Perceptual similarity in autism

Perceptual similarity in autism

Lewis Bott, Jon Brock, Noellie Brockdorff, Jill Boucher, and Koen LambertsUniversity of Warwick, Coventry, UK

People with autism have consistently been found to outperform controls on visuo-spatial tasks such asblock design, embedded figures, and visual search tasks. Plaisted, O’Riordan, and others (Bonnel et al.,2003; O’Riordan & Plaisted, 2001; O’Riordan, Plaisted, Driver, & Baron-Cohen, 2001; Plaisted,O’Riordan, & Baron-Cohen, 1998a, 1998b) have suggested that these findings might be explainedin terms of reduced perceptual similarity in autism, and that reduced perceptual similarity couldalso account for the difficulties that people with autism have in making generalizations to novel situ-ations. In this study, high-functioning adults with autism and ability-matched controls performed alow-level categorization task designed to examine perceptual similarity. Results were analysed usingstandard statistical techniques and modelled using a quantitative model of categorization. This analy-sis revealed that participants with autism required reliably longer to learn the category structure thandid the control group but, contrary to the predictions of the reduced perceptual similarity hypothesis,no evidence was found of more accurate performance by the participants with autism during thegeneralization stage. Our results suggest that when all participants are attending to the same attributesof an object in the visual domain, people with autism will not display signs of enhanced perceptualsimilarity.

Autism is characterized by deficits in reciprocalsocial behavior, communication, and behaviouralflexibility (American Psychiatric Association,1994; Wing, 1996), but is also associated withcertain strengths, particularly in performance onvisuo-spatial tasks such as block design (e.g.,Shah & Frith, 1993; Tymchuk, Simmons, &Neafsey, 1977), the embedded figures task(Jolliffe & Baron-Cohen, 1997; Shah & Frith,1983; but see Brian & Bryson, 1996), and visual

search tasks (O’Riordan & Plaisted, 2001;O’Riordan et al., 2001; Plaisted et al., 1998b). Ithas been argued that studying these cognitivestrengths in autism may be particularly informa-tive because, unlike deficits, they cannot readilybe explained in terms of general mental retardation(Happe, 1999b).

Frith (1989; see also Frith & Happe, 1994;Happe, 1999a) proposed that many of the cogni-tive strengths and weaknesses in autism could be

Correspondence should be addressed to Lewis Bott, Department of Psychology, New York University, 6 Washington Place,

New York, NY 10003, USA. Email: [email protected]

Lewis Bott is now at the Department of Psychology, New York University. Noellie Brockdorff is now at the Centre for

Communication Technology, University of Malta, Malta. Jon Brock is now at the Department of Experimental Psychology,

University of Oxford. Research was supported by an RDTF grant to Lamberts, Boucher, and Bott from the University of

Warwick. Lewis Bott was supported by a studentship from the Brain and Behavioral Sciences Research Council and by a grant

from the Centre National de la Recherche Scientifique (France) as part of Action Thematique et Incitative. Jon Brock was supported

by a studentship from the Williams Syndrome Foundation.

# 2006 The Experimental Psychology Society 1http://www.tandf.co.uk/journals/pp/17470218.html DOI:10.1080/02724980543000196

THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY

2006, 59 (0), 1–18

understood in terms of what she called “weakcentral coherence”—a tendency to focus on localdetails at the expense of the global “big picture”and to process information in isolation from itscontext. According to this account, people withautism perform well on the embedded figurestask because this involves searching for a localtarget in a global picture. Similarly, they performwell on the block design task because they find itrelatively easy to break the global target patterndown into its constituent local parts. However,weak central coherence struggles to explainenhanced visual search performance in autism.Moreover, evidence from other paradigms tendsto show evidence for enhanced local processingbut little evidence for impaired global processing(e.g., Mottron & Burack, 2001).

More recently, Plaisted and colleagues(O’Riordan & Plaisted, 2001; O’Riordan et al.,2001; Plaisted, 2001; Plaisted et al., 1998a,1998b) have proposed that findings previouslyattributed to weak central coherence may bebetter understood in terms of reduced perceptualsimilarity.1 Thus, detection of the target shape inthe embedded figures test is relatively easybecause the target is seen as being relatively dis-similar to other shapes in the overall picture(Plaisted, 2001). Similarly, superior performanceon visual search tasks can be explained in termsof enhanced ability to discriminate betweenthe target item and similar distractor items(O’Riordan & Plaisted, 2001, cf. Wolfe, Cave, &Franzel, 1989). According to this account,reduced perceptual similarity also entails a deficitin generalization such that novel objects andevents are seen as highly distinctive from previousexperiences. Plaisted (2001) therefore suggestedthat reduced similarity could explain poor

generalization of social training to real-life situa-tions (e.g., Ozonoff & Miller, 1995; Swettenham,1996) and the restricted range of interests shownby many people with autism.

In support of this account, Plaisted et al.(1998a) reported a perceptual learning study inwhich, unlike controls, high-functioning adultswith autism failed to benefit from preexposureto similar stimuli. The authors assumed thatsimilarity between stimuli was an increasingfunction of the number of shared features and adecreasing function of the unshared features(cf. Tversky, 1977). They argued that individualswith autism failed to generalize from thepreexposure phase to the test phase because theywere preferentially attending to the features thatdiscriminated between stimuli and thereforefailed to notice the similarity between the sets ofstimuli in the preexposure phase and the testphase of the task. The difficulty for this accountis that the two stimuli in the preexposure phaseand the two stimuli in the test phase all differedon the same features. Thus, by attending to thefeatures that discriminated between stimuli inthe different phases, participants with autismwould also have been attending to the featuresthat enabled discrimination between the twostimuli within the transfer phase and should there-fore have shown an enhanced transfer effect. Analternative explanation for this finding is that par-ticipants with autism were simply more likely thancontrols to follow the experimenter’s instruction totreat the test phase as a new task. This wouldreflect differences in strategy (or naivety to experi-mental manipulations) rather than perceptualabnormalities.

The aim of the current study was to formalizeand then test the reduced perceptual similarity

1 Plaisted and O’Riordan generally express their theory in terms of “enhanced discrimination” rather than “reduced perceptual

similarity”. Although these two expressions are not identical, there is general agreement in the categorization literature that the per-

ceptual similarity of two objects is monotonically related to the extent to which they can be discriminated (see, for example, Medin &

Schaffer, 1978; Nosofsky, 1986, 1987; Shepherd, 1987). Put simply, an individual who finds it relatively easy to discriminate between

two objects would also consider them relatively dissimilar. However, the concepts of discrimination and (dis)similarity are not used

interchangeably in the categorization literature, and the theory described by Plaisted et al. (1998a) corresponds better to the categ-

orization expression “reduced perceptual similarity” than to the term “enhanced discrimination”. Because the techniques and theory

employed in this article are based on standard paradigms used in perceptual categorization, we adopt the term “reduced perceptual

similarity” for the remainder of the article.

BOTT ET AL.

2 THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY, 2006, 59 (0)

hypothesis in terms of exemplar models of categor-ization (e.g., Medin & Schaffer, 1978; Nosofsky,1986, 1990, 1992; Nosofsky & Palmeri, 1996).According to such models, people representcategories by storing individual exemplars with alabel indicating their category membership. Anovel item is categorized by computing its simi-larity to all the exemplars in memory, summingthe similarities for each category and then assign-ing it to the category with the highest summedsimilarity to all items in memory. In fact, olditems are also categorized in the same way, bycomputing similarities to all exemplars (includingthemselves). This means that items that aresimilar to members of another category are likelyto be misremembered. Exemplar models thereforeprovide an explicit link between perceptualsimilarity, categorization, and generalization, andthey allow predictions to be made about theeffect of reduced perceptual similarity on categor-ization performance (see Hayes & Taplin, 1992,1993a, 1993b, for a discussion of categorizationfrom a developmental and clinical perspective).

