ConceptsandCategories
Percepts,
B. Burns(Editor)
B.V. All rightsreserved.
@ 1992ElsevierSciencePublishers
PerceptualSimilarityand
ConceptualStructure
LINDA B. SMITH
DIANA HEISE
Indiana University
I.
II.
m.
ry.
V.
VI.
Introduction
The CaseAgainstPerceptualCategorization
Categorization
A. From Basicto Superordinate
VersusEssences
B. PerceptualCategorization
C. CategoryInduction
In Defenseof PerceptualSimilarity
A. PerceptualSimilarityIs Dynamic
B. Knowledgeand PerceptualSimilarity
1. The Caseof Eyesand Texture
2. Th'e Caseof Lexical Form-Class
C. ConceptuallyRelevantPerceptualProperties
PerceptualSimilarityand CausalTheories
A. Perceptualand ConceptualStructureAre Not the Same
B. Perceptualand ConceptualStructureAre CausallyRelated
C. How ConceptualStructureDependson PerceptualStructure
Structureand Process
Conclusions
References
I. INTRODUCTION
of categoriesand conceptscontrast the
Many recent discussions
pick up a
perceptualand conceptualbasesfor categories.Thesediscussions
persistenttheme in developmental
theory. This theme posits a trend from
234
LindaB. SmirhandDianaHeise
perceptionto conceptionand it hasbeenplayedthroughoutthe worksof Piaget
(1929),Vygotsky(193411962)
and Werner(1948). It is echoedby Flaveil (L970),
Wohlwill (1962),and Bruner and Olver (1963)and more recentlyby Gentner
(1989)and Keil (1989). The ideais that childrenshift from perceptually-bound
representations
of objecs that are globaland holisticto onesthat are principled
and articulatedalongabstractdimensions.So,for example,Flavell(1970)wrote
that conceptualdevelopmentgrows from "equivalences
based on the more
concreteand immediatelygivenperceptual,
situational,
and functionalattributes
of objectsto equivalences
of a more abstract,verbal-conceptual
sort." Since
developmentis directional,this themeimpliesthat conceptionis in some way
"betterthan"perception.
Many developmentalists
argueagainsta shiftfrom perceptualcategories
to conceptualones (e.g.,Brown, 1990;Mandler & Bauer, 1988;Wellman &
Gelman,1988). They arguethat categories
are conceptually
structuredfrom the
very beginning.But even for thesetheorists,the researchagendais definedin
termsof categories
structuredby "mereappearance"
versuscategories
with a "rich
theoreticalstructurethat goesbeyondsuperficial
similarity."In thesediscussions,
perceptualand conceptualcategorization
are presentedas mutuallyexclusive
processes.
Moreover,conceptualcategorization
is presented
assmartand mature
and perceptualcategorization
is presented
as deficient,and to be abandoned.
We believethatthisperception-versus-conceptionorganization
of research
works to the detrimentof empiricaland theoreticalprogress.This definitionof
the researchagenda dismissesthe contributionof perceptualstructure to
conceptualstructureand doesnot studyit seriously.Our purposein this chapter
is to defendperception. We counterthe perception-versus-conception
approach
by arguingthat the contributionof perceptualcategories
to conceptual
categories
is considerable
and continuousthroughoutdevelopment.Conceptualstructure
does not replaceor even overrideperceptualcategories.Instead,conceptual
structureis basedon perceptualstructure.
This view that perceptionis the groundingforcefor conception
hasbeen
argued recentlyby Johnson(1983)and lakoff (1987). This view is also gaining
attentionin the cognitivedevelopment
literature(seeGelman,1990;Gentner&
Rattermann,1990;Mandler,1990). In this chapter,we specifically
considerthe
role of perceptualsimilarityin categorydevelopment.Our defenseof perceptual
similarityrestson threepoints:
1. Perceptualsimilarityis dynamic.It varieswith the attributesattended
to.
Structure
andConceptual
Similarity
Perceptual
235
2. Experiencein perceivingthe relationsin the world influencesattention
and perceptualsimilarity.Becauseof this,perceptualsimilaritymoves
about in meaningfulways.
canbe abstractand canconvey-- indeedbe the
3. Perceptualcategories
sourceof -- conceptuallyrelevantknowledge.
Before presentingthe defense,we assessthe caseagainstperceptual
similaritv.
II. THB CASE AGAINST PERCEPTUAL CATEGORIZATION
contrasts
an "animal
Keil (1989)buildingon thewritingsof Quine(1.970)
"conceptual
similarityspace."Animal similarityaccording
similarityspace"and a
to Keil, is original similarity. It is uninterpretedfeature counts and lists of
correlationsbetween perceptualfeatures. According to Keil, this animal or
tabulationof informationwe getwhenwe look
originalsimilarityis the atheoretical
and thereforestupid. Fishand whales
at the world. It is constantand unchanging
we
what objectslook like is not always
of
what
know.
But
look alike regardless
what they really are. Thus, developmentis away from "the immediate,the
subjective,animalsenseof similarityto the remotersenseof similarirydetermined
and posis and constructs"
(Quine,1977).In thisview,the
by scientifichypotheses
onewe argueagainst,perceptualsimilarityis replacedby belief systemscomposed
properties.
of causallyrelated featuresand nonperceptual
We offer three examplesof how ideassuchas theseorganizethe current
researchon categoriesand concepts.
A. From Basic to Superordinate Categories
Medin and Ortony (1989;see also,Medin et al., 1990) suggestthat
perceptualsimilarity althoughnot a very intelligentmechanismmay preciselyfit
her first categories.Researchon perceptual
the needsof the infant in discovering
development(e.g.,Kemler Nelson,1990;Smith,1989)indicatesthat very young
objectswholisticallyacrossall dimensions
childrenoftencomparemultidimensional
at once. Medin and Ortonyarguethat thisoriginalsimilarityspacefits the child's
task of making a first partitioningof the world'sobjectsinto categories.
LindaB. Smithand DianaHeise
Medin and Ortony point to the fit betweenoverall perceptualsimilarity
and the basic level caregoriesdescribed by Rosch & Mervis (1975). The
categories dog, cat, car, house, and bird are basic level categoriesand these
categoriesdo seem to be organizedby perceptualsimilarity or in Rosch'sterms
by family resemblance. Dogs look alike,or at least most dogs seem to be mostly
alike in most of their perceptualproperties(see,Rosch,et al. I976: Biederman.
reSs).
In contrastto basiclevelcategories,
superordinate
categories
seemto be
structuredby a few generalabstractproperties. Animals,for example,differ
widely and are not as a group perceptually
like eachother. Moreov-er,
by the
consensus
view, there are no specifiableperceptualpropertiesthat all (or even
most) animalspossess.The consensus
view is that superordinate
categories
are
structuredby nonperceptual
properties(carey,19g5;Gelman,19gg;Mandler&
Bauer,1988;Markman,1939).
There is muchdevelopmental
dataconsistent
with a trendfrom basicto
superordinatecategories. In classification
tasks,young children form spatial
groupingsof basiclevel categoriessuchas shoesversusdogsbeforethey form
superordinategroupings(Roschet al., 1976). Namesfor basiccategoriesare
learned fast and considerablybefore names for superordinatecategories
(MacNamara,1982). Thesedata and arguments
suggest
a developmental
trend
that proceedsfrom Quine's"animal"similaritywhichgivesus basiCcategories
to
conceptualsimilarirywhichgivesus superordinate
categories.
