition. Pigeon1’s stimulus space is displayed in the left pan-

el of Figure 6 because Pigeon 1, along withPigeon 5, had an accuracy intermediate be-tween the best (Pigeon 2) and the worst (Pi-geon 4), and because Pigeon 5’s results arepresented below for other conditions. Figure6 gives an estimate of Pigeon 1’s decisionbound: It shows the median of the five equal-ly best fitting estimated bounds. (Table 1gives the numerical values of the correspond-ing slope and y intercept.) Pigeon 1 catego-rized most moving targets generally in accordwith the optimal decision bound. The rightpanel of Figure 6 shows each individual pi-geon’s median estimated decision bound andalso shows the group average of those indi-vidual medians. Pigeon 1’s behavior was notatypical: Estimated and optimal bounds weregenerally similar, especially so for the groupaverage.

Averaged over pigeons, 81.1% of categori-zations were to the optimal key (Table 1).The theoretical optimal decision bound, as-suming appropriate integration of informa-tion from both speed and direction, account-ed for a higher percentage of categorizationsthan did either optimal decision bound as-suming selective attention purely to eitherspeed or direction alone (Table 1). Pigeonstherefore to some degree integrated infor-mation across dimensions, as required by theoptimal decision bound for the task in Con-dition 1. The slopes of the best fitting esti-mated decision bounds ranged for individualpigeons from approximately 0.92 to 4.76,with an average of 2.07, which correspondedwell to the optimal value of 1.87. The y inter-cepts ranged from 246.97 to 22,948.57, withan average of 2908.08, which roughly ap-proximated the optimal value of 2669.52.

Table 1 also shows two additional estimatedbest fitting decision bounds, best fitting speedonly and best fitting direction only. The bestfitting speed-only bound was obtained by ex-amining the fit of all possible bounds thatwere perpendicular to the speed axis andthen selecting the bound that maximized thenumber of categorizations that were the sameas a pigeon’s. The best fitting direction-onlybound was obtained in the correspondingmanner. The best fitting integration bound,described previously, fit better for each pi-geon than either of these selective attentionbounds. This comparison provides additionalsupport for the view that in Condition 1, the


Table 1

Accuracy and best fitting decision bounds.

Condition Pigeon 1 Pigeon 2 Pigeon 4 Pigeon 5 Average

1: Optimal decision bound, Y 5 1.87X 2 669.52% of choices that were reinforced% of choices that were optimal% of choices accounted for by speed 5 800% of choices accounted for by direction 5 785% of choices accounted for by best fitting bound

best fitting speed onlybest fitting direction only

79.380.077.773.881.678.474.6

82.884.383.273.986.384.474.6

77.177.676.271.581.080.373.1

80.382.381.173.883.181.974.0

79.981.179.673.383.081.374.1

Number of best fitting boundsSlope of median best fitting boundIntercept of median best fitting bound

54.76

22,948.57

501.50

2543.76

230.92

246.97

441.09

293.04

30.52.07

2908.08



70.570.371.665.576.675.571.4

81.884.081.475.585.883.376.4

76.678.077.769.281.582.373.5

76.377.476.970.181.380.473.8


623.39

3,026.44

721.18

1,778.82

2120.76

1,166.41

11.321.78

1,990.56

3: Optimal decision bound, X 5 785% of choices that were reinforced% of choices that were optimal% of choices accounted for by speed 5 800% of choices accounted for by direction 5 785% of choices accounted for by best fitting bound


67.666.550.866.578.571.378.1

82.583.652.483.684.154.184.0

76.876.154.276.177.655.477.4

75.675.452.575.480.160.379.8

X 5 AY 1 BNumber of best fitting boundsSlope of median best fitting boundIntercept of median best fitting bound

220.03921.60

720.06851.20

120.02816.00

3.320.04862.93

4: Optimal decision bound, Y 5 800% of choices that were reinforced% of choices that were optimal% of choices accounted for by speed 5 800% of choices accounted for by direction 5 785% of choices accounted for by best fitting bound


73.974.574.551.875.674.857.5

80.581.581.550.885.485.460.3

79.480.080.053.381.380.954.4

77.978.778.752.080.880.457.4


921.02

1,588.80

260.84

216.81

121.12

1,733.03

12.020.43

1,101.67



72.072.672.365.474.172.866.5

82.581.980.074.486.684.976.8

82.182.081.474.384.181.975.1

78.978.877.971.481.679.972.8


192.78

21,244.44

71.85

2808.89

401.09

2318.26

22.01.91

790.52


Fig. 6. Left: the obtained stimulus space for Pigeon 1 in Condition 1, with filled and open circles correspondingto individual responses categorizing stimuli as belonging to Category A or B, respectively, the dashed line showingthe estimated decision bound for Pigeon 1, and the solid line showing the optimal bound. Right: all four individual-pigeon estimated decision bounds, the average decision bound (dashed line), and the optimal decision bound (solidline). Scale units are arbitrary numbers according to which smaller numbers refer to faster speeds and to movementto the right (see text for details).

pigeons were appropriately integrating infor-mation from both speed and direction.