One prediction that can be derived from thereduced similarity hypothesis is that people withautism will show a reduced prototype effect(Plaisted, 2001). The prototype of a category isthe category member whose features representthe average of the group. As such, its summedsimilarity to all the category members is higherthan that for any other item, and it is thereforemore likely to be correctly classified than anyother category member. However, if all themembers of a category are seen as highly distinc-tive then this effect will be reduced because othergroup members will exert less of a reinforcingeffect on the prototype.

Consistent with this prediction, Klinger andDawson (2001; see also Klinger & Dawson,1995) reported that children with autism showeda reduced prototype effect. Participants weretaught to discriminate between two sets ofcartoon animals. They were then presented withtwo novel members of one category, one ofwhich was the prototype, and were asked todecide which was the best example of that cat-egory. Typically developing children reliably

chose the prototype over the nonprototypicalanimal, but children with autism performed atchance levels. However, one potential problemwith this study is that, although some dimensionsdiffered within categories, others dimensionsvaried between categories only. For example, allmembers of one category had octagonal bodies,while members of other categories had bodies ofdifferent shapes. Participants with autism maytherefore have learnt to distinguish between thecategories by attending to such invariant featuresand would then have been unable to choose themost representative example of the category.This would have led to a reduced prototypeeffect but would not imply reduced perceptualsimilarity.

The current study also used a categorizationparadigm to investigate the reduced perceptualsimilarity hypothesis. However, we avoided thedifficulty suggested above by employing a designwhereby participants were forced to attend tomultiple dimensions of the stimuli in order tolearn the category structure. In the trainingphase of our experiment, participants learned toclassify 10 rectangles varying in height and widthinto two experimenter-determined categories,referred to as Categories A and B. Rectangleswere chosen because they could vary along con-tinuous dimensions, thus discouraging the use ofverbalized rules. The stimuli were presented oneat a time on a computer screen, and participantswere given feedback informing them whethertheir classification was correct or not. Onceparticipants had learned to correctly classify all10 rectangles accurately, they proceeded onto atest phase in which they classified the same 10 rec-tangles together with 6 novel rectangles, but thistime without any feedback.

Figure 1 shows the coordinates of the stimuli interms of height and width, together with theirappropriate classification in the training phase(either A or B). It can be seen that the majorityof the Category B rectangles are located in thebottom left triangle of the rectangle space, whilethe A rectangles are in the top right triangle.However, there is one B rectangle (B6) that isthe exception to the general rule because it is

SIMILARITY IN AUTISM

THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY, 2006, 59 (0) 3

located in the top right corner of the spaceand is consequently surrounded by Category Arectangles. We refer to this rectangle as the excep-tion item, an important part of the design that wediscuss below. Note that it is not possible to learnthe correct labels of the rectangles by attending toone dimension of the stimuli only, whether this beheight, width, area, or shape.2

Once participants had learned to correctlyclassify all 10 rectangles, they proceeded onto atest phase in which they classified the same 10rectangles together with 6 novel rectangles, but thistime without any feedback. According to exemplarmodels, classification of the exception iteminvolves computing the similarity of the exceptionitem to itself, to the Category A rectangles and tothe remaining Category B rectangles. Because thesimilarity to Category A members is relativelyhigh, participants are likely to make more errors

in classifying the exception item. However, theextent of this effect will be determined by howmuch an individual is generally able to discrimi-nate the different rectangles: If they have lowdiscrimination abilities, then the Category Arectangles will be considered highly similar tothe exception item, so the probability of misclassi-fication will be relatively high. But if an individ-ual can easily discriminate the rectangles, theinterference from the surrounding A exemplarswill be low, and performance on the exceptionitem will be good. Thus, if people with autismjudge the rectangles to be less similar than con-trols, then they should be better at discriminatingthe exception item in the testing phase of theexperiment.

The experimental design also allowed quantitat-ive modelling of categorization performance usingthe general context model (GCM; Nosofsky,1986). The GCM is an exemplar-based model ofcategorization, which assumes that items areclassified on the basis of their similarity to allitems whose category membership is known.Similarity is defined using Shepard’s (1987) uni-versal law of generalization, which states that thesimilarity between two objects is determined bytwo factors: first the distance between the objectsin psychological space, and second a scalingparameter representing a participant’s level ofmemory sensitivity in discriminating among dis-tinct exemplars. A high value means that theparticipant distinguishes easily among items inmemory, while a low value means that exemplarsare difficult to discriminate. The reduced per-ceptual similarity hypothesis therefore predictshigher values of the scaling parameter for individ-uals with autism than for controls.

We have so far emphasized predictions of thereduced perceptual similarity hypothesis for thetesting phase of the experiment and not the train-ing phase. This is because we wish all participantsto be equated on the extent to which they knowthe training exemplars before being tested on

Figure 1. Stimuli structure for the categorization task. Each label

refers to a rectangle of a given height and width. Rectangles

marked with an A or a B refer to items that were presented in

the training phase (and belonged to category A or B respectively).

Rectangles marked with question marks indicate rectangles seen

only in the testing phase of the experiment. The numbers adjacent

to the rectangles relate to the stimuli numbers shown in Figures 2

and 3. The rectangle B6 is surrounded by A category members

and is therefore referred to as the exception item.

2 If a participant tried to classify the rectangles on the basis of height alone, then it would not be possible to correctly classify

rectangles B7 and A4 (amongst others) because they have the same height but are in different categories. Similar arguments hold

for width (e.g., rectangles B9, B8, and A2), area (rectangles A1 and B8), and shape (B8 and A4 are both square).

BOTT ET AL.


generalization. By training participants until theyreach a given criterion of accuracy, we eliminatethe possibility that the different groups aregeneralizing from different knowledge bases.Nonetheless, it is possible to consider what predic-tions could be made from the reduced perceptualsimilarity hypothesis concerning category learning,as opposed to generalization.

Recall that participants who have high discri-mination abilities would treat the rectangles asrelatively distinct, individual items associatedwith the same category label. This implies thatthe exception item would be relatively easy tolearn because there would be less interferencethat could arise from the surrounding rectanglesof the opposing category. However, the normalrectangles would benefit from a high similaritybecause, overall, similar rectangles are placedwithin the same category. Thus, high discrimi-nation would facilitate learning on the exceptionitem but would harm learning on the normalrectangles. The reduced perceptual similarity hypo-thesis would therefore predict that the participantswith higher functioning autism (HFA) wouldhave less difficulty in acquiring the exceptionitem than would controls, but more difficulty inlearning the normal items. Predictions for theoverall learning times are difficult to generatebecause there are both costs and benefits to highsimilarity, and the precise ratio cannot be deter-mined in advance.

It is important to realize that there are manyother factors aside from exemplar similarity thatdetermine the speed with which participantsacquire exemplars and the overall category struc-ture. These include individual factors such asgeneral confidence level and learning strategies,factors that are difficult to control for and likelyto vary across our two groups. In general, learninga category is a far more complex cognitive activitythan naming the exemplars, suggesting that theclearer test of reduced perceptual similaritywould be the responses in the testing phase andnot those from the training phase.