However,in their full complexity,
thedevelopmental
datado not squarely
fit a unidirectionaltrend from basic to superordinatecategories(see, Mervis,
1987). An early sensitivityto superordinatecategorystructurecan be seen in
young children's overgeneralizationsin word learning. Young children's
overgeneralizations
honor superordinate
categorydistinctionsin that children
sometimesmistakenlycall cows"doggie"but do not mistakenlycall cars "doggie"
(see waxman, 1980,for relevantdata). Mandler and Bauer (19gg;see also,
Mandler, Bauer & McDonough,in press)have shownthat even prelinguistic
childrenare sensitiveto superordinate
categories.Theyobserved12-month-olds'
"categorizations"in a free play
situation. Their measure of categorization
consistedof the sequenceof touchingobjecs. Same categoryobjectswere
touchedin rapid succession
more frequentlythan differentcaregoryobjecs (see
Sugarman,1983).By thismeasure,Mandlerand Bauerfoundthat infantsreadily
madesuperordinate
classifications
of the sort "dogand horseversuscar." Indeed,
in their task,16-month-olds
madesuperordinate
classifications
more readilythan
theymadebasiclevelones(e.g.,poodleand collieversushorse).
Perceptual
Similarity
andConceptual
Structure
231
Mandlerand Bauertakefor grantedthat perceptualpropertiesaloneare
not enoughto organizesuperordinate
categories.Under this assumption,
their
data providesupportfor earlyconceptually
basedcategorization.Mandlerand
Bauer point to the categoriesformedby the children-- chicken,fish,cow,and
turtleversusmotorcycle,
airplane,van,andtrainengine-- as proofthatperceptual
similaritywas not involved. They point to the way the children playedwith the
toys -- makingthe animalswalk and talk and makingthe vehiclesspeedabout
with "vroom-vroom"sounds as evidencefor nascentconceptualstructure.
Perceptualsimilarity,they conclude,does not control categorization
even at its
beginning.
We are not sure that thiSconclusion
is warranted.Perceptualstructure
may well be sufficientto explainMandler and Bauer'sdata. Specifically,there
maybe heretoforeundiscovered
perceptualpropertiesthat distinguishanimalsand
vehicles,and babiesmightwell be sensitiveto theseproperties.Of course,to find
theseproperties,we have to look. Later, in this chapter,we will presentsome
evidenceof suchpropertiesthat we foundwhenwe did look.
B. Perceptual Categorization versus Essences
Keil (1989)recentlyarguedthat perceptualsimilarityis not sufficientto
explain the psychologicalstructureof even basic level categories. A thought
experimentmakesthe point. Considera cow. Is it still a cow if somepsychologist
coversit in sheepskin,paintsit purple,cutsoff a leg,and addsmoosehorns? Keil
(1989)hasshownthat peoplestronglybelievethat naturallyoccurringobjectsdo
not changetheir identitywithchangesin their perceptualappearance-- no marter
how severe. Peoplepossessbeliefsabout naturallyoccurringobjecs (or natural
kinds) that attribute their identityto the processes
of their origin or rhar imbue
them with an undefinedessence.
In this research,Keil affirms the distinctionbetween the perceptual
processeswe use to identifyobjectsin the world and our conceptsof them (see
Smith& Medin, L981,for more on this importantdistinction).The idea is that
when we comeacrosssomeunknownobject,we may useperceptualsimilarityto
-- our conceptsof what
identi$ it and assignit to a category.But our categories
it means to be a particularkind of object -- are much more than lisrs of
perceptualor even functionalproperties. Peoplehave organizedsetsof causal
beliefs,or theories,that distinguish
kindsof objectsaccordingto whethertheyare
naturallyoccurringor manmade,
aliveor not,terrestrialor water livingand so on
(seeCarey,1985;S. Gelman,1988;& Keil, 1989).Accordingto Keil, it is these
238
LindaB. Smithand DianaHeise
theoriesabout the causalconnections
betweenperceptualpropertiesand their
originsthat form the conceptualcoreof a category.And thus,alterationsin the
"mereappearance"
of an objectin ways(e.g.,painting)that do not violatecentral
theoreticalbeliefs(e.g.,an objectis a cow if its motherwasa cow),do not alter
the object'sidentity. This theoreticalcore is sometimesreferred to as the
category's"essence"in recognitionof the relation betweenthese views and
in philosophy(seeKeil, 1989;and also Murphy & Medin,
nominalessentialism
and
Medin
&
Ortony,
1989).
1985
Causaltheoriesincreasein complexityand scientificaccuracywith age
with this age-relatedgrowthin
(e.g.,Carey,1985;Gelman,1988). Consistent
conceptualstructure,Keil (1989)hasshownthat the beliefin "categoryessences"
alsoincreases
with age. Three-and 4-year-olds
often
that transcendappearance
in appearance.Thus,
maintainthat an object'sidentitydoeschangewith changes
trend
is described
as movingfrom perceptualsimilarity
again,the developmental
to conceptualstructure.
There are a number of ongoingdisputesin this literature. They all
for any
is enoughto explainperformance
concernwhetherperceptualappearance
developmental
level.
For
example,
Wellman
and
Gelman
(1988)
stimuliat any
"essences"
too and ignoreappearance
arguethat very youngchildrenbelievein
in domainsin which they have sufficientknowledge. Carey,Gelman,and Keil
that transcend
perceptualappearance
are characteristic
eacharguethat essences
of peoples'beliefsaboutnaturallyoccurringobjectsbut not manmadeobjects.By
these arguments,artifacts are characterizedby an impoverishedtheoretical
structure and thus their categorystructuresare more controlledby superficial
of naturalkinds. Greerand'
perceptualsimilarifythan arti the categorystructures
in press)arguein contrast
Sera(1990;seealso,Mandler,Bauerand McDonough,
that there is no principleddistinctionbetweenartifacs and naturalkinds. They
report that peoplehavecausaltheoriesaboutthe internalpartsand workingsof
complexartifactssuch as computersand radiosand will maintainthat these
despiteradicalchangesin outwardappearance.
objects'identitiesgo unchanged
are all definedin termsof
The researchagenda,the theoreticaldisputes,
perceptionversusconception.Is this the most usefuldefinitionof the research
question? When we come uponan objectin the world,whenwe comeface-toto determine
perceptualprocesses
facewith a dog or a chair,we usesophisticated
what it is (e.g.Biederman,1985).But if we are to take the dataand claimsabout
essences
seriously,our conceptsof objecs are largelyunrelatedto and largely
unaffectedby the perceptualsystemsthat haveevolvedto recognizereal objects
in the real world.
Perceptual
Similarityand Conceptual
Structure
239
C. CategoryInduction
When asked to make inductiveinferencesfrom what is known about
objecs in one categoryto objecs in another,conceptualstructureand not
perceptualsimilari{ is again said to be the preeminentforce. The typical
solutions.The state-of-the
empiricalstudyagainpits perceptualand conceptual
by
a
thought
experiment
derivedfrom the work of
field is cogentlysummarized
Careyand Gelman. Imaginea real monkey,a mechanicalmonkey,and a real
snake. We tell you that the real monkeyhas a duogleenaninside. Is the
mechanicalmonkeyor the real snakemore likely to also have a duogleenan
monkeylookalike,but the real monkey
inside?A real monkeyand a mechanical
and real snake share "deeper" properties that place them in the same
superordinatecategories.
Carey(1985)and GelmanandMarkman(1986,1987,seealso,Gelman,
1988)haveshownthat people(sometimesas youngas three yearsof age) make
inductiveinferencesabout the nonperceptualinsidesof objecs in accordwith
In twoimportantstudies,
GelmanandMarkman
superordinatecategorystructure.
with
pictures
(1986,1987)presented
children
triadsof
as in Figure7.1. The crow
and bat were black and the flamingo was colored pink. By Gelman and
Markman'sanalysis,the crow and bat were similaroverallwhereasthe flamingo
and the crow belongto the sameconceptualcategory.The childrenwere told the
namesof eachobject(bird,bird, and bat) and thenwere told that the crow laid
eggs. They were askedwhich other object,bat or flamingo,also laid eggs.