Condition 2: Divided AttentionAcross Both Speed and Direction

Figure 7 shows the stimulus space and es-timated decision bound for Pigeon 5, whichwas intermediate in terms of accuracy of cat-egorizations. (Recall that Pigeon 2 died be-fore the end of Condition 2.) All estimatedslopes reversed sign from positive in Condi-tion 1 to negative in Condition 2, in accordwith the corresponding change in optimal de-cision bound slopes from 1.87 to 21.87 (Ta-ble 1). Individual slopes varied from 20.76 to23.39, with the average of 21.78 closely cor-responding to the optimal value of 21.87. In-dividual y intercepts varied from 1,166.41 to3,026.44, with an average of 1,990.56 com-pared to the optimal 2,269.52. Like theslopes, the y intercepts changed drasticallyand thereby moved toward the correspond-ing value in the changed optimal bound. Ingeneral, the pigeons continued to integrateinformation from both dimensions, as re-quired by the task, and also now appropriate-

ly associated movement to the left with Cat-egory A rather than with Category B.

Table 1 shows estimated best fitting speed-only and direction-only decision bounds. Thebest fitting integration bound fit better thaneither of these selective attention bounds forPigeons 1 and 4, but for Pigeon 5, there wasa best fitting speed-only bound that describedthe data better than the integration bound.For Pigeon 5, that is, there was a selective at-tention criterion in terms of speed, slightlydifferent from the average speed, that led toa slightly better fit than did any estimated di-vided attention bound. Therefore, in thissense Pigeon 5 selectively attended to speed.Otherwise, Condition 2, like Condition 1,generally showed appropriate integration ofinformation from both speed and direction.

Condition 3: Selective Attention toDirection, with Speed Irrelevant

Pigeon 5 was again intermediate in termsof accuracy so we again display, in Figure 8,that pigeon’s stimulus space and median es-timated decision bounds. All 3 pigeons pro-duced decision bounds that approximated


Fig. 7. Left: the obtained stimulus space for Pigeon 5 in Condition 2, with open and filled circles correspondingto individual responses categorizing stimuli as belonging to Category A or B, respectively, the dashed line showingthe estimated decision bound for Pigeon 5, and the solid line showing the optimal bound. Right: all three individual-pigeon estimated decision bounds, the average decision bound (dashed line), and the optimal decision bound (solidline). Scale units are arbitrary numbers according to which smaller numbers refer to faster speeds and to movementto the right (see text for details).

Fig. 8. Left: the obtained stimulus space for Pigeon 5 in Condition 3, with open and filled circles correspondingto individual responses categorizing stimuli as belonging to Category A or B, respectively, the dashed line showingthe estimated decision bound for Pigeon 5, and the solid line showing the optimal bound. Right: all three individual-pigeon estimated decision bounds, the average decision bound, and the optimal decision bound. Scale units arearbitrary numbers according to which smaller numbers refer to faster speeds and to movement to the right (see textfor details).


Fig. 9. Left: the obtained stimulus space for Pigeon 5 in Condition 4, with filled and open circles correspondingto individual responses categorizing stimuli as belonging to Category A or B, respectively, the dashed line showingthe estimated decision bound for Pigeon 5, and the solid line showing the optimal bound. Right: all three individual-pigeon estimated decision bounds, the average decision bound, and the optimal decision bound. Scale units arearbitrary numbers according to which smaller numbers refer to faster speeds and to movement to the right (see textfor details).

the vertical optimal bound (Table 1): In thissense, when the task demanded it, pigeonswere more nearly able to attend exclusivelyto direction and to ignore speed. The group-average estimated slope of 20.04 comparesfavorably to the optimal 0, and the group-av-erage estimated y intercept of 862.93 com-pares reasonably well to the optimal 785. (Inthis condition, we estimated speed as a func-tion of direction to avoid the problem of aninfinitely large slope.) It should be noted,however, that for none of the pigeons was theselective attention bound, even with directionfree to vary, quite as good a fit as the esti-mated integration bound. Therefore, to someresidual extent, one can see in this sense thatall 3 pigeons did not completely selectivelyattend to direction. In this sense, Condition3 was somewhat less successful in generatingselective attention purely to direction thanConditions 1 and 2 were in generating divid-ed attention to both dimensions.

Condition 4: Selective Attention toSpeed, with Direction Irrelevant

Pigeon 5 was again intermediate in termsof accuracy, and we therefore again chose

that pigeon’s estimated decision bound asrepresentative. Figure 9 shows a substantialchange in estimated decision bound fromthat in Figure 8, corresponding to the changein tasks. The average slope changed fromroughly vertical to roughly horizontal: The av-erage slope of the individually estimated de-cision bounds was 20.43, which is quite closeto the optimal value of 0. The average y in-tercept was 1,101.67, compared to the opti-mal 800.

In Condition 4, as in Condition 3, estimat-ed decision bounds assuming divided atten-tion between speed and direction fit catego-rizations as well as or better than theestimated decision bound based on selectiveattention to the relevant dimension, directionin Condition 3 and speed in Condition 4 (Ta-ble 1).