Fitting the GCM required that participantsperform a large number of trials to eliminateas much noise as possible from their data.

This consideration, together with the inherentdifficulty of the task, entailed that it was not pos-sible to test children or low-functioning indi-viduals with autism. The participants in thisstudy were therefore high-functioning adultswith autism and nonautistic controls matched onverbal mental age and performance mental age.

Another requirement for modelling the datawas that we needed to know how participantsperceive the stimuli used in the experiment.Participants may view the stimuli in the same wayas we have constructed them (as in Figure 1), or itmay be that they perceive the rectangles in differentway, by treating the rectangles as if they vary alongonly one dimension: area, for example. The GCMrequires a set of psychological coordinates for therectangles to describe categorization performanceaccurately. The most common approach to deter-mining this set of coordinates is by performing asimilarity ratings experiment using the stimulithat will be used in the principal experiment, inour case the 16 rectangles shown in Figure 1.This involves presenting participants with all pos-sible pairs of stimuli and asking them to ratehow similar they think each of the pairs are.These data are then analysed using a multi-dimensional scaling (MDS) analysis that convertsthe “distances” (similarity judgements) betweenstimuli into a “map” where each of the stimuli isgiven its own coordinates. The set of coordinatesfor all the rectangles is chosen so that the distancesbetween stimuli on the map are as close as possibleto the original distances given by participants. Theresulting map is not restricted to two dimensions,but can be of any dimensionality from 1 to N 2

1, where N is the number of different stimuli usedin the experiment. Choosing the most appropriatenumber of dimensions for the space is generallydone a priori or on grounds of parsimony, in asimilar way to the approach taken in factoranalysis. The final map provides a representationof the psychological space and can be used by theGCM. In order to determine the psychologicalcoordinates of our stimuli, therefore, we performeda similarity ratings experiment with our HFA andour control participants, before commencing thecategorization study (see Kruskal & Wish, 1978,



and Young & Harris, 1994, for useful introduc-tions to MDS).

In addition to being a useful step in the model-ling of the GCM, this task also allowed investi-gation of potential differences between groups inthe representation of the stimuli. In particular,participants might choose to encode the stimuliusing spaces of differing dimensionality. Forexample, those in the HFA group may base theirsimilarity judgements on only one dimension ofthe stimuli, whereas those in the control groupmight use two dimensions. Note, however, thatit is not our aim to test the reduced perceptualsimilarity hypothesis using the similarity ratingstask: Although we might expect the theory topredict lower similarity ratings overall for theHFA participants, differences that occur betweenthe groups might also arise because of responsebiases that do not necessarily reflect underlyingperceptual similarity. Because the effects ofreduced perceptual similarity and response biaswould be confounded, we restrict our investigationto differences involving the number of dimensionson which participants represent the stimuli.Response biases variation across individualswould not pose a problem for our analysisbecause individual ratings are normalized as partof the MDS algorithm.

EXPERIMENT

Method

ParticipantsA total of 12 high-functioning adults with autismwere recruited via personal connections and local

support groups and services. They had all beendiagnosed by psychiatrists or clinical psychologistsas having Asperger syndrome. However, it was notclear from the available information that earlylanguage development had been entirely normalin all cases as required by American PsychiatricAssociation (1994) and World Health Organiza-tion (1992) definitions of Asperger syndrome.The more open term higher functioning autism(HFA) was therefore used to describe theseparticipants. All were given travel expenses and agift voucher worth £10. A total of 17 controlswere also recruited. These participants wereundergraduate students at the University ofWarwick and were paid £10 for taking part.Verbal mental age was assessed using the vocabu-lary and comprehension subtests of the WechslerAdult Intelligence Scale (WAIS; Wechsler, 1986).Performance mental age was assessed using thepicture completion and object assembly subtests.Participant details are summarized in Table 1.There were no significant group differences in per-formance mental age or verbal mental age (ts, 1),although the HFA group were significantly olderthan controls, t(27) ¼ 2.11, p ¼ .043.

StimuliStimuli were 16 rectangles presented on the screenof a computer monitor. The outline of each rec-tangle was drawn in green single pixel lines ona black background. The dimensions of therectangles represented the factorial combinationof four heights and widths, which were 30, 60,90, and 120 pixels. The same stimuli were usedfor the similarity ratings task and for the categor-ization experiment.

Table 1. Participant details

Agea Verbal mental agea

Performance

mental agea

n No. male No. female M Range M Range M Range

HFA 12 10 2 30 20–62 27.10 17–36 19.5 7–31

Controls 17 6 11 21 19–45 27.12 20–35 21.6 17–31

aIn years.

BOTT ET AL.


ProcedureParticipants were tested individually in a sound-proof booth with the experimenter in an adjoiningroom. The similarity ratings task was performedfirst, followed by the categorization task andfinally the WAIS subtests.

In the similarity ratings task, participants weretold that they would see pairs of rectangles pre-sented on a computer screen and that they wouldhave to judge how similar they thought each pairwas. They were told to use the numbers 1–9 onthe computer keyboard to make their judgement,1 being the least similar, and 9 being the mostsimilar. They then completed 480 trials in whichthey saw each of the possible rectangle pairstwice. A rectangle was never paired with itself,and the second time a pair was presented theirpositions to the left and right of the screen werechanged. Rectangle pairs were presented in adifferent random order for each participant.

In the categorization task, participants weretold that one of the experimenters had chosensome of the rectangles to be his, while the otherexperimenter had chosen some of the remainingrectangles. Participants were then told that theyhad to learn which rectangles belonged to eachexperimenter and that no one rectangle couldbelong to both experimenters. They wereinformed that they would receive feedback in thefirst part of the experiment (the training phase),but later on that feedback would disappear in thesecond phase (the testing phase). They were alsotold that they would see “new” rectangles in thetesting phase and that they should classify theserectangles on the basis of the classifications thatthey had made in the training phase.

In each block of the training phase, the 10training items were each presented once in arandom order. A typical trial consisted of the pres-entation of a fixation point (1 second) followedby the rectangle appearing on the screen andremaining until the participant had made theirresponse. All responses were made using astandard button-box, with two buttons corre-sponding to the two different categories ofrectangle. Feedback was provided with a highbeep if the participant responded correctly and a

low beep if they responded incorrectly. Trainingcontinued until participants could correctly ident-ify each of the 10 rectangles without error for fourcomplete blocks, whereupon they moved onto thetesting phase after a short break. There were 40blocks in the test phase, with all 16 rectanglesbeing presented once in a random order in eachblock, making a total of 640 trials. The procedurewas the same as that in the training phase, apartfrom the absence of any feedback.

Results

Similarity ratingsThe goal of this experiment was to provide a suit-able set of coordinates for the GCM and to estab-lish whether there might be any differences betweenthe perceptual representations of individuals in theHFA and control groups. To this end, a type ofMDS known as individual scaling (INDSCAL,Carroll & Chang, 1970) was applied, assumingordinal data. The INDSCAL analysis results in asingle solution for all participants together with aset of weights for each participant dictating theextent that the participant relies on each dimension.The GCM can then be applied using the averagesolution while differences between groups can beanalysed using the dimensional weights.

We report the results of applying MDS assum-ing a one-, two-, or three-dimensional solution.The stress and R2 figures that we report are themean of each individual participant’s fit to thesingle group solution. In the two-dimensionalcase, the INDSCAL analysis resulted in stressand R2 values of .3 and .46 respectively, and forthe three dimensional case they were .23 and .52.A replicated MDS analysis was applied to obtaina one-dimensional solution, with accompanyingstress and R2 values of .45 and .40, respectively(INDSCAL cannot be applied because there isonly a single dimension). We did not try toselect the solution with the most appropriatedimensionality because we were able to fit theGCM using each of the solutions and comparethe results (see the Model Fitting section below).