Previousresearchhas shownthat young children often freely classiff
objectslike theseby perceptualsimilarity(Tversky,1985).Gelmanand Markman
found, however,that childrenmade categoryinductionsalong conceptuallines.
Gelmanand Markman concludedfrom theseresultsthat wordsaccessconceptual
structureand that categories-- evenfor very youngchildren-- are organizedby
a rich theoreticalstructurethat goesbeyondappearance.We have no quarrel
with this conclusion.
We do, however,question the ancillary assumptionthat perceptual
properties have no role in conceptualstructure. Gelman and Markman's
assumptionthat perceptiondoes not matter is highlightedby their designof
experimentsand choice of stimuli. Gelman and Markman chosestimuli that
distortedthe similarityrelationsthat actuallyexistin the world. The bat and crow
are drawn and colored to emphasizethe few featuresthat real bats and crows
havein commonand to de-emphasize
the manyreal differencesthat existbetween
Linda B. Smith and Diana Heise
Figure7.1. A sampletriadof stimuliusedby Gelman& Markman(19g6).
bats and crowsand to also de-emphasize
the many real similaritiesthat exist
betweencrowsand flamingos.Gelmanand Markmanselectedtheir stimulifor
a goodreason-- to providea strongtestof the hypothesis
that childrenwoulduse
conceptual
structureevenin thecontextof strongcountering
perceptualsimilarity.
Their goalwasto choosebetweentwo hypotheses
-- conceptual
structurecontrols
performanceor perceptualstructurecontrolsperformance. Our questionis
whetherthis is a theoretically
sensiblegoal.
Considertwo more thoughtexperiments
thatare variantson Gelmanand.
Markman'sstudy. In the first thoughtexperiment,the triad of stimuli looks like
that in Figure 7.2. Objects A and B are clearly most similar overall. The
"conceptual"informationthat
is providedis that insideA is a bird, insideB is a
bat, and insidec is anotherbird. Now, if the objectin A layseggs,which other
object,B or C, is alsolikely to lay eggs?Giventhe impoverishedcharacrerof the
perceptualinformation,one wouldbe wise to generalize
accordingto the labels
provided. This experimentif actuallyconducted,would providea perfectlygood
test of whetherchildrencan ignoreperceptualinformation. Like Gelmanand
Markman's original studies,perceptualsimilarityis pitted againstconceptual
similarity; and also like Gelman and Markman'sstimuli, the perceptualstimuli
provide little information about how the objectsrealtylook. Although the stimuli
in Figure7.2 mightprovidean adequatetestof the perception-versus-conception
Perceptual
Similarity
and Conceptual
Structure
241
Figure7.2. Threestimulifor a thought
experimenr.
hypothesis,
they providea poor test of the usesof perceptualand conceptual
informationin real categoryand conceptformation.
Now considera secondversionof Gelmanand Markman'sstudy. In this
version,the stimuliconsistof a real livingflamingo,a livingcrow,and a livingbat.
The objectsare not named. Nothingis saidabout them exceptthe fact that one
(the crow) lays eggs. In the richnessof the real perceptualinformation,there
seemslittle doubt that eggJayingwouldbe generalizedfrom one bird to the next.
Of course,thesestimuli are inappropriateto the task of determiningwhether
children useperc€ptualor conceptualinformationin categoryinductionbecause
the rich perceptualstructureand the conceptualstructureconverge. However,
thesestimuli might be usefulif our researchgoal is to determinehow children
form the categories
they do. The point is that determininghow real categories
are formedrequiresan understanding
of perceptualsimilarity.
III.
IN DEFENSE OF PERCEPTUAL SIMILARITY
Defendingthe role of perceptualsimilarity in conceptualstructure
requiresa clearstatementof just what perceptionand conceptionare. We take
242
LindaB. Smithand DianaHeise
perceptionto referto thestructureof immediateexperience.We takeconceptual
structureto refer to knowledgethat is (or can be madeto be) explicit-- that is
knowledgethat is easilytalkedaboutandthoughtabout.Thesedefinitionsappear
to the sameas thoseof Keil (1989)and Mandler(1990). However,wherewe
differ is that we believeperceivedsimilarityembodiesand reflectsmuchimplicit
knowledge.Many theorists(e.g.Keil, 1989)write aboutperceptualsimilarityas
if it were givenin the stimulus.This naiverealismassumesthat the perceived
just as the physical
similarirybetweennvo objectsis constantand unchanging
measuremensof the propertiesof objectsare constantand unchanging.But
perceivedsimilarityis the resultof psychological
processes
andit is highlyvariable.
The centralprocessthatchanges
perceptual
similarityis attention.The perceived
similarityof n'voobjecs changeswith changesin selectiveattentionto specific
perceptualproperties.
The importanceof attentionin categorization
has been recognizedby
other investigators.
MurphyandMedin(1985)persuasively
pointedout thatwhen
people categorizeobjectsand reasonabout categories,
they shift attentionin
principledwaysamongsetsof perceptualfeatures.For example,peopleseemto
use different perceptualpropertieswhen they classifyan object as an animal
versuswhentheyclassifythat sameobjectas a pet. Murphyand Medin (seealso,
Keil, 1989)suggestthat people'sexplicitcausaltheoriesor beliefsaboutobjects
organizeand drive the selectionof relevantperceptualfeatures.The ideais that
althoughperceptionis involved,perceptionalone -- animal similarity-- is not
enough. Accordingto Murphyand Medin,the real forcein categorization
is the
conceptualstructurethat organizesand interpretsperception.That is, by their
view,conceptualstructureis the causeand changing
attentionweightson specific
dimensionsis the effect.
Our viewis quitetheopposite:thedynamicnatureof perceptualsimilariry
is a causalforcein the development
beliefs.
of conceptual
A. Perceptual Similarity is Dynamic
Among thosewho studyperceptualsimilarity,there is one agreedupon
fact, the perceptualsimilarity betweenany two objectsvaries considerably.
Perceptualsimilarityvaries with the attributesattendedto (Shepard,L964;
Nosofsky,1984). The dynamicnatureof similariryis evidentin the empirical
and in formal theoriesof
researchon perceptionand perceptualcategorization
similarity(Goldstone,
Medin,& Gentner,in press;Nosofsky,1984;and Tversky,
1,977). In formal theories,similarity is some function of some weighted
Structure
Similarity
andC,onceptual
Perceptual
243
combinationof featuresand attributes. As the feature weightschange,so
similaritychanges.
as in
For example,in Smith's(1989)modelof perceptualclassification
Nosofsky(1984)and Shepard(1987),similarityis calculatedas an exponential
space. The
decayfunctionof the distancebetweenstimuliin the psychological
similaritybetweentrvo objects,O; &rd Q, then is
.,)"u
e
sij
Distance,d;;,is defined as the sum of the weighteddimensionaldifferences
I\/
r
dij
/-t
WKlOik -
Ojkl
K - 7
where O;pminus O;r is the differencebetweenobjectsi and j on dimension k, N
is the number of dimensions,W1 is the weight given dimensionK and O < W. <
1.00and
N
E
w
k
K - t
1-.00.
The present point is that this formula definesperceptual similarity;
perceptualsimilaritys just this (or somecalculationlike it). And if perceptual
similarity is some weighted combinationof dimensionalsimilarities,then
perceptual similarity necessarilyvaries with the magnitudeof the difference
betweenstimuli on the dimensionsand with the dimensionweights.
Considerthe triad of stimulion the left handsideof Figure7.3. These
stimuli are representedin termsof their coordinateson two varyingdimensions.