Condition 5: Replication of IntegrationTask of Condition 1

Comparisons of Figures 6 and 10 and en-tries in Table 1 for the results of Condition 1and corresponding entries for the replicationin Condition 5 show considerable consistencyacross pigeons in accuracy and estimated de-


Fig. 10. Left: the obtained stimulus space for Pigeon 4 in Condition 5, with filled and open circles correspondingto individual responses categorizing stimuli as belonging to Category A or B, respectively, the dashed line showingthe estimated decision bound for Pigeon 5, and the solid line showing the optimal bound. Right: all three individual-pigeon estimated decision bounds, the average decision bound, and the optimal decision bound. Condition 5 rep-licated Condition 1. Scale units are arbitrary numbers according to which smaller numbers refer to faster speeds andto movement to the right (see text for details).

cision bounds; the original integration taskwas well replicated. The pigeons, after revers-ing the relation between vectors and catego-ries in Condition 2, and after selectively at-tending mostly to either speed or directionalone in Conditions 3 and 4, once again di-vided attention between both speed and di-rection as they did in Condition 1.

Pigeons in Condition 5 were particularlysuccessful in integrating information fromboth dimensions, rather than selectively at-tending to just one dimension; in every case,the estimated integration bound fit betterthan any estimated selective attention bound(Table 1).

Overall Optimality

Figure 11 summarizes the degree to whichaverage performance across the five selectiveand divided attention conditions conformedto the optimal decision bounds. Figure 11shows the slope and y intercept of averageestimated decision bounds as a function ofthe slope and intercept of the optimal deci-sion bound, respectively. There was a remark-able correspondence between the optimalvalues and average performance. The close fit

between obtained and optimal performanceis remarkable because (a) the slopes and yintercepts of the optimal bounds were knownbefore the experiment began (the optimalbound had no free estimated parameter) and(b) the estimated bounds were not obtainedsimply by varying a parameter within thesame task, say, for example, by varying thebase rate of one of the two categories withinan experimental condition, but were ob-tained instead over tasks involving important-ly different attentional mechanisms, dividedor selective attention, and in one case (Con-dition 2) involved reversing the relation be-tween vectors and categories. The data in Fig-ure 11 therefore imply not only that pigeons,on the average, can be remarkably optimal incategorizing moving targets in terms of speedand direction, but that they can remain near-ly optimal even when important attentionalrequirements of the categorization taskchange.

DISCUSSION

We asked if pigeons can categorize movingtargets on the basis of movement, without any


Fig. 11. The average of the slopes of the individually estimated decision bounds as a function of the optimal slope(left) and the average of the y intercepts of the individually estimated decision bounds as a function of the optimaly intercept (right).

other visual identifying characteristics of ex-emplars of two categories. Not only was theanswer in this task affirmative, but in addi-tion, pigeons categorized moving targetswhile attending to either speed or direction,or both, as the task demanded. In this task,pigeons switched among attentional strate-gies to a nearly optimal degree, by which wemean the following. Slopes and average y in-tercepts of linear decision bounds estimatedfor individual pigeons suggest that they adaptalmost optimally, averaged across pigeons, tochanging attentional demands when they cat-egorize exemplars of two ill-defined two-di-mensional categories involving the same tar-get moving in different directions and atdifferent speeds. When optimal performancedemanded selective attention to a single di-mension, either speed or direction, estimateddecision bounds reflected correspondinglypowerful control, although not always exclu-sive control, by the appropriate dimension,and when optimal performance demandeddivided attention between both dimensions,estimated decision bounds reflected corre-sponding appropriate integration of infor-mation from both dimensions. Previous re-search with various types of static stimuli hasshown that pigeons attend either to one orthe other, or both, of two dimensions, suchas color, line tilt, and tone frequency(Blough, 1972; Chase & Heinemann, 1972,2001; Herbranson et al., 1999; Riley & Brown,

1991; Shimp et al., 2001). More particularly,Blough, Chase and Heinemann, and Her-branson et al. discussed theoretical decisionbounds in two-dimensional stimulus spaces asa means of identifying control by individualdimensions or combined control by both.The present results extend these previousfindings to show that pigeons similarly eitherselectively attend to just one of two features(speed or direction) of a moving stimulus ordivide attention across both in a more holisticmanner, depending on task demands.

Issues in the Analysis of Optimality

Few issues are more difficult or controver-sial than general claims to the effect that be-havior is, in some sense, ‘‘optimal’’ (Gould& Lewontin, 1979; Herbranson et al., 1999).We therefore hasten to note that it wouldnot yet be safe to assume that the presentresults, showing that average categorizationtracked attentional demands to an almostoptimal degree, will necessarily generalize toother experimental settings. Many experi-mental variables might affect categorizationperformance, some of which might improveperformance and many of which might de-grade it. The generality of the present re-sults, therefore, will depend on the outcomeof much additional research, giving carefulattention to the issues described below.

First, although the present results showthat pigeons changed how they attended to


speed and direction as the task demanded,and although Figure 11 shows an almost as-tonishing degree of optimal categorization,on the average, it should be recalled that thetwo selective attention conditions (Condi-tions 3 and 4) also showed some evidencethat pigeons continued to divide attention tosome extent instead of attending exclusivelyto the relevant dimension. In general, esti-mated bounds fit the categorization dataslightly better in divided attention tasks thanin selective attention tasks. This difference issmall, however, and obviously more researchwill be required, with speed and direction var-ied in many different ways, to determine theextent to which pigeons are better able to di-vide attention across speed and directionthan to selectively attend to either speed ordirection.