We now examine whether there are differencesbetween the two groups in how they perceive the



rectangle space. One possibility is that participantsin the two groups might be making their similarityjudgements using different numbers of dimen-sions. For example, the HFA group might preferto use only one dimension on which to base theirsimilarity ratings, while the controls might prefertwo dimensions. This hypothesis can be investi-gated by looking at the participant weightings onthe dimensions, as a function of group. In addres-sing this question, we are assuming that at leastsome of our participants encoded the rectanglesusing two or three dimensions whereas, in fact,the R2 results reported above do not lead us toreject the hypothesis that all participants usedonly a one-dimensional solution. We proceedwith the analysis of higher dimensional solutionsbecause individual variation in the dimensionalityof encoding (the very question we are investigatinghere) contributes to overall variance and may havemasked the extent to which it is possible to deter-mine the best group solution.

To examine the extent to which different par-ticipants used a different number of dimensions,we analysed a transformation of the dimensionalweights referred to as the w-score3 (MacCallum,1976; Young & Harris, 1994). The w-score indi-cates how far the individual’s weightings differedfrom the group average: The more a participantrelies on one dimension, the more extreme theweight ratio becomes and the higher theirw-score. Consequently, if the HFA group weremaking their similarity judgements based on

fewer dimensions than were the controls,their w-scores should be higher. For the two-dimensional solution, the mean w-scores were0.242 (SD ¼ 0.158) for the HFA group, and0.156 (SD ¼ 0.100) for the control group. A two-tailed, equal-variance t test revealed this differenceto be narrowly nonsignificant, t(27) ¼ 2.00, p ¼

.055. Analysis of the weights for the three-dimen-sional solution revealed a similar story, withmeans of 0.27 for the HFA group and 0.2 for thecontrols, t(27) ¼ 1.6, p ¼ .1. These resultsprovide some tentative support for the idea thatthe HFA group were making their similarity judge-ments on the basis of fewer dimensions than werethe controls.4 However, due to the difficulty ofestablishing the most appropriate group space,and the ambiguous results of the t tests, werefrain from drawing firm conclusions regardingdifferences in the representation of therectangle space between groups.

Categorization taskFor both training and testing phases, an arcsinetransformation was carried out on all choice pro-portions to improve the conformity of the datato the standard assumptions of analysis of variance(ANOVA; Howell, 1997). In the training phase,4 of the HFA participants failed to reach thecriterion of four consecutive blocks of correctresponses. Of these, 2 failed to perform abovechance on a single block and were excluded fromall further analyses.5 The other 2 participants

3 Referred to as the “weirdness” index by Young and Harris (1994).4 An alternative to analysing the participant weights might have been to fit two completely different solutions, one for the HFA

group and another for the control group, and to examine the degree to which each solution fitted the data. One might have expected

that, if the HFA group were basing their judgements on fewer dimensions, then the most optimal dimensionality for the HFA group

would be lower than that of the controls. However, there are two problems with this method, which would make any result difficult to

interpret. The first is that formal methods for deciding dimensionality, such as Lee’s (2001) BIC, favour a low dimensional solution

for data that has a high variance. Thus, different ideal dimensionalities could be the result of differences in the variance across groups,

rather than a difference in the perceived rectangle space. The second problem is that even if the HFA solution requires a high-

dimension solution, it could be because different participants within the group choose to base their judgements on different

dimensions. Hence, a high-dimensional solution is required to account for between-participant variation in the choice of dimensions,

even though any single participant might be best modelled with a low-dimensional solution. The INDSCAL technique that we used

avoids both of these potential problems.5 The removal of these two HFA participants did not substantially alter the matching between our two groups. The new means

and ranges for the HFA group were: 31.2 (20–62), 27.3 (20–35), and 22 (17–31) years for age, verbal mental age, and performance

mental age, respectively. As before, there were no reliable differences between groups on the verbal and performance mental age

measures, t(25)s , 1, ps . .83, and the difference on age remained, t(25) ¼ 2.6, p ¼ .014.

BOTT ET AL.


were very close to criterion (for example, gettingone response wrong out of a run of 4 blocks) andproceeded onto the test phase after 100 trainingblocks. All analyses were therefore carried outboth with and without these two noncriterion lear-ners (NCLs). However, only the results includingthe NCLs are reported, unless there were qualita-tive differences between the conclusions of the twoanalyses. One HFA participant (not an NCL)dropped out three quarters of the way into thetest phase because he was tired. His results weretherefore based on the responses that were col-lected (27 blocks out of 40).

Training phaseParticipants in the HFA group took more blocksthan did controls to reach criterion in the trainingphase t(23) ¼ 2.74, p ¼ .012, with averages of 46and 30 blocks, respectively. NCLs were notincluded in this comparison because the precisenumber of blocks required for them to reachcriterion was unknown (although assuming learn-ing times of greater than 100 blocks wouldincrease the difference between the two groups).Turning to the proportion correct as the depen-dent measure, Figure 2 shows the accuracy ofresponses given to each rectangle in the trainingset (NCLs included). These results were subjectedto an ANOVA with rectangle as a repeatedmeasure and group as a between-subjects factor.The HFA group performed significantly worsethan controls, F(1, 27)¼ 4.43, p ¼ .045, reflecting

the greater number of blocks required to reach cri-terion. However, this effect became nonsignificantwhen the NCLs were removed, F(1, 23) ¼ 3.7, p¼ .067. There was a significant effect of rectangle,F(1, 9) ¼ 12.07, p ¼ .001, demonstrating that theexception item was learnt with more difficulty thanother rectangles. In fact, 7 out of the 10 partici-pants in the HFA group and 11 of the 17 controlsfound the exception item (Stimulus Number 6) themost difficult to learn. There was no significantinteraction between group and rectangle, F(9,243) ¼ 1.15. In summary, the HFA group foundthe category structure significantly more difficultto learn than did the controls, but there was no evi-dence of differences in the rate at which they learntthe exception item.

The above analysis was conducted on the datafrom the entire training phase, during whichdifferent participants received different amountsof training. However, it could be argued that theextra blocks of exemplars seen by the HFAgroup might obscure differences in discriminationor general learning strategy. For example, theHFA participants might have learned the excep-tion item better than did controls earlier on learn-ing but this effect could then have been obscuredby requiring a large number of trials to learn theother items. To investigate this issue, we analysedthe data from the first 25 blocks only. Up until thispoint, all participants were still in the trainingphase and had therefore received the sameamount of training. We divided the 25 blocksinto five divisions and, for each participant,found the mean accuracy for the exception itemand for the normal rectangles. We averagedtogether responses to the normal rectangles,unlike in the analysis above, because we had farfewer data points than previously and we alsohad an extra factor, that of block. We thenperformed an ANOVA on these figures, with rec-tangle (normal and exception) and block (1 to 5) asrepeated measures factors, and group (HFA andcontrol) as a between-participants factor. Theresults of this analysis were very similar to resultsof the analysis carried out using the data fromthe entire training set. We found that bothgroups learned the exemplars better over time,

Figure 2. Proportion correct during the training phase. Error bars

are the standard errors for each stimulus, within developmental

group. Stimulus numbers refer to the category structure shown in

Figure 1, where 6 is the exception item.