Object A is identicalto ObjectB on dimensionX but differsfrom B considerably
on dimensionY. Object B differsfrom ObjectC by a small amount on both
dimensions.Usingthe equationsabove,the right handsideof Figure7.3 shows
theperceivedsimilairlesbetweenobjectsA and B, objectsB and C, and objects
A and C as the dimensionweighs changefrom perfect selectiveattentionto
244
Linda B. Smith and Diana Heise
1.00
.90
.80
.70
'=
.60
.50
.40
a .30
.20
.10
0
s
.E
q
q)
E
cttmenston x
Wx= 50
v1/x=1.m
Attention
valuesin a
in termsof theircoordinate
Figure7.3. Left -- A triadof stimulirepresented
pairsin the triadasa functionof
between
space.Right-- the similarity
2-dimensional
to dimension
x.
selective
attention
dimensionY (W, - 1.00,W* = 0) throughequalweightingof the trvodimensions,
(W* = .50, Wy : .50) to perfect selectiveattentionto dimensionX given
dimensionX (W- : 1.00,Wv : 0). As is apparent,the absolutevalueof the
with changesin the feature
similarityof ObjectsA and B changesconsiderably
weights. Moreover,thereis a dramaticchangein the relativesimilaritiesof AB,
AC and BC. Which two objectsin the triad are the most similar dependson
which featuresare attendedto.
If we extendthesenotionsto Markmanand Gelman'sstimuliin Figure
7.1, a perceptualinterpretationof their resultsis possible. Under one set of
overallshapeand color, the crow is
featureweights,say one that emphasized
perceptuallymore similar to the bat than the flamingo. Under anotherset of
beaksor feet,the crowandthe
featureweights,for example,onethatemphasized
percepnally
pair. Importantly, then, a
similar
more
be
the
flamingo might
demonstrationthat children perceivea bat and crow to be similar in a
taskand a crowand flamingoto be similarin the categoryinduction
classification
taskneednot meanthat childrenshiftedfrom perceptualsimilariryto conceptual
similaritywhen asked to make inductions. They may only have shifted the
similaritymayhaveplayedthe keyrole in
perceptualfeatureweights.Perceptual
both judgments.
Nosofsky(1984) representsthe effectsof shiftingfeatureweightson
space
of the psychologicalsimilarity
perceptualsimilarityin termsof thestretching
in one direction or another.Borrowingthis idea, we representhypothetical
Perceptual
Similarity
and Conceptual
Structure
245
Figure7.4. A theoretical
depiction
of the similarity
relationsbetweenGelmanand
Markman'sstimuliundertwodifferentsetsof featureweights.
similarityrelationsbetweenGelmanand Markman'sstimuliin Figure7.4. For
easeof discussion,
themany-dimensionalspace
is compressed
into twodimensions.
The X axis representssomecombinationof variationalong overallshapeand
color and the Y axisrepresents
variationon somecombinationof headand feet
features. The three dots representthe locationin that spaceof Gelmanand
Markman'scrow,bat, and flamingo.The distances
betweenthe dos represent
their perceptualsimilarityto eachother. Figure7.4 top illustratesthe casefor an
equalweightingof all the dimensions;
for theseparticulardrawings,the bat and
246
Linda B. Smith and Diana Heise
crow are the most perceptuallysimilar objects among the three. Figure 7.4
bottom representsthe similarity relationswhen attention is focussedon the head
and the feet. The space is stretched so that the crow and flamingo are now
closer,that is, more similar than the bat and crow.
One mightarguethatsuchstretchedsimilarityspacesare not the original
similarityor "animal"similaritymeantby Keil and Quine. Thesetheoristsmeant
"raw" similarity unconfounded
by knowledgeor whatevermechanismpushes
featureweightsand similarityaround. Consistent
with theseviews,Smith(1939)
defined raw similariryas the similarityrelationsthat result from the equal
weightingof all perceptualfeatures.This overallsimilarityseemsrighrlyraw in
that it is the defaultsimilarity,the one that dominatesperceptionwheneverlimits
are placedon performanceand/orthere is no previousexperience.Even for
adults, an equal weightingof dimensionsdominateswhen processingtime is
limited,whenstimuliare complex,and whenthereis lack of relevantknowledge
(seeJ. D. Smith& Kemler-Nelson,
1984;Smith,1981).This is alsothe similariry
that dominates2- and 3-year-olds'perceptual categorization
in most task
circumstances
(smith, 1989).
Thus,the equalweighting
of all dimensions
doesseemto be the pointof
origin for perceptualsimilarityin the senseof beingthe zerostateof the system.
However,overallsimilarityis neithermechanistically
nor developmentally
distinct
from a stretchedsimilarityspace. There is but one system,one manner of
calculatingperceptualsimilarity-- the one that is mathematically
describedby the
equations(seeSmith,1989,for an empirical demonstration
of this claim). The
possibilityof changingfeatureweightsis not an add-onto somemore primitive
form of similarity. Separatefeaturesand changingfeatureweightsare inherent
in the very natureand processof perceptualsimilarity.This claimis supported
by the facts of animal perceptionand by the factsof perceptualsimilarityin
human infancy.
Shiftsin featureweightsare characteristic
of perceptualsimilarityin nonhuman animals- i.e.,in real "animalsimilarity"(Mackintosh,1965;Sutherland,
7964). The phenomenonof cue-blocking(or overshadowing)
provides a
particularlygood example.Consideran illustrativecue-blocking
experiment.In
originallearning,the organismis trainedto respondto red objectsfor a reward
but not to respondto blue ones. The shapesof the objectsvary and do not
matter in the response-reward
contingencies.
If animalsformedconceptual
rules,
"red
a reasonable
onewouldbe
wins,blueloses,andshapedoesn'tmatter."If we
Perceptual
Similarity
and Conceptual
Structure
247
Figure7.5. An example
of a stretched
similarity
spacein whichcoloris weighted
more
thanshapein the calculation
of similarity.
stoppedthe experimentalprocedurehere and testedthe organismwith novel
shapesandcolors,the patternof response
wouldfit the rule. The organismwould
respondto all and only red objectsand ignoretheir shape. Figure7.5 depictsthe
effectsof learningon the perceptualsimilarityspace. Originallearningtakesthe
organismfrom a similarityspacein whichcolorsand shapesmatrer equally(top)
to one in which color is (virtually)all that matters(bottom).
The phenomenonofcue-blockingshows
that thisinterpretationis correct.
In a cue-blockingexperiment,a redundancy
phaseis added. In this phase,shape
is correlatedwith color. For example,all the red objecs are now trianglesand
all the blue ones are circles. The organismis given some lengthyset of these
redundancytrials. The questionis whether,giventhe previousoriginallearning,
the animal noticesthe redundancy.If during originallearning,the organism
learnedto selectivelyattendto color and ignoreshapesuchthat the similarity
spacehasbeenseverelystretchedin one direction,thenthe organismshouldnot
notice the redundancy.The resultsshowthat in taskssuchas theseanimalsdo
not learnaboutthe addedredundanry.After the redundanryphase,
the organism
respondsto all .andonly red objecsand ignoresshapejust as if the redundanry
248
Linda B. Smithand DianaHeise
training had not occurred. Selectiveattentionis clearly a force in ,'animal
similarity." The similarityspacefor pigeonsand rats is easilystretched.
Another literature that suggeststhat "raw" similarirymay rarely be
realizedis the infant habituationliterature.Considerwhat happensto perceived
similarityin a prototypichabituation
study(for real andsimilarstudies,
see,e.g.,
Bornstein,1985). In a typicalexperiment,infantsmight be shownrepeated
examplesof objecs that are alike on some dimensionbut vary on a second
dimension. For example,theymightbe shownred squares,red circles,and red
triangles. Lookingtime is recordeduntil the babiesno longerlook muchar the
red squares,circles,and triangles.Then, the infantsare shownnew objects-either one that differson the previously
constant(or relevant)dimension(e.g.,a
green square)or one that offers a new value on the previouslyvarying(or
irrelevant)dimension(e.g.,a red cross). The standardresult,given repeated
examplesof the samecolor duringhabituation,is that infantslook more to the
novelcolor (greensquare)thanthe novelform (red cross)duringdishabituation.