Second, the present task required pigeonsto shift attention only across conditions last-ing several days. This task therefore did notnecessarily involve the dynamic shifts of atten-tion that might occur when stimuli and atten-tion interact in a real-time manner, as theydid, for example, in Skinner (1960). Hownearly optimally pigeons can shift attention ina more dynamic situation remains an openquestion.

Third, when pigeons peck, they often brief-ly close their eyes (Smith, 1974), so theymight not observe the initial moments of tra-jectories of moving stimuli to which they arerequired to peck. The overall performance inthe present experiment suggests any such ef-fect here was negligible, but one can imagineother categories with much shorter stimulusdurations or nonlinear trajectories for whichconsequences of the pigeon’s natural peckingtopography might prevent the high level ofperformance obtained here, provided a keypeck is required to initiate a target’s move-ment.

Fourth, categorization of a target’s move-ment over some nontrivial period of timeseems automatically to involve working mem-ory for the target’s recent trajectory, with abuilt-in retention interval for any given partof a target’s trajectory equal to the remainingexposure duration preceding the categoriza-tion response. In the present experiment, be-cause both target speed and direction withina trial were held constant, this might not haveimposed as serious a problem as it might were

speed and direction to vary in complex wayswithin a single observation period. Indeed, inour experiment, because both speed and di-rection were constant within a trial, longerviewing times presumably both degraded cat-egorization, due to the greater memory loadfor earlier parts of longer trajectories, and fa-cilitated categorization, because workingmemory performance generally improveswith viewing time, as in delayed matching-to-sample tasks (Maki, Moe, & Bierley, 1977). Inshort, the present experiment involved onlylinear trajectories, and it is an open questionfor future research to determine what kindsof more complex trajectories can also serveas the basis for nearly optimal categorylearning (see Rilling & Neiworth, 1987, foradditional discussion of relations betweenmovement discrimination and short-termmemory).

Fifth, consider the role of the initial spatiallocation of the target. In the present experi-ment, it seems unlikely that the pigeons usedspatial locations on the monitor, rather thanspeed and direction, as cues upon which tobase their categorizations. There are at leasttwo reasons why this possibility seems unlike-ly: The target initially appeared on differenttrials in different locations on the monitor,and the target moved on different trials overmany different regions of the screen, so thatany given location or region would have beenunlikely to have served as a useful discrimi-native cue. In general, however, literature onavian spatial attention (Shimp & Friedrich,1993) and on attentional switches between lo-cal and global levels of perceptual analysis(Cavoto & Cook, 2000; Fremouw, Herbran-son, & Shimp, 1998) suggest the way in whichattention is directed in advance of target pre-sentation to spatial regions and levels of anal-ysis relevant to target observation may be ex-pected to affect discrimination of movingtargets. Thus, manipulations that increase ordecrease spatial attention to a region of thescreen in which a target is to appear mightimprove or degrade categorization, respec-tively.

Sixth, pigeons have two visual systems, afrontal system and a lateral system (Husband& Shimizu, 2001; Martinoya et al., 1983), withthe frontal system more adapted to a pigeon’seating strategies, involving small, static stim-uli, and the lateral system more adapted to


stimuli that are peripheral and dynamic. It isunclear what roles these systems played in thepresent experiment. Given the random vari-ations in initial target location, and given thevariability in duration and length of the stim-uli and the speed of the target, it seems un-likely that a pigeon was always capable ofmoving its head to optimize viewing. It is notunlikely that a pigeon might use differentviewing strategies for different stimuli that ap-pear in different locations and have differentspeeds and viewing times. It is an interestingquestion for future research to determinewhat roles the two systems play in the discrim-ination of different kinds of dynamic stimuli,and how pigeons use head movements to af-fect their discriminative capacities. It seemslikely that how any given task encourages theuse of different visual systems will affect ac-curacy of categorization of moving targets(Husband & Shimizu, 2001; Lea & Dittrich,2001). Teuber (1960) noted that human mo-tion thresholds ‘‘obviously vary with size andpattern of moving stimulus, with pattern ofsurround, with illumination and with numer-ous other factors’’ (p. 1642). Presumably, thekind of categorization required in the presentexperiment depends on all these variables;therefore, the conditions under which thenearly optimal categorization obtained herewill be obtained remain to be determined.

Seventh, the ‘‘oblique effect’’ (e.g., Donis,1999), according to which pigeons discrimi-nate horizontal and vertical lines better thanthey do oblique lines, might affect categori-zation accuracy in experiments of the presenttype. That is, depending on whether fre-quently occurring stimuli involve trajectorieson major axes or at oblique orientations, pi-geons might perform more or less accurately.