F(4, 100) ¼ 32.08, MSE ¼ 0.29, p , .0005,and that classification of the exception rect-angle was poorer than that for the others,F(1, 25) ¼ 29.40, MSE ¼ 0.42, p , .0005. Thiseffect was present in both the control participants,F(1, 16)¼ 19.18, MSE ¼ 0.40, p , .0005, and theHFA group, F(1, 9) ¼ 11.67, MSE ¼ 0.45, p ¼

.008. The difference between the exception itemand the normal rectangles diminished overblocks, F(4, 100) ¼ 5.14, MSE ¼ 0.30, p ¼

.001. The HFA group performed worse than con-trols overall, F(1, 25)¼ 3.6, MSE ¼ 1.24, p ¼ .07,and this difference diminished over learn-ing, F(4, 100) ¼ 2.34, MSE ¼ 0.29, p ¼ .061.However, there was no interaction of groupwith rectangle, F(1, 25) ¼ 0.045, MSE ¼ 0.42,p ¼ .834, nor was there an interaction of group,rectangle and block, F(4, 100) ¼ 0.771, MSE ¼

0.303, p ¼ .547. As we found when we analysedthe complete training set, the HFA groupappeared to learn more slowly overall, but therewas no evidence of any learning differencesrelated to the exception item.

Test phaseFigure 3 shows the proportion of correct responsesin the test phase to the 10 rectangles that partici-pants had learnt to classify during the trainingphase. Performance on the exception item(Stimulus 6) was clearly worse than that on theother nine items, despite the fact that all itemshad been learnt to criterion in the training phase.

This observation was confirmed by performing arepeated measures ANOVA on the proportioncorrect with group and rectangle as factors. Asignificant main effect of rectangle was found,F(9, 225) ¼ 17.12, MSE ¼ 0.1, p , .001, indicat-ing that the exception item was reliably different toother items (see Figure 3), but there was no maineffect of group, F(1, 25) ¼ 0.9, MSE ¼ 0.56, p ¼

.352. Importantly, there was no significant inter-action between group and rectangle, F(9, 225) ¼1.69, MSE ¼ 0.1, p ¼ .094, a finding that failsto support the reduced perceptual similarityhypothesis.

Despite our failure to find a reliable interaction,we analysed performance between groups on theexception item in particular because the reducedperceptual similarity hypothesis makes a prioripredictions regarding this item. Mean proportioncorrect for the HFA group was .73 (SD ¼ 0.10),while for the controls it was .78 (SD ¼ 0.050).Thus, the means were in the opposite directionto that predicted by the reduced perceptual simi-larity hypothesis. Analysis of confidence intervalsdemonstrated that we can be 95% certain thatthe mean HFA correct score cannot be morethan .014 greater than that of the controlgroup, assuming t(25), one-tailed, because thereduced perceptual similarity hypothesis predictsMHFA�Mcontrol. Thus, these results suggest thatany effects of reduced perceptual similarity arenegligible in such a task as ours.

We also conducted a correlational analysisinvolving the factors of age, verbal mental age,and performance mental age. This analysisrevealed that classification of the exception itemwas not associated with age (r ¼ .194, p ¼ .33)nor with verbal mental age (r ¼ .196, p ¼ .53),but was positively correlated with performancemental age: r ¼ .51; Z(27) ¼ 2.748; p ¼ .006. Aone-way analysis of covariance (ANCOVA) wastherefore performed with classification of theexception item as the dependent variable, groupas the between-subjects factor, and performancemental age as the covariate. The main effect ofperformance mental age approached significance,F(1, 23) ¼ 4.00, p ¼ .058, and there was asignificant interaction between the group and

Figure 3. Proportion correct for responses to training items, during

the testing phase. Stimulus numbers refer to the category structure

shown in Figure 1, where 6 is the exception item. Error bars are

the standard errors for each stimulus, within developmental group.

BOTT ET AL.


performance mental age, F(1, 23) ¼ 4.6, p ¼ .043.Crucially, however, once the variance due to per-formance mental age had been factored out,control participants showed reliably better responseson the exception item, F(1, 23) ¼ 4.7, p ¼ .041. Areliable interaction between the two groups ofparticipants and performance mental age waspresent, F(1, 23) ¼ 4.6, p ¼ .043, although themain effect of performance mental age was narrowlynonsignificant, F(1, 23) ¼ 4.00, p ¼ .058.

In summary, the results indicate that partici-pants in the HFA group did not perform moreaccurately than those in the control group. Ifanything, the small effects present in theANCOVA analysis point to an advantage for thecontrol group—effects that are in the oppositedirection to those predicted by the reduced per-ceptual similarity hypothesis.

Figure 4 shows the classification of the six novelrectangles in the test phase as indexed by the pro-portion of A responses. Results were subjected toan ANOVA with rectangle as a repeated measureand group as a between-subjects factor. Therewere no reliable effects involving group (Fs , 1).

Model fitting

The results from the test phase were modelledusing the GCM (Nosofsky, 1986). This modelassumes that exemplars are represented as pointsin multidimensional psychological space, withexemplars from a particular category tending tocluster together because they generally havesimilar values on each dimension. The coordinates

of the exemplars are the psychological coordinatesderived from the MDS analysis of similarityratings.

The distance, dij, between two items Xi and Xj

is defined by the equation

dij ¼X

m

wmjxim � x jmjr

" #1=r

(1)

where xim and xjm are the coordinates of the twoitems on dimension m. The r parameter dictatesthe metric of the space. If r is equal to 1 thenthe distance is determined by summing the differ-ences between the coordinates of the two objectson each dimension, whereas if r is equal to 2(Euclidean space) then distance is simply theshortest straight line between the two points.The similarity, sij, between Xi and Xj is thendefined as an exponential function of the distancebetween them:

si, j ¼ exp (� c � Di, j ) (2)

where c is a scaling parameter that determines howeasy the individual objects are to discriminate.Figure 5 shows how similarity varies as a function

Figure 4. Proportion of ‘A’ responses for the rectangles presented in

the test phase only. Error bars are the standard errors for each

stimulus, within developmental group.

Figure 5. Similarity as a function of distance and the value of the c

parameter. At high values of c, objects are less similar than at low

values.



of c and the distance between two items asdetermined by the coordinates of the object inpsychological space. When c is high, the functiondrops off sharply so that at a given distanceapart, two items are relatively dissimilar and canbe easily discriminated. Conversely, when c islow, the same two items appear more similar andare relatively difficult to discriminate. The c para-meter is generally a free parameter estimated byminimizing the difference between the model’spredictions and an individual participant’sresponses. Put in these terms, the c parametercan be seen as a useful way of measuring a partici-pant’s sense of perceptual similarity. As such, thereduced perceptual similarity hypothesis wouldpredict increased c parameters in the HFA grouprelative to controls.

The final stage of the model involves assigningthe item to a particular category. In the GCM, theprobability that the object, Xi, is assigned to aparticular category Ck is given by the sum of thesimilarities between Xi and each item in Ck

divided by the summed similarity of Xi, to allitems in memory.

p(Ck j Xi) ¼

Pj[k sijPNj¼1 sij

(3)

In summary, the algorithm first calculates thedistance between the unknown item and theknown exemplars using Equation 1, then trans-forms these distances into similarities usingEquation 2, and finally calculates the probabilitythat the object belongs in a particular categoryusing Equation 3.

The GCM was fitted to the data by adjustingthe c parameter value to minimize the summedsquare error between each participant’s responsesand the model predictions, resulting in a single cparameter value for each participant. Becausethere were no a priori assumptions regarding themost appropriate MDS solution, or the most

appropriate GCM distance metric, this analysiswas performed for all possible combinations ofthese factors.