Importantly,if infantswere shownrepeatedexamplesof the sameshapebut
varyingcolorsduring habituation,
theywouldlook more at the novelshapethan
the novelcolor.
Presumably,themechanism
behindsuchhabituation-dishab
ituationeffects
is short-termchangesin featuresweights.If the habituation
stimuliare all red but
with varyingshapes,
the commonredness
of the habituationstimuli
mustsomehow
stretchthe similarityspacealongthecolordimension(or thevaryingshapescause
a shrinkageof the spacealongthe shapedimension).The habituationphase
causesa change in the perceivedsimilarity of objects; in our example,the
habituationphasemade differencesin color count more in the calculationof
perceptualsimilaritythan differencesin form. Originalsimilarity,rhe perceptual
similarity of infanry, is inherentlydynamic;it moves about becauseof the
mechanismthroughwhichit is computed.
The mechanism
is one in whichfeature
weightsmayvary from contextto context.The useof habituation
as a methodto
study infant perceptiondependson the reality of malleablefeatureweightsin
psychological
similariry.
If perceptualsimilarityshifs so readilyin the artificiallaboratorycontexts
of operant conditioningand habituationexperimens,might not perceptual
similarityshiftsystematically
andmeanin{ultyin response
to thecontingenciesand
correlationsthat existbetweenperceptualfeaturesin the real world?
Perceptual
Similarityand ConceptuaI
Structure
249
Figure7.6.A distorted
region
of a similarity
space
asmightresultfromincreased
attention
to a particular
conjunction
of features
(therelevant
conjunction
is thecompressed
area).
B. Knowledge and Perceptual Similarity
The relevance
of perceptual
similarityfor conceptualstructure
stemsfrom
the role of real world experience
on featureweights.One kind of knowledgethat
pushesfeature weightsaround and stretchesthe similarityspace is implicit
knowledgeaboutrelationsbetweenperceptual
features.As Rosch(1,973)
argued,
perceptualfeaturesdo not vary orthogonallyin the world. They come in causally
relatedclusters.Birdswith web feet tend to havebills. Objectswith dog-likefeet
tend to havedog-likeheads.Evidencefrom laboratoryexperiments
indicatesthat
both adults (Medin et al., 1982)and older infants(Younger& cohen, 1983;
Younger, 1990) are sensitiveto such correlations. This empirical evidence
indicates that experiencewith correlationscilusesincreasedattention to the
combinationsof featuresthat enter into correlations.
Figure 7.6 representshow sucha correlationmight organizea similariry
space. The grid linesrepresent"equal"physicaldistances.Perceptualsimilarity
is representedby real Euclideandistancein the space. In the depictedspace,
objectsthat possessa particularconjunctionof valuesare closer togetheror
perceptuallymore similarthanobjectsthat possess
other particularconjunctions
of values. Thus the compactedupperright of the figuremight be the regionof
the stimulusspacein whichobjectswith dog-likefeet and dog-llkeheadsfall. If
this figurewere an accuratedepictionof perceivedsimilarity,it wouldmeanthat
the similaritybetweenobjectswith bothdog-feetanddog-heads
is grearerthanthe
?.50
Linda B. Smith and Diana Heise
EXEIIPLAR
l-L
\.ro
I
ril
m
lo-\
sHApEcHAr{cE
.s3LJ\
lb-3
.76
Lru
TEXTURE
CHANGE .la
lo--3
l
t
.82
SIZECHAIIGE. 8 0
-
l
t
|
t
L.tL.l
Figure7.7. Stimuliand meanpropostion
nameextensions
to teststimuliin Jones,Smith
& Landau(1990).
similarity between objectswith both dog-feetand pig-heads.Do young children,
in their everydayinteractionsin the world, learn about such correlationsin ways
that systematicallyand to good purpose distort the similarity space?
1. The Caseof Eyesand Texture
Recently,Jones,Smith,and [,andau(1990)discovereda co-relationeffect
in children'snovelword extensions
that suggests
a powerfulrole for combinations
of perceptualfeaturesin naturalcategoryformation.Joneset al. foundthat the
perceivedsimilarityof objectsfor 36-month-olds
changeddramaticallywith the
addition of a constantperceptualproperty. In the experiment,childrenwere
showna novelobjectand it wasnamed. For example,theymightbe shownthe
objectatthe top left of Figure7.7andtoldthatitwasa Dax. Childrenwerethen
askedwhat other testobjectswerealsoDaxes.As shownin the left columnof the
figure, one test object differedfrom the exemplaronly in shape,one only in
texture,and one onlyin size. Whiletheexemplar(the experimenter-named)
Dax
remainedin view,childrenwereaskedseparately
whethereach one of the test
Perceptual
Similarity
andConceptual
Structure
Z5I
objectswas also a Dax. The proportionof timesthe childrenagreedthat the
exemplar'snamewas alsothe nameof the testobjectis givennext to eachtest
object. As the resultsattest,thesechildrencalled the test objectsby the same
nameasthe exemplarwhentheobjectwasidenticalin shapeto the exemplarobject.
Shapechangesfrom the exemplarmatteredconsiderably
more than changesin
textureand size.
The right side of Figure7.7 showsthe stimuli for a second(benweensubjects)condition. In this condition,toy eyeswere affixed to eachobject.The
addition of a constantstimulusproperty(the eyes)radicallychangedchildren's
judgmentsas is evidentin the proportionsof nameextensions.Wheneyeswere
added,the exemplar'snamewasextendedto a testobjectonly if the test object
was the sameshapeand sametertureas the exemplar.
Theseresultsdemonstrate
thesynergistic
relationthat mustexistbetween
perceptualpropertiesand knowledge.The presenceof eyesactivateschildren's
knowledgeabout kinds of categories.That knowledgethen guidesattention,
changingthe weightsaccordedto the variousdimensions,and thus shifting the
perceptualsimilaritiesamongthe objects.
The kind of co-relationthat existsbetweeneyesand textureis different
from the correlationsbetweenpropertiesthat havebeenstudiedby Medin et al.
and Younger and Cohen. The kind of correlationstudied by previous
investigators
is betweenspecificproperties,for example,betweenfeathers,webbed
feet, and bills. The kind of co-relationthat underlies3-year-oldsnovel word
extensionsdoesnot seemto be betweeneyesand a specifictexture.
In the world, if havingeyesis correlatedwith a specifictexture,it is with
one that is soft and pliable. In the Joneset al. study,the Dax exemplarwasmade
of wood. In one experiment,the texture-change
test item was a sponge,in
another,it was a soft cloth. If the presenceof eyeshad signalledthe importance
of particular textures,then the childrenshouldhave been more likely to accept
these softer Daxes in the Eyes conditionthan in the No-eyescondition even
thoughthey had a differenttexturethan the exemplar. However,the resultsare
just the oppositeof this specific-feature-correlation
hypothesis.Childrenin the
No-eyesconditionreadilycalledthesesoft objectsDaxes. However,childrenin
Eyesconditionrejectedtheseitems. They requiredobjecs with eyesto havethe
same wooden texture as the exemplarto be called a Dax. Apparently, the
presenceof eyes indicatedthat texture-- whateverthat texture might be -- is
criterial in categorization.