Eighth, how fast a stimulus must move fora pigeon to be able to detect its motion is notyet known. Early evidence suggested that thepigeon’s velocity threshold is relatively poorcompared to the human’s (Hodos et al.,1975; McKee, 1981), but more recent evi-dence suggests that the pigeon’s thresholdfor movement may be better than originallyestimated (Lea & Dittrich, 2001). It is alsonow known that the pigeon’s movementthresholds may be different for frontal andlateral viewing fields (Lea & Dittrich). Be-cause viewing times varied in the present ex-periment, and because we know only that the

pigeons’ heads were somewhat near the cen-tral region of the chamber while the targetmoved but know nothing of how a pigeon ori-ented its head toward the target while viewingit, estimating viewing angles or knowingwhether or how a pigeon used different visualfields for different vectors is difficult. Wheth-er attentional allocation would continue toapproximate optimality, as in the present ex-periment, if target speeds were considerablyfaster or slower, or if a pigeon were requiredto use a particular visual field, is not known.

Ninth, presumably there are limitations im-posed by context on a pigeon’s ability to at-tend selectively to a nearly optimal degree. Itshould be interesting, for example, to deter-mine further if a kind of ‘‘intensive’’ atten-tional effect discussed by Washburn and Put-ney (1998) in the context of researchinvolving moving stimuli and nonhuman pri-mates can be obtained in avians: Our exper-iment deals with the selective nature of atten-tion to various features of a moving stimulus,but it leaves open the question of whetherthose features vary in terms of their ability toattract attention when a participant is also re-quired to attend to other stimuli in a second-ary task. Initial evidence from a simulatednaturalistic foraging task developed by Dukasand Kamil (2000) suggests that blue jays infact do not attend as effectively to a multidi-mensional central target if they are also re-quired to attend to a peripheral target. To thedegree to which the tasks of Washburn andPutney and Dukas and Kamil are analogousto the present situation, it seems unlikely thata pigeon could selectively attend in an almostoptimal way to various features of a targetmoving in the distance if at the same time itwere foraging for food on the ground.

These issues raise questions about how wellthe present results characterize the generalcategorization of moving targets. The presentresults, nevertheless, allow us to state that onetask has been identified in which pigeons al-locate attention to a moving target in waysthat are, on average, nearly optimal.

Does Categorization of Moving TargetsInvolve Abstract Visual Concepts?

Dittrich and Lea (1993) and Lea and Dit-trich (2001) examined relations between avi-an motion discrimination and the idea of ageneral concept. Motion is, in at least a prim-


itive way, inherently conceptual, because bydefinition it involves integration of differentstimuli across time (see also Cook & Katz,1999, p. 195; Emmerton, 1986). Perhaps inthe present experiment, categorization wasnecessarily conceptual in the specific sensethat categorization logically required somekind of many-to-one unification of diverse im-ages that appeared over successive moments.Whether this logical unification and construc-tion of coherence are in terms of perceptualprocesses, memory processes, perceptual in-teractions, or in other terms, is, as far as weare aware, unclear not only in the nonhumananimal literature but in the analogous humanliterature as well (see Berthoz, 2000, for aninteresting discussion of this issue). The pos-sibility discussed by Emmerton, Cook andKatz, and Lea and Dittrich is that such unifi-cation takes place, whatever the underlyingprocesses. On the other hand, it must be ac-knowledged that from the perspective of be-havioral neuroscience, it might seem increas-ingly likely that such integration might bepossible at a fairly low level of neural pro-cessing, because much evidence suggests in-dividual striatal neurons of monkeys respondto movement and direction (see Goldstein,1996, for a review). Single-cell recording inpigeons also has identified cells sensitive tomotion and direction, and these cells are notin brain regions usually viewed in humans asrelated to abstract thought (Frost et al.,1994). It therefore remains to be determinedwhether apparent unification of successiveimages required for the perception of move-ment involves high-level abstractions or canbe explained by low-level motion and direc-tion detectors.

There is a second way in which perceptionof movement may be necessarily conceptual.Customarily, with static naturalistic stimuli,one says a pigeon has acquired the visual con-cept of a fish if it generalizes training to novelexemplars (Herrnstein & Loveland, 1964;Huber, 2001; Wright, 2001). In the presentexperiment, even though pigeons saw manydifferent exemplars of both categories, manyexemplars were nevertheless novel becausethey were selected from an immense pool ofpossible exemplars numbering in the hun-dreds of thousands (see Herbranson et al.,1999). By the customary definition, the pres-ent performances were therefore automati-

cally conceptual. However, just as there wasambiguity in whether movement is encodedin terms of abstractions or lower level sensoryevents, there is also ambiguity in the natureof the kind of generalization needed to estab-lish a learned behavior as involving concep-tual learning. For instance, would the presentperformances have generalized if the targetwere changed? Specifically, if the target werechanged from a moving white circle to a mov-ing object of changing shape and color, or toan expanding two-dimensional projection ofa peregrine falcon seeming to loom ever clos-er, would the pigeons have categorized move-ment in the same way? We do not know, butwe tend to agree with Dittrich and Lea (1993)that there is more to naturalistic motion per-ception than motion vector perception.

Our position on the question of whetherthe present performances were conceptual istherefore somewhat agnostic. Badly needed,in our opinion, is an adequate theory interms of which the basic nature of categori-zation would be described and explained(Shimp, 1984). Contemporary theories of cat-egorization deal with static prototypes andstatic specific exemplars (Ashby & Maddox,1998; Estes, 1994), however, and therefore donot as yet deal with the implications for cat-egorization of real-time processing of movingstimuli. Quantitative control of behavior bydynamic stimuli by its very nature raises im-portant questions about categorization behav-ior, including the conventional distinctiondrawn between the concrete, in the form ofindividual memories of specific exemplars,and the abstract, in the form of memories ofprototypes or rules.