We first used the MDS solutions as the set ofexemplar coordinates for the GCM. All threesolutions were tested, using both r ¼ 1 and r ¼ 2,which resulted in a value of R2 that variedbetween .56 and .70. However, there was noevidence for any group differences in c parametervalues. In fact, the GCM demonstrated the bestfits to the data using the objective coordinates ofthe rectangles (i.e., their heights and widths)rather than psychological coordinates derivedfrom MDS. With the GCM distance metric, r,set to 1, the R2 value was .74, and when r wasset to 2, the R2 value was .75. Figure 6 showsthe c parameters for each participant with r ¼ 2(results were similar with r ¼ 1).6 The reducedperceptual similarity hypothesis predicted thatthe HFA group would have higher c parametersthan those in the control group. In fact, the differ-ence was in the other direction, although notreliably so, U 0(17) ¼ 161, p ¼ .29. Furthermore,all the participants in the HFA group appear tohave c parameters well within the range of thecontrol participants, indicating that individualdifferences are not obscuring a group effect.

Figure 6. The c parameter scores after fitting the GCM to the

categorization response, using the objective MDS coordinates.

6 One participant in the control group had an extremely high c value (the maximum possible value tested by our optimization

algorithm). This was because he had a near-perfect score for the training items during the testing phase, which the GCM fits

best by using an infinitely high c parameter value.

BOTT ET AL.


These results do not provide evidence in support ofthe reduced perceptual similarity hypothesis.

One aspect of our results that might seem unusualis that we obtained better model fits with the objec-tive solution than with the MDS solutions. Webelieve this stems from two factors of our design.First, because we were using a clinical sample,there was likely to be considerable individual vari-ation in the similarity judgements. This wouldhave resulted in a noisy set of coordinates from theINDSCAL analysis. If the “true” perceptual rep-resentation resembled the objective coordinate setmore than the results of the INDSCAL analysis,then a better fit would be expected from the GCMwith the objective set. Second, the GCM was mod-elled on responses that took place after the trainingphase of the experiment. If there were participantswho chose to rely on only one dimension duringthe similarity ratings task, they would have had tochange their representational coordinates duringtraining, in order to learn the category structure.Thus, the true perceptual space may well havechanged from that described by the MDS analysisto the objective space by the time participantsmade their responses during the testing phase.

Discussion

The current study investigated the hypothesisthat autism is associated with a reduced sense ofperceptual similarity. High-functioning individ-uals with autism and nonautistic controls firstperformed a similarity ratings task, designed todetermine the psychological coordinates of therectangle stimuli that were used in the subsequentcategorization task. Analysis by MDS revealed atrend suggesting that individuals in the HFAgroup represented the stimuli on fewer psycho-logical dimensions than did controls, althoughthis pattern was not significant. Participants thenperformed a categorization task in which theylearnt to classify rectangles into two arbitrarilydefined categories. Individuals with autism tookreliably longer to learn the category structureduring the training phase of this task. However,the main focus of this study was performance onthe test phase of this task in which participants

were required to categorize the learned stimuliwithout feedback and generalize their responsesto stimuli that had not previously been categor-ized. As expected, performance was relativelypoor on an exception item that was similar tomembers of the opposite category but, contraryto predictions, there was no evidence that thiseffect was reduced among members of the HFAgroup. Moreover, generalization of responses tonovel stimuli was similar in both groups.

Similarity ratings taskThe main objective of the similarity ratings taskwas to provide psychological coordinates for mod-elling of the categorization task. Nevertheless, theresults of the MDS analysis suggested potentiallyinteresting differences between individuals withand without autism in terms of their initialrepresentations of the stimuli. Specifically, partici-pants in the HFA group showed a trend towardshigher w-scores in the INDSCAL analysis, indi-cating that they might be basing their ratings ofthe stimuli on fewer dimensions than were con-trols. However, given that the effect was non-significant, this result should of course be treatedwith caution. Furthermore, the current data donot demonstrate that the control participantsencoded the data using more than one dimension;a finding that would be a necessary precursor forusing the w-score analysis as evidence that partici-pants with autism were representing the data onfewer dimensions than were controls. Moreover,the current study was not specifically designed totest the hypothesis that individuals representstimuli on fewer dimensions than were controls,and, consequently, all interpretation of the MDSanalysis is post hoc. Future experiments couldinvestigate this issue using the same methodology,although we would recommend collecting moredata points per participant to reduce within-participant noise and using high-dimensionalstimuli, rather than rectangles, to test the inte-gration ability of people with autism more fully.

Category learningAlthough category learning was not the mainfocus of this study, we note that the training



phase of the categorization task provided the firstreported investigation of the process of category-learning abilities of people with autism undercontrolled conditions. Perhaps unsurprisingly,given that individuals with autism appear to havedifficulties forming categories, the HFA grouptook significantly longer than the control groupto learn the category structure. Advocates of thereduced perceptual similarity hypothesis may betempted to explain this in terms of difficulties innoticing the similarities between category exem-plars. This would imply that for the HFA group,the costs of not being able to group togethersimilar rectangles outweighed the benefits oflearning the exception item relatively easily.However, both groups found that the exceptionitem was reliably more difficult to learn than theother rectangles, and there were no differences inthe rate in which different groups learned thedifferent types of rectangle. Thus, we found noevidence to suggest that a reduced perceptualsimilarity hindered learning of the categorystructure.

The fact that the exception item produced mosterrors during training means that participants wererequired to change their initial responses to theexception item in order to learn the category struc-ture. There is considerable evidence that peoplewith autism have specific difficulties on tests ofexecutive function when they are required tochange the dimensions of a stimulus to whichthey are attending (e.g., Ciesielski & Harris,1997; Hughes, Russell, & Robbins, 1994;Ozonoff & Jensen, 1999), and, arguably, suchdifficulties could explain the relatively slowlearning of the category structure. However, it isimportant to note that, whereas tests of executivefunctioning require participants to inhibitresponses that have previously been reinforced,in our task the correct response (and feedback) tothe exception item was always the same. Thus, ifexecutive deficits are to explain poor categorylearning then one must assume that individualsin the HFA group had difficulty inhibiting theexception item response because of the compe-tition from the similar rectangles. This suggestionappears plausible but rests on the assumption that

individuals in the HFA group are sensitive to thesimilarity of the surrounding items—an assump-tion that is at odds with the reduced perceptualsimilarity hypothesis.

Finally, it is possible that participants in theHFA group started the categorization task withinappropriate assumptions about the best way ofrepresenting the rectangles, and this hinderedtheir learning because they had to change therepresentation during the learning process. Thispossibility is compatible with the idea, discussedabove, that individuals in the HFA group initiallyrepresented the stimuli on fewer dimensions thandid controls. Clearly, further research with para-digms specifically designed to investigate categorylearning will be necessary to distinguish betweenthese different explanations for slow categorylearning in autism.

Categorization and generalizationAs expected, performance in the testing phase wasreliably poorer on the exception item than on theother rectangles, despite the fact that participantshad learnt to correctly classify the exception itemin the training phase of the experiment. Thispresumably was a result of interference fromsurrounding rectangles (see Figure 1). However,the reduced similarity hypothesis predicted thatthose in the HFA group would perform moreaccurately than controls on this exception itembecause they would be less influenced by thesurrounding rectangles. In the event, there wasno significant advantage for the HFA group.