252
Linda B. Smithand DianaHeise
This kind of co-relation-- berweena specificproperty (eyes)and a
dimensionof variation(texture)-- shouldprovea powerfulaid in learningnew
categorystructures,
one that wouldsupplement
the useof correlations
bet'ween
specificproperties(e.g.,wingsand feathers).A learnedcorrelationbetweenthe
specificpropertiesof wingsand feathersis a localdistortionof thesimilariryspace
as shown in Figure 7.6. Sucha local distortionwould help the categorizer
distinguish
objectsthat do or do not belongto an already-learned
category.
A learnedcorrelation
benveen
a cueanddimension
of variation(likethat
betweeneyes and texture),however,causesa more widespreaddistortionof
similarity. The entire similariryspacewould be stretchedin one direction
exaggerating
texturedifferences
in the contextof eyesrelativeto the contextof
no eyes. Suchan expansionof the spaceshouldhelp the child organizemultiple
categoriesof birds, and bats and walruses,and of real monkeys,mechanical
monkeys,and snakesbecausethe objects in thesecategoriesdiffer from one
anotherin the textureof their surfaces. The idea, then, is that the similarity
spacereflectsthe factthatsurfacetexturedifferences
mattermorefor categorizing
all objectswith eyesthansuchdifferences
matterfor categorizing
objectswithout
eyes. A generalexpansion
of the similarityspacealongthe texturedimensionfor
objectswith eyeswould increasethe similarig of real crowsand flamingos(which
both havefeatheredsurfaces)
anddecrease
the similarityof crowsand bats(since
crowsare featheredand bats are furry) relativeto eyelesscounterparts.
If we imaginemultiplesof local and dimension-wide
distortionsof the
-similarity space distortionsresultingfrom real-worldcorrelationsbetween
specificproperties,from co-relations
betweenmaterialkind and kind of motion,
eyesand texture,eyesand kind of motion,shapeand motion,and so on -- then '
what emergesis a bumpy and irregularsimilarityspacethat organizes
itself into
multiplesof categoiesat multiplelevelsin contextdependentways.
We suggestthat the resultof spendingtime in the world -- of lookingat
it, hearingit, and feelingit -- is a structuredand contextdependentsimilarity
space.The resultis a dynamicsimilarityspacethat fis the structureof the world.
Rosch (1973) suggestedthat many natural categoriesare given in the world in
correlatedfeaturessets. Her proposalseemsright. The reason that the
thousands
of categoriesnamedby concretenounsemergeso rapidlyin the first
three yearsof life and are conceptually
well understood
by four or five yearsof
age is that they are givenin a contextdependentperceptualsimilarity.
This idea that perceptualsimilarityreflectrthe co-relationsbetween
perceptualpropertiesas theyexistin theworld presumes
a particularmechanism.
Perceptual
Similarityand Conceptual
Structure
253
We presumewhat most formal theoriesof perceptualsimilarityand perceptual
categoruationposit -- that learned atheoreticalassociationsshift the features
waysthat reflectthe structure
weightsand the computedsimilarities
in systematic
of theworld (seeNosofsky,1986;Gluck& Bower,1988;Medin & Schaffer,1978).
There is another explanationof the effect of eyes on 3-year-olds'
-- one built on represented
causaltheories.For example,3-yearcategorizations
structuresof the sort: objects
olds'judgmentscouldbe organizedby knowledge
classhavethesame
with eyesare naturalkinds,membersof thesamenatural-kind
genetic structure, the same genetic structureproducesthe same material
substance;the same material substanceproducesthe same surfacetexture;
set of beliefs
therefore,texturemattersfor objecswith eyes.Sucha complicated
seemsimplausibleto usbut moreto the point,we haveevidencethat suchbeliefs
are not necessary
for the co-relationeffectsreportedby Joneset al.
We (Heise& Smith,1990)havedemonstrated
thatthesamephenomenon
observedin children'snovelword extensions
occursin adults'learningof arbitrary
categories.Billman(1989)alsohasreportedcompellingevidencethat makesthe
point. In our study,adultslearnednvodifferentclassification
systemsin a single
experiment.For example,subjecsmightbe givenred bugsand blue bugsand
learn to classiff the objecs into four groupssuch that the red bugs were
partitionedinto npo subgroupsby shape(circularvs. dropshape)and the blue
bugswere partitionedinto nvo subgroups
by numberof legs(2 versus4). Given
aduls proclivity for selectiveattentionand the formation of criterial property
categories(Medin,Wattenmakerand Hampson,1987),one mightexpectadults
in this taskto learn only to attendto shapeand/ornumberof legs. Transfertasks,
however,showedthat adultslearneda moregeneralco-relationbetweencolor and
the other dimensions.In the transfertask,adultswere givennewsetsof bugsand
askedto freely classi$ them into groups. Unlike typicaltransfertasks,then,we
did not ask how well adults put new items into the just learnedcategories,we
askedhow well they formed new categoriesof new items -- on their own -- and
without instruction.
Adults'transfer performancein this free classification
task suggeststhey
"rule"
had learnedan implicit
of the sort: red bugsare distinguished
into subkinds
by their shapebut blue bugsare distinguished
into subkindsby their numberof
legs. Giventhe trainingdescribed
above,aduls classified
red bugsby their shape
evenwhen the novelbugshad shapesneverseenbeforeand they classifiedthe
blue bugsby their numberof legsevenwhenthe numbersof legswere different
from thosein the learningphase.Thisacquisition
of a contextdependentshiftin
254
Linda B. Smith and Diana Heise
feature weightswas obtainedwith various transpositionsof the cuing and transfer
dimensions.
These resultssuggestthat context-dependent
distortionsof the similarity
space are easily set up. Moreover, the adults in these experimentshad no preexperimental naive causaltheoriesabout how one dimensionshould be related to
another in classiffing cartoon bugs. The subjectsalso do not appear to have
developed such causal theories in the courseof the experiment. Clearly, causal
theoriesor well-developedbelief systemsare not necessaryfor contextdependent
and useful shifts in attention to perceptualproperties.
2. The Caseof Lexical Form-Class
thatoneperceptual
Thecaseof eyesandtextureindicates
propertymay
influenceattentionto anotherperceptualpropertyand systematically
alter the
perceivedsimilaritiesof objects.But knowledge
of perceptualpropertiesis only
one forceon perceptualsimilariry.Other nonperceptual
forces(thoughprobably
not explicitknowledge)alsoplaya role. For example,youngchildren,shift their
attentionamongdimensionsin novelword extensiontasksas a functionof the
syntacticform classof the word.
In the Jones,Smith,and Landau(1990)study,(seealso,Landau,Smith,
& Jones,1988),youngchildren'sselectiveattentionto shapewith eyelessstimuli
and their attentionto shapeand texturewith eyedstimulioccurredonlyin a word
extensiontask. When childrenwere askedto freelyclassiffthe samestimuli,they
of size,shape,and texture.
showedno differentialattentionto the dimensions
The presenceof a novelword, therefore,organizedthe featureweightsand shifted
them from the defaultvalueof equalattentionto all dimensions.In thosestudies,
the novelword wasa countnoun.
Recently,Smith, Jones,and l,andau (1990) examined36-month-olds'
shiftingattentionto shapeandcolorwhenthe novelwordwasan adjectiveaswell
as whenit wasa countnoun. In one study,half the childrenwere presentedwith
an exemplarobjectand told it was"a dax"and half the childrenwere presented
with the sameexemplarobjectbut were told it was "a dax one." The exemplar
a zig-zagshapeand was coloreda glittery (and
was made of wood, possessed
of
highlyreflective)combinationof silverygold. The criticaltestitemsconsisted
(1) nvo uniqueobjectsthat were the sameshapebut differentcolor and (2) two
uniqueobjectsthatwere the samecolorbut differentshape.Childrenwereshown
eachtestobjectone at a time and asked whether it was a "dax"in the noun
PerceptualSimilarityand ConceptualStructure
255
Table 7.1. Mean proportion extensionsto critical test trials in the Adjective and Noun
conditionsin Smith, Jones & Landeau(1990).(Standarddeviationsare in parentheses.)