Implications for Current Theories

One might be forgiven for assuming thatthe ability of a theory to deal with perfor-mances in naturalistic conditions is one rea-sonable criterion for evaluating the generalityand overall significance of the theory. Fewcontemporary theories of learning, memory,and conditioning, however, are rigorouslyheld up to that criterion. By that criterion,most current theory, when applied to perfor-mances involving moving stimuli, looks muchweaker than it customarily does. Although itis true that research on nonhuman animallearning and cognition presently includesmany tasks that explicitly address time, in-


cluding delayed matching-to-sample taskswith which to study memory (Maki et al.,1977; Wright, 2001) and many tasks withwhich to study timing itself (Church & Kirk-patrick, 2001; Gallistel & Gibbon, 2000), it isnevertheless the case that, as Skinner (1976/1996) so poignantly lamented, theoreticalanalyses of behavior often do not attempt todescribe or explain the real-time processesthat are presumably responsible for behavior-al outcomes averaged over time. Relativelyfew theories of nonhuman animal behavior,that is, predict real-time behavior. Corre-spondingly, few theories of nonhuman ani-mal behavior address discrimination of, gen-eralization over, memory for, attention to, orcategorization of dynamic visual stimuli. Aproblem is that there is no clear solution tothe problem of how to represent an animal’sreal-time ‘‘knowledge’’ of a moving visualstimulus theoretically. It is not entirely clear,at least not to us, how well-known theories ofPavlovian conditioning (Rescorla & Wagner,1972), choice behavior (Davison & McCarthy,1988; Herrnstein, 1997), short-term memory(Wright, 2001), or timing (Gallistel & Gib-bon, 2000) can address this question. We sus-pect considerable progress will result from re-search in nonhuman animal timing, visualsearch, memory, categorization, and choicethat addresses the exciting challenge of de-termining the implications of dynamic stim-uli.

The virtual absence from research on non-human animal behavior of such a striking fea-ture of the natural world as movement is sure-ly due in part to technological limitations.Until recently, few researchers have had theinterest or resources Skinner (1960) hadwhen he worked on the Pigeons in a Pelicanproject. He designed, constructed, and testedtechnology with which to present movingstimuli and even employed a servo mecha-nism by which moving stimuli were made todepend on a pigeon’s behavior. Modern com-puter technology now makes it relatively sim-ple to arrange such tasks. This rapid changein technology may not necessarily produce,however, a correspondingly rapid change inits use, because there are now established em-pirical and intellectual traditions and conven-tions in terms of which dynamic stimuli, letalone real-time interactions between dynamicstimuli and behavior, are not a part (Shimp,

2001). We therefore might be facing a mo-ment in the development of research on an-imal behavior when unusual strides forwardmight be made by paying increasing attentionto dynamic issues. We may have a rare op-portunity to increase generalizability of dataand theory by extending conventional meth-ods and theory to more naturalistically validcases involving moving stimuli. Perhaps thegenerality of contemporary theory for learn-ing and cognition might be improved if eval-uative criteria for such theories included is-sues of stimulus dynamics, the continuousbehavior stream, or how motion can improvediscrimination of stimuli, recognition of ob-jects, and detection of predators and prey(Cook & Katz, 1999; Jitsumori, Natori, &Okuyama, 1999; Kirkpatrick-Steger, Wasser-man, & Biederman, 1998; Lea & Dittrich,2001).

REFERENCES

Ashby, F. G., & Gott, R. E. (1988). Decision rules in theperception and categorization of multidimensionalstimuli. Journal of Experimental Psychology: Learning,Memory, and Cognition, 14, 33–53.

Ashby, F. G., & Maddox, W. T. (1998). Stimulus catego-rization. In M. H. Birnbaum (Ed.), Measurement, judg-ment, and decision making: Handbook of perception andcognition (pp. 251–301). San Diego: Academic Press.

Berthoz, A. (2000). The brain’s sense of movement (G.Weiss, Trans.). Cambridge, MA: Harvard UniversityPress.

Blough, D. S. (1972). Recognition by the pigeon of stim-uli varying in two dimensions. Journal of the Experimen-tal Analysis of Behavior, 18, 345–367.

Cavoto, K. K., & Cook, R. G. (2000). Cognitive prece-dence for local information in hierarchical stimulusprocessing by pigeons. Journal of Experimental Psychol-ogy: Animal Behavior Processes, 27, 3–16.

Chase, S., & Heinemann, E. G. (1972). Choices basedon redundant information: An analysis of two-dimen-sional stimulus control. Journal of Experimental Psychol-ogy, 92, 161–175.

Chase, S., & Heinemann, E. G. (2001). Exemplar mem-ory and discrimination. In R. G. Cook (Ed.), Avianvisual cognition. Retrieved September 6, 2002, fromhttp://www.pigeon.psy.tufts.edu/avc/

Church, R. M., & Kirkpatrick, B. (2001). Theories ofconditioning and timing. In R. R. Mowrer & S. B.Klein (Eds.), Handbook of contemporary learning theories(pp. 211–253). Mahwah, NJ: Erlbaum.