The results of the test phase were modelledusing the GCM (Nosofsky, 1986), whichallowed us take into account responses to allthe items, rather than just the exception item,and provided us with a measure of individualparticipant performance. The reduced similarityhypothesis predicted that individuals in the HFAgroup would have higher scaling parameters(i.e., c parameters) than would the controls.Again, however, the trend was in the oppositedirection—there were no differences betweenthe groups, and no individual participant in theHFA group differed from the sample of controls.

BOTT ET AL.


Overall, therefore, the results failed to supportthe reduced perceptual similarity hypothesis andcontrast with a number of other recent findings.There are a number of potential explanations forthis. First, it might be argued that there wasinsufficient statistical power in the design.However, participants were tested 40 times eachon the exception item with little evidence ofceiling or floor effects that might have maskedan advantage for the HFA group. Furthermore,confidence intervals for performance on the excep-tion item indicated that any effect of reduced per-ceptual similarity is likely to be very small. Asecond potential criticism concerns the matchingof controls: The mean age of the HFA groupwas higher than that of the control group, andperformance mental age was lower than that ofthe control group. However, there was no evidencefor a correlation between age and classification ofthe exception item, and controlling for perform-ance mental age by covariation resulted in anadvantage for the control group on the exceptionitem. This latter result is in the reverse directionto the reduced perceptual similarity hypothesis.

A third and more theoretical concern mightbe that exemplar models are not appropriate toapply to this experimental design and that it isinvalid to assume, for example, that performanceon the exception item is determined by perceptualsimilarity, or that the results can be modelledwith the GCM. The most plausible alternativemodel would be some form of rule-plus-exceptionmodel (e.g., Erickson & Kruschke, 1998;Nosofsky & Palmeri, 1998), which would takethe form of a rule (e.g., “respond ‘A’ if therectangle is in some corner of the space, ‘B’ if itis in another corner”), but with some rectanglesremembered on an individual basis. However,these models still require an aspect of generaliza-tion around the exception item to account forthe fact that it is remembered significantly lesswell than other old items during the testingphase (both in the control participants and inthe HFA group). The task therefore assessesperceptual similarity even under the assumptionsof a different model. Consequently, even ifthe modelling of the data using the GCM is not

appropriate, the standard inferential statisticsused to analyse responses to the exceptionitem still provide evidence against reducedsimilarity.

A further issue concerns the generalizability ofthe current results. The complexity of the taskand the large number of trials that had to be com-pleted combined to ensure that it was only possibleto test high-functioning adults with autism. Onepossibility is that reduced perceptual similarity isonly found in younger or less able individualswith autism, and this could potentially accountfor the discrepancy between the current resultsand the reduced prototype effect demonstratedby the children with autism in the study conductedby Klinger and Dawson (1995, 2001). While thispossibility cannot be firmly ruled out, our resultsdo strongly suggest that findings from otherstudies with high-functioning adults withautism, such as the reduced transfer effect notedby Plaisted et al. (1998b) and the relatively goodperformance on the embedded figures taskreported by Jolliffe and Baron-Cohen (1997),cannot be explained purely in terms of reducedperceptual similarity, as has been previouslyargued (Plaisted, 2001; Plaisted et al., 1998b).

Finally, we note that our findings are at oddswith those of Bonnel et al. (2003), who recentlyfound evidence for enhanced sensitivity to pitchin people with HFA—a result consistent withthe reduced perceptual similarity hypothesis. Oneobvious difference between the studies is the typeof stimuli used. Thus, it is possible that reducedperceptual similarity in autism is restricted to theauditory domain. A second potentially crucialdifference is in the dimensionality of the stimuli;the current study employed multidimensionalrectangles, whereas Bonnel et al. (2003) usedstimuli that varied on a single dimension. It ispossible that difficulties in integrating dimensionsin the current study could mask otherwise superiordiscrimination abilities. However, it is importantto note that by the time our participants hadcompleted the training phase of our experiment,they had succeeded in integrating the multi-dimensional stimuli. Clearly, further researchusing simple and complex stimuli from different



modalities is required to determine the strengthsand weaknesses in categorization and discrimi-nation abilities in autism.

The results of the current study provided noevidence to support the hypothesis that autism isassociated with a reduced sense of perceptual simi-larity. Instead, we propose the novel hypothesisthat individuals with autism have a tendency torepresent objects on fewer dimensions than dotypically developing individuals, as we suggestedabove when discussing the similarity ratings task.This “reduced dimensions” hypothesis couldexplain the reduced prototype effect in autismreported by Klinger and Dawson (1995, 2001).It was suggested that individuals with autism inthat experiment were not able to choose themost representative exemplar from the categorybecause they had not coded the dimensions thatdefined the prototype. This situation may havearisen because children with autism chose tocode the objects on the fewest dimensions possibleto perform the training task—that is, on thebetween-category dimension—whereas controlschose to represent the objects more completely.The reduced dimensions hypothesis could alsoexplain why the HFA group required moreblocks to learn the category structure than did con-trols in this experiment—the HFA group wouldhave had to change their perceptual space inorder to learn to classify the rectangles, whereasthe controls were already attending to the appro-priate dimensions (see Palmeri & Nosofsky,2001, for evidence that the perceptual space canbe different after a categorization task). Morespeculatively, the hypothesis could also explainsome of the difficulties in generalization in every-day situations that Plaisted (2001) has attributedto reduced perceptual similarity. Objects and situ-ations in real life are typically complex and multi-dimensional. If one focuses on a very limitedaspect of a stimulus (such as the colour of anobject) then there is a high probability that thiswill be a superficial rather than a defining featureof the object. Subsequent encounters with othersuperficially different members of the same cat-egory will then be treated as an entirely novelexperience.

GENERAL CONCLUSIONS

The experiments presented in this paper weredesigned to test the hypothesis that people withautism have reduced perceptual similarity. Wefound no evidence in support of this hypothesis.In addition to our principal finding, we havemade several other contributions. First, we haveintroduced a precise definition of reduced percep-tual similarity based on computational modelsfrom the study of categorization. Second, wehave provided evidence that people with autismmay have difficultly in learning categories.Third, we have suggested a novel hypothesis con-cerning the cause of some of the generalizationdifficulties in autism: that of coding objects onfewer dimensions than controls in circumstancesin which such a strategy can be effective. Theresults of the current study are consistent withthis hypothesis but it clearly demands furtherinvestigation.

Original manuscript received 13 May 2004

Accepted revision received 1 October 2004

PrEview proof published online day month year

REFERENCES

American Psychiatric Association. (1994). Diagnostic

and statistical manual of mental disorders (4th ed.).Washington, DC: Author.

Bonnel, A., Mottron, L., Peretz, I., Trudel, M.,Gallun, E., & Bonnel, A.-M. (2003). Enhancedpitch sensitivity in individuals with autism: A signaldetection analysis. Journal of Cognitive Neuroscience,

15, 226–235.Brian, J. A., & Bryson, S. E. (1996). Disembedding per-

formance and recognition memory in autism/PDD.Journal of Child Psychology and Psychiatry, 37,865–872.

Carroll, J. D., & Chang, J. J. (1970). Analysis ofindividual differences in multidimensional scalingvia an n-way generalization of “Eckart–Young”decomposition. Psychometrika, 35, 283–319.

Ciesielski, K. T., & Harris, R. J. (1997). Factors relatedto performance failure on executive tasks in autism.Child Neuropsychology, 3, 1–12.

BOTT ET AL.