Noun
Same Color
Diff. Shape
Adjective
SameShape
Diff. Color
No
Cave
.20
(.23)
.85
(.11)
Cave
.29 \
(.22)
.94
(.08)
Same Color
Diff. Shape
.51
( .31)
-a
( .18)
SameShape
Diff. Color
.78 \
(.2e)
.27 1
(.20)
conditionor a "daxone" in the adjectivecondition.The exemplarwas in view
throughout the entireprocedure.
The resultsare givenin Table7.7in the row labelled"No cave".Children
in the nounconditionattendedto shapecallingsameshapeitemsa Dax; children
did not call samecolor but differentshapeitemsa Dax. Childrenalsoshoweda
weak shapebiasin the adjectivecondition.However,in the adjectivecondition,
individual children'sperformancesdiffered from each other. Most children
maintainedthat the same-shapeobjecs were "dax ones." But some children
believed same+olor objects were ndax ones" and some children responded
haphazardly.The shinycolor of the exemplarwas highlynoticeableand indeed
was spontaneously
commentedon by all the children.Nonetheless,
childrendid
not systematicallyinterpret the novel adjectiveas referring to this novel and
salientfeature.Theseresultsshowthat the presence
of a novelword,particularly
a count noun,organizesattentionto shape.
The resultsof a secondexperiment
showedthatchildren'sattentionin the
adjectiveconditioncould be as highlyorganizedas their attentionin the noun
condition -- if additionalforces directed their attention among the varying
dimensions.The procedurein thissecondexperiment
wasidenticalto that in the
first with the exceptionthat the stimuliwere presentedinsidea smalldark cave.
The child and experimenterlookedthroughthe openingof the caveto view the
exemplarand eachtestitem. A spotlightin the caveilluminatedthe stimuli. The
caveand spotlighthad the effectof heightening
the alreadysalientglitterycolor.
256
LindaB. SmithandDianaHeise
The resultsfrom this secondexperimentare alsogivenin Table 7.1 in the row
labelled"Cave".Again,in the nouncondition,childrenextendedthe novelword
to new itemsaccordingto shapealone. But in the adjectivecondition,children
attendedto color and calledobjects"daxones"only if they possessed
the same
glitterycolor as the exemplar.
Theseresultsshowthat both lexicalknowledge
and localforcesconspire
to control children'sattention to dimensionsand the perceivedsimilarityof
objects. Presumably,
in the courseof learninglanguage,
childrenhave learned
that shapeis the principledeterminerof membershipin the categorieslabelled
by count nouns(seeLandauet al., 1988). The act of namingobjectsthusmay
serye as a cue to stretch the similarityspace along the shape dimension.
Apparently,however,adjectivecategoriesare not well constrainedby kind of
propertyand thus the contextof an adjectivedoesnot systematically
distortthe
similarityspace.But local,context-specific,
ad hoc forcescanwork synergistically
with the adjectivecontextto forcefullyorganizeattentionand therebyconstrain
possibleinterpretations
of the noveladjective.
We do not find these resultssurprising. They fatl right out of the
mechanismthat underliesperceptualsimilarity. However,sinceit seemsthat
manykindsof knowledgeshift featureweights-- associative
connections
between
perceptualproperties,knowledge
aboutsyntactic
formclasses,
localad hoceffects,
question
whether
some might
distortionsof similariryare truly perceprual
phenomenon.Is the shiftynatureof perceptualsimilaritytoo contaminated
by
(implicit) knowledgeto be rightly consideredperceptual? We believe there is
little to be gainedfrom pursuingthisquestion.The criticalpoint is that perceived
similarity is dynamicand shiftsin meaningfulways.
C. ConceptuallyRelevantPerceptualPrcperties
The powerof perceptualsimilarityand shiftingattentionamongfeatures
whenwe considerabstract
becomesmoreimpressive
perceptualproperties.There
are lots of perceptualpropertiesthat are not easyto talk about but that may
shapeour conceptionsof the world from very early in development.We are
thinkingof the perceptualpropertiesthatdistinguish
biologicalfrom nonbiological
motion (e.g.,Berthental,Proffitt,Spertner& Thomas,1985;R. Gelman,Spelke,
& Meck, 1983;Bullock,R. Gelman,& Baillageon,
1982),that allowus to predict
what novelobjectscanand cannotmovealone(Massey& R. Gelman,1988),and
to determinewhetheran actionby one objecton anotherwasintentional([rslie,
Structure
Perceptual
Similarity
andConceptual
257
L984;1988;Spelke,1982). There is a growingresearcheffort to discoversuch
conceptualcategories.
perceptualpropertiesthat mustunderlie,indeedorganize,
Briefly, we would like to considerone perceptualproperty that may
underlie the contrastbetweennaturallyoccurringobjectsand manufactured
andnaturallyoccurringobjecsdiffer
is thatmanufactured
objects.Our hypothesis
wayin theirsurfacegradients.Considerthe two starsshownat the
in a perceptible
top of Figure 7.8. They havethe sameglobalshape. But if one of theseis a
holidaystar and the othera livingstarfish,it is the one on the right
manufactured
that was madein a factoryand the one on the left that was foundat the beach.
Naturallyoccurringobjectstend to have an intricatesurfacegradientthat is
objects.
distinctfrom that of manufactured
Mandelbrot(1983)describedthe surfacesof naturalobjecs as fractal
curves.Thesesurfacesare self-similaracrosschangesin scale. Thus, if we
Figure 7.E. Two objectswith the sameglobalshapebut differentsurfacecradientsand
enlargements
of portionsof thoseobjects.
258
Linda B. Smithand DianaHeise
magnifiedthe starfish,we wouldseebumpson thesurface.If we magnifiedthose
bumps,we would see bumps on the bumps and so on. We may speculate,
followingMandelbrot,that the complexand seemingly
irregularsurfacegradient
of naturalobjectsstemsfrom the factthat theygrowandare causedby a multiple
of convergingforceswhoseeffectsaccruein time. In contrast,the scale-specific
structureof manufactured
objectspresumably
derivesfrom the factthat theywere
made at a particularscaleleveland at a particularpoint in time.
Are children sensitiveto the surfacegradient differencesbetween
naturallyoccurringand manufactured
objects?DianaHeise,SusanRiveraand
I have preliminaryevidencethat 12-month-old
childrenare. In this ongoing
research,the taskis preferentiallooking.In the baselineexperiment,the stimuli
were toy vehiclesand life-like model animals. Model animals,of course,are
and so it mightbe arguedthat theyare poor stimulifor a studyof
manufactured
the perceptualpropertiesaffordedby naturallyoccurringobjects.However,the
perceptualpropertiesthat makeonemodelanimalseemlife-likeandanothertoylike or robot-likeare presumablythe sameonesalongwhich naturallyoccurring
and (typical)manufactured
objectsdiffer.
a l o
z
ANIMAL
U
H
8
2
c
tr
-l
=
l-t
6
z
Y
xY
I
VEHICLE
r
E lx-lx
O .1,N.VEH
I YEH.VEH
(t VEH-AN
F IRST
FA}l
LAST FAH
TEST
TRIAT
Figure 7.9. Mean looking time on the first and last of 14 familiarizationtrials showing
(AN-VEHA/EH-AN)
Animals or Vehicles and on the test trial in the Between-category
and Within-category(AN-ANA/EH-VEH) conditionswhen the stimuli were intact toy
animalsand vehicles.