Cook, R. G., & Katz, J. S. (1999). Dynamic object per-ception by pigeons. Journal of Experimental Psychology:Animal Behavior Processes, 25, 194–210.

Davison, M., & McCarthy, D. (1988). The matching law.Hillsdale, NJ: Erlbaum.

Dittrich, W. H., & Lea, S. E. G. (1993). Motion as a nat-ural category for pigeons: Generalization and a fea-


ture-positive effect. Journal of the Experimental Analysisof Behavior, 59, 115–129.

Dittrich, W. H., Lea, S. E. G., Barrett, J., & Gurr, P. R.(1998). Categorization of natural movements by pi-geons: Visual concept discrimination and biologicalmotion. Journal of the Experimental Analysis of Behavior,70, 281–299.

Donis, F. J. (1999). The oblique effect in pigeons (Co-lumba livia). Journal of Comparative Psychology, 113, 107–115.

Dukas, R., & Kamil, A. C. (2000). The cost of limitedattention in blue jays. Behavioral Ecology, 11, 502–506.

Emmerton, J. (1986). The pigeon’s discrimination ofmovement patterns (Lissajous figures) and contour-dependent rotational invariance. Perception, 15, 573–588.

Emmerton, J. (1990). Pigeons’ perception of complexmovement patterns. In M. L. Commons, R. J. Herrn-stein, S. M. Kosslyn, & D. B. Mumford (Eds.), Quan-titative analyses of behavior: Vol. 8. Behavioral approachesto pattern recognition and concept formation (pp. 67–87).Hillsdale, NJ: Erlbaum.

Estes, W. K. (1994). Classification and cognition. New York:Oxford University Press.

Fremouw, T., Herbranson, W. T., & Shimp, C. P. (1998).Priming of attention to local or global levels of visualanalysis. Journal of Experimental Psychology: Animal Be-havior Processes, 24, 278–290.

Frost, B. J., Wylie, D. R., & Wang, Y. C. (1994). The anal-ysis of motion in the visual systems of pigeons. In M.N. O. Davies & P. R. Green (Eds.), Perception and motorcontrol in pigeons (pp. 248–269). Berlin: Springer-Ver-lag.

Gallistel, C. R., & Gibbon, J. (2000). Time, rate, andconditioning. Psychological Review, 107, 289–344.

Goldstein, E. B. (1996). Sensation and perception (4th ed.).Pacific Grove, CA: Brooks/Cole.

Gould, S. J., & Lewontin, R. (1979). The spandrels ofSan Marco and the Panglossian paradigm: A critiqueof the adaptationist programme. Proceedings of the Roy-al Society, B205, 581–598.

Herbranson, W. T., Fremouw, T., & Shimp, C. P. (1999).The randomization procedure in the study of cate-gorization of multidimensional stimuli by pigeons.Journal of Experimental Psychology: Animal Behavior Pro-cesses, 25, 113–135.

Herrnstein, R. J. (1997). The matching law: Papers in psy-chology and economics. Cambridge, MA: Harvard Uni-versity Press.

Herrnstein, R. J., & Loveland, D. H. (1964). Complexvisual concepts in the pigeon. Science, 146, 549–551.

Hodos, W., Smith, L., & Bonbright, J. C., Jr. (1975). De-tection of the velocity of movement of visual stimuliby pigeons. Journal of the Experimental Analysis of Behav-ior, 25, 143–156.

Huber, L. (2001). Visual categorization in pigeons. In R.G. Cook (Ed.), Avian visual cognition. Retrieved Sep-tember 6, 2002, from http://www.pigeon.psy.tufts.edu/avc/

Husband, S., & Shimizu, T. (2001). Evolution of the avi-an visual system. In R. G. Cook (Ed.), Avian visualcognition. Retrieved September 6, 2002, from http://www.pigeon.psy.tufts.edu/avc/

Jitsumori, M., Natori, M., & Okuyama, K. (1999). Rec-ognition of moving video images of conspecifics bypigeons: Effects of individuals, static and dynamic mo-

tion cues, and movement. Animal Learning & Behavior,27, 303–315.

Kirkpatrick-Steger, K., Wasserman, E. A., & Biederman, I.(1998). Effects of geon deletion, scrambling, andmovement on picture recognition in pigeons. Journalof Experimental Psychology: Animal Behavior Processes, 24,34–46.

Lea, S. G., & Dittrich, W. H. (2001). Motion discrimi-nation and recognition. In R. G. Cook (Ed.), Avianvisual cognition. Retrieved September 6, 2002, fromhttp://www.pigeon.psy.tufts.edu/avc

Maki, W. S., Moe, J. C., & Bierley, C. M. (1977). Short-term memory for stimuli, responses, and reinforcers.Journal of Experimental Psychology: Animal Behavior Pro-cesses, 3, 156–177.