Erickson, M. A., & Kruschke, J. K. (1998). Rulesand exemplars in category learning. Journal of

Experimental Psychology: General, 127, 107–140.Frith, U. (1989). Autism: Explaining the enigma. Oxford:

Blackwell Science.Frith, U., & Happe, F. (1994). Autism: Beyond “theory

of mind”. Cognition, 50, 115–132.Happe, F. (1999a). Autism: Cognitive deficit or cogni-

tive style? Trends in Cognitive Science, 3, 216–222.Happe, F. (1999b). Understanding assets and deficits in

autism: Why success is more interesting than failure.Psychologist, 12, 540–546.

Hayes, B. K., & Taplin, J. E. (1992). Developmentalchanges in categorization processes: Knowledge andsimilarity-based models of categorization. Journal of

Experimental Child Psychology, 54, 188–212.Hayes, B. K., & Taplin, J. E. (1993a). Developmental

differences in the use of prototype and exemplar-specific information. Journal of Experimental Child

Psychology, 55, 329–352.Hayes, B. K., & Taplin, J. E. (1993b). The development

of conceptual knowledge in children with mentalretardation. American Journal on Mental Retardation,

98, 293–303.Howell, D. C. (1997). Statistical methods for psychology

(4th ed.). Belmont, CA: Wadswoth PublishingCompany.

Hughes, C., Russell, J., & Robbins, T. R. (1994).Evidence for executive dysfunction in autism.Neuropsychologia, 32, 477–492.

Jolliffe, T., & Baron-Cohen, S. (1997). Are people withautism and Asperger syndrome faster than normal onthe embedded figures test? Journal of Child Psychology

and Psychiatry, 38, 527–534.Klinger, L. G., & Dawson, G. (1995). A fresh look at

categorization abilities in persons with autism. InE. Schopler &G. B.Mesibov (Eds.), Learning and cog-

nition in autism. Current issues in autism (pp. 119–136).New York: Plenum Press.

Klinger, L. G., & Dawson, G. (2001). Prototype for-mation in autism. Development and Psychopathology,

13, 111–124.Kruskal, J. B., & Wish, M. (1978). Multidimensional

scaling. Beverly Hills, CA: Sage Publications.Lee, M. D. (2001). Determining the dimensionality of

multidimensional scaling representations for cogni-tive modeling. Journal of Mathematical Psychology,

45, 149–166.MacCallum, R. C. (1976). Transformation of a three-

mode multidimensional scaling solution toINDSCAL form. Psychometrika, 41, 385–400.

Medin, D. L., & Schaffer, M. M. (1978). Contexttheory of classification learning. Psychological

Review, 85, 207–238.Mottron, L., & Burack, J. A. (2001). Enhanced percep-

tual functioning in the development of autism.In J. A. Burack, T. Charman, N. Yirmiya, &P. R. Zelazo (Eds.), The development of autism:

Perspectives from theory and research (pp. 131–148).Hillsdale, NJ: Erlbaum.

Nosofsky, R. M. (1986). Attention, similarity, and theidentification–categorization relationship. Journal

of Experimental Psychology: General, 115, 39–57.Nosofsky, R. M. (1987). Attention and learning

processes in the identification and categorization ofintegral stimuli. Journal of Experimental Psychology:

Learning, Memory, and Cognition, 13, 87–108.Nosofsky, R. M. (1990). Relations between exem-

plar-similarity and likelihood models of classifi-cation. Journal of Mathematical Psychology, 34,393–418.

Nosofsky, R. M. (1992). Exemplars, prototypes, andsimilarity rules. In A. M. Healy, S. F. Kosslyn, &R. M. Shiffrin (Eds.), From learning theory to connec-

tionist theory: Essays in honor of William K. Estes

(pp. 149–167). Hillsdale, NJ: Lawrence ErlbaumAssociates, Inc.

Nosofsky, R. M., & Palmeri, T. J. (1996). Learning toclassify integral-dimension stimuli. Psychonomic

Bulletin & Review, 3, 222–226.Nosofsky, R. M., & Palmeri, T. J. (1998). A rule-plus-

exception model for classifying objects in continuous-dimension spaces. Psychonomic Bulletin & Review, 5,345–369.

O’Riordan, M., & Plaisted, K. (2001). Enhanceddiscrimination in autism. Quarterly Journal of

Experimental Psychology, 54A, 961–979.O’Riordan, M. A., Plaisted, K. C., Driver, J., & Baron-

Cohen, S. (2001). Superior visual search in autism.Journal of Experimental Psychology: Human

Perception and Performance, 27, 719–730.Ozonoff, S., & Jensen, J. (1999). Brief report: Specific

executive function profiles in three neurodevelop-mental disorders. Journal of Autism and

Developmental Disorders, 29, 171–177.Ozonoff, S., & Miller, J. N. (1995). Teaching theory of

mind: A new approach to social skills training forindividuals with autism. Journal of Autism and

Developmental Disorders, 25, 415–433.Palmeri, T. J., & Nosofsky, R. M. (2001). Central

tendencies, extreme points, and prototype enhance-ment effects in ill-defined perceptual categorization.



Quarterly Journal of Experimental Psychology, 54A,197–235.

Plaisted, K. (2001). Reduced generalization in autism: Analternative to weak central coherence. In J. A. Burack,T. Charman, N. Yirmiya, & P. R. Zelazo (Eds.), The

development of autism: Perspectives from theory and

research (pp. 149–169). Hillsdale, NJ: Erlbaum.Plaisted, K., O’Riordan,M., & Baron-Cohen, S. (1998a).

Enhanced discrimination of novel, highly similarstimuli by adults with autism during a perceptuallearning task. Journal of Child Psychology and

Psychiatry, 39, 765–775.Plaisted, K., O’Riordan, M., & Baron-Cohen, S.

(1998b). Enhanced visual search for a conjunctivetarget in autism. A research note. Journal of Child

Psychology and Psychiatry, 39, 777–783.Shah, A., & Frith, U. (1983). An islet of ability in autis-

tic children: A research. Journal of Child Psychology

and Psychiatry, 24, 613–620.Shah, A., & Frith, U. (1993). Why do autistic individ-

uals show superior performance on the block designtask? Journal of Child Psychology and Psychiatry, 34,1351–1364.

Shepard, R. (1987). Towards a universal law ofgeneralization for psychological science. Science,

237, 1317–1323.

Swettenham, J. (1996). Can children be taughtto understand false belief using computers?Journal of Child Psychology and Psychiatry, 37,157–165.

Tversky, A. (1977). Features of similarity. Psychological

Review, 84, 327–352.Tymchuk, A. J., Simmons, J. Q., & Neafsey, S. (1977).

Intellectual characteristics of adolescent childhoodpsychotics with high verbal ability. Journal of

Mental Deficiency Research, 21, 133–138.Wechsler, D. (1986). Wechsler adult intelligence

scale—revised, UK edition. Windsor, UK: NFER-Nelson.

Wing, L. (1996). The autism spectrum: A guide for parents

and professionals. London: Constable.Wolfe, J. M., Cave, K. R., & Franzel, S. L. (1989).

Guided search: An alternative to the featureintegration model for visual search. Journal of

Experimental Psychology: Human Perception and

Performance, 15, 419–433.World Health Organization. (1992). International

classification of diseases and disorders (10th ed.).Geneva, Switzerland: Author.

Young, F. W., & Harris, D. F. (1994). Multi-dimensional scaling. In M. J. Norusis (Ed.), SPSS

professional statistics 6.1. Chicago: SPSS, Inc.

BOTT ET AL.


Documents

Perceptual similarity in autism