Structure
Similarity
andConceptual
Perceptual
259
the babieswere repeatedlyshownin alternationnvo
In the experiment,
toys and then after familiarizationwere showna noveltest toy. In nvo Betweenclass
Categoryconditions,the noveltest toy was from a differentsuperordinate
the noveltesttoy
conditions,
toys.In the Within-Category
than the familiarization
stimuli. For example,childrenin
was from the sameclassas the familiarization
conditionmight see a goat and cow during
the AnimaVBetrveen-Category
during
test. Children in the corresponding
boat
and
a
familiarization
seea goatandcowduringfamiliarization
Animal/Within-Categoryconditionwould
results
in Figure7.9 showclear evidencethat
test.
The
and an elephantduring
infantsdiscriminateanimalsfrom vehicles.Lookingtime duringtestgoesup more
conditions.
conditionsthanin the Within-Category
in the Benveen-Category
There are a numberof perceptualpropertiesalongwhich babiescould
(and probablydo) distinguishtoy animalsand toy vehiclesthat includeeyes,
surfacegradient.We examinedthe
mouths,globalshape,andby our hypothesis,
infants' ability to use surfacegradientcues in a subsequentexperimentby
removingall potentialcuesexceptsurfacegradientandaskingif babiescouldstill
distinguish animals from vehicles. Specifically,we cut up the toys into
2" x 2" pieces,removinganypieceswith eyes,mouths,or parLsof
approximately
10
rt
6
=
x
t{
u
7
tr
lrl
I
I
tr
at
o
o
2
Y
R
x
E ln-nx
O er-vBn
I vEH-vEH
I veu-rr
FIX,ST.FAM
I.AST FAM
TES"T
TRIAL
trialsshowing
Figure 7.10. Mean lookingtime on the first and lastof 8 familiarization
(AN-VEHA/EH-AN)
Animalsor Vehiclesand on the testtrial in the Between-category
thestimuliwerecutcondition(AN-ANA/EH-VEH)conditionswhen
andWithin-category
up toy animalsand vehicles.
260
Linda B. Smith and Diana Heise
faces (and all wheels from the vehicles). We separatelymounted the remaining
pieces of each toy in a circle for viewing by the infant and repeated the baseline
experimentwith thesecut-upstimuli. The resuls shown in Figure 7.10 indicate
that children discriminatedcut-up animalsfrom cut-up vehicles. Contrary to the
first experiment,children looked more at the within-category test stimulus than
the Benveen-category
stimulus. Although a preferencefor the familiar as opposed
to novel was shown with the "parts," lhe results, nonetheless,show babies
discriminateanimal texturesfrom vehicle textures. Thus, there do appear to be
perceptual properties that distinguishanimals from vehicles. Along at least one
perceptible property, surface gradient, animals and vehicles do not look alike.
Perhaps,then, the babiesin Mandler & Bauer'sstudyweresimplyclassiffingthe
objecs by what they looked like.
Our finding that babiesare sensitiveto a texturecue that may distinguish
animals from vehicleshas implicationsbeyond Mandler & Bauer's results. The
findings provide a useful starting point for thinking about the relation between
perceptual and conceptualstructure. The resuls show a perceptual dimension
along which many naturally occurring and manufactured objects may be
distinguishedand to which infantsare sensitive. Moreover, it is the nature of the
human apparatus for perceptual similarity to link up this dimension with cooccurring properties and in this way to selectivelyenhance attention to this
dimension in the stimuluscontextsin which it most matters -- for example,in the
context of stimuli with eyes,or stimuli that move in a certain biologicalway.
Our claim here is not that surfacegradient is a perfect or fool-proof cue
by which to distinguishnaturallyoccurringfrom manufacturedobjecs. There are
stonesso smooth from the continual action of the waves that they are as perfect
spheres as ball bearingsand there are talented artistswho can fool the eye with
their creations. Our idea, though, is that surfacegradient is an important force
in starting, maintaining,and groundingconceptualstructure to reality.
This concludesour defenseof perceptualsimilarity as a major player in
category development. Perceptualsimilarity is context dependent. It is shifted
about by implicit knowledgeof relationsbetweenperceptualproperties. And, the
perceptual properties that can be emphasizedor de-emphasizedin similarity
judgments may, like the presenceof eyesand like the surfacegradient differences
between animals and vehicles,be highly conceptuallyrelevant -- of a kind worth
developinga causaltheory to explain.
Structure
Perceptual
Similarityand Conceptual
26I
Iv. PERCEPTUAL SIMILARITY AND CAUSAL THBORIES
similarityare not,by our view,the samething.
andconceptual
Perceptual
similarity.But perceptualand
doesnot reduceto perceptual
Conceptualsimilarity
conceptualsimilarityare also not independent. We will considerfirst how
perceptualand conceptualsimilarityare different. Then,we will considerthe
benveenthe nvo.
causaldependencies
I
A. Perceptual and Conceptual Similarity are Not the Same
When we look at objects,what we perceivereflects much implicit
benveenperceptualfeatures,linguistic,
knowledge--- a myriadof associations
social,and physicalcontext.Perceptualsimilarityis thusdynamicand smartbut
structureis. Perceptualsimilarity
it is not smartin the samewaythat conceptual
and
explanations;
it
embodiescausalrelationsbut it
causal
beliefs
is not a set of
doesit not representthem. Thus,when childrenlook at objectswith eyesthey
in textureand their perceptionis smart
perceivethe similaritiesand differences
of why
and adaptive. But there is in the perceptionno causalunderstanding
texturesmatter for objecs with eyes.
distinguished
between
Rozin(1976)in hislandmarkpaperon intelligence
birds
to
migrate
and
welded implicit knowledge of the sort that allows
on ships.Rozin
explicitknowledgesuchas that usedby navigators
transportable
view
as
and
inflexible.
Our
of the implicit
characterized
implicit knowledge rigid
knowledgethat is embodiedin the dynamicsof perceptualsimilarirydoesnot fit
implicit knowledgeas beingweldedto
this description.Rozin alsocharacterized
for the knowledgeto showitself.
particularcontextualfactorsthat are necessary
would seemto be weldedin the
The knowledgebehindperceptualcategorization
propertiespresent,past
senseof beingtightlydeterminedby the actualperceptual
specificlearning,and context.Explicit knowledge,true intelligenceaccordingto
Rozin, is knowledgethat is more context free -- that may be voluntaily
structuremaybe
transportedacrosscontextsto newproblemspaces.Conceptual
smart in this way.
This distinctionbetween implicit perceptualstructure and explicit,
conceptualknowledgeis attestedto by the difficultyone has in
transportable,
similaritymay
consciously
controllingperceivedsimilarities.Whereasconceptual
be stronglyinfluencedby beingtold somenew fact,perceptualsimilaritymay be
movedalongonly a little by "meretalk." Beingtold that a purple,three-legged
beastwith moosehornsis a cow.doesn'tmake the unfortunateanimallook llke
2.62
LindaB. Smithand DianaHeise
a cow (although one might, if given that information by experts,claim to believe
the odd beast to be a cow despiteis looks). The separateness
of perceptionand
our easily talked about conceptualbeliefs is evident in linguistichedges(Lakoff,
1972) and statementssuch as, "It looks like a fish but it's really a mammal." This
obvious distinction between perceptual knowledge,which is not transparent to
consciousthoughtprocesses,
and explicitconceptualknowledgeprobablyunderlies
the temptationto empiricallypit perceptionagainstconceptionin developmental
experiments.
B. Perceptualand ConceptualStructuneaneCausallyRelated
The separatenessand different status of perceptual and conceptual
structure does not mean they are not causallyrelated. There are causalrelations
that go in both ways. Explicit conceptualknowledgecan push perception only so
far (no amount of