Malloy, T. E., Jensen, G. C., Song, T., Herbranson, W.T., Fremouw, T., & Shimp, C. P. (1991). Interactivedemonstration of avian categorization. RetrievedJune 18, 2002, from http://www.psych.utah.edu/shimp/Avian_Categorization.htm

Martinoya, C., Rivaud, S., & Bloch, S. (1983). Compar-ing frontal and lateral viewing in the pigeon: II. Ve-locity thresholds for movement discrimination. Behav-ioral Brain Research, 8, 375–385.

McKee, S. P. (1981). A local mechanism for differentialvelocity detection. Vision Research, 21, 491–500.

McVean, A., & Davieson, R. (1989). Ability of the pigeon(Columba livia) to intercept moving targets. Journal ofComparative Psychology, 103, 95–99.

Neiworth, J. J., & Rilling, M. E. (1987). A method forstudying imagery in animals. Journal of ExperimentalPsychology: Animal Behavior Processes, 13, 203–214.

Potts, W. (1984). The chorus-line hypothesis of manoeu-vre coordination in avian flocks. Nature, 309, 344–345.

Rescorla, R., & Wagner, A. (1972). A theory of Pavlovianconditioning: Variations in the effectiveness of rein-forcement and nonreinforcement. In A. H. Black &W. F. Prokasy (Eds.), Classical conditioning: Vol 2. Cur-rent research and theory (pp. 64–99). New York: Apple-ton-Century-Crofts.

Riley, D. A., & Brown, M. F. (1991). Representation ofmultidimensional stimuli in pigeons. In G. R. Lock-head & J. R. Pomerantz (Eds.), The perception of struc-ture (pp. 227–245). Washington, DC: American Psy-chological Association.

Rilling, M. E., & LeClair, T. L. (1989). Visually guidedcatching and tracking skills in pigeons: A preliminaryanalysis. Journal of the Experimental Analysis of Behavior,52, 377–385.

Rilling, M., LeClair, L., & Warner, M. (1993). A compar-ative, hierarchical theory for object recognition andaction. In T. R. Zentall (Ed.), Animal cognition: A tributeto Donald A. Riley. Comparative cognition and neuroscience(pp. 313–333). Hillsdale, NJ: Erlbaum.

Rilling, M. E., & Neiworth, J. J. (1987). Theoretical andmethodological considerations for the study of imag-ery in animals. Learning and Motivation, 18, 57–79.

Rosch, E., & Mervis, C. B. (1975). Family resemblance:Studies in the internal structure of categories. Cogni-tive Psychology, 7, 573–605.

Sekuler, R. (1975). Visual motion perception. In E. C.Carterette & M. P. Friedman (Eds.), Handbook of per-ception (Vol. 5, pp. 387–424). New York: AcademicPress.

Shimp, C. P. (1973). Probabilistic discrimination learn-


ing in the pigeon. Journal of Experimental Psychology, 97,292–304.

Shimp, C. P. (1984). The question: Not shall it be, butwhich shall it be? Commentary on ‘‘Are theories oflearning necessary?’’ by B. F. Skinner. Behavioral andBrain Sciences, 7, 536–537.

Shimp, C. P. (2001). Behavior as a social construction.Behavioural Processes, 54, 11–32.

Shimp, C. P., & Friedrich, F. J. (1993). Behavioral andcomputational models of spatial attention. Journal ofExperimental Psychology: Animal Behavior Processes, 19,26–37.

Shimp, C. P., Herbranson, W. T., & Fremouw, T. (2001).Avian visual attention: Laboratory and cultural per-spectives. In R. G. Cook (Ed.), Avian visual cognition.Retrieved September 6, 2002, from http://www.pigeon.psy.tufts.edu/avc/

Skinner, B. F. (1960). Pigeons in a pelican. American Psy-chologist, 16, 28–37.

Skinner, B. F. (1996). Farewell, my lovely. In A. C. Ca-tania & P. H. Hineline (Eds.), Variations and selections:An anthology of reviews from the Journal of the ExperimentalAnalysis of Behavior (p. 230). Bloomington, IN: Society

for the Experimental Analysis of Behavior. (originalwork published 1976)

Smith, R. F. (1974). Topography of the food-reinforcedkey peck and the source of 30-millisecond interre-sponse times. Journal of the Experimental Analysis of Be-havior, 21, 541–551.

Teuber, H.-L. (1960). Perception. In Handbook of physi-ology (Vol 3., pp. 1595–1668). Washington, DC: Amer-ican Physiological Society.

Walls, G. (1942). The vertebrate eye and adaptive radiation.Bloomfield Hills, MI: Cranbrooke.

Washburn, D. A., & Putney, R. T. (1998). Stimulus move-ment and the intensity of attention. The PsychologicalRecord, 48, 555–570.

Watanabe, S. (1991). Effects of ectostriatal lesions onnatural concept, pseudoconcept, and artificial patterndiscrimination in pigeons. Visual Neuroscience, 6, 497–506.

Wright, A. A. (2001). Learning strategies in matching-to-sample. In R. G. Cook (Ed.), Avian visual cognition.Retrieved September 6, 2002, from http://www.pigeon.psy.tufts.edu/avc/

Received October 22, 2001Final acceptance June 24, 2002

CATEGORIZING A MOVING TARGET IN TERMS OF ITS SPEED ...

Documents

Transcript of CATEGORIZING A MOVING TARGET IN TERMS OF ITS SPEED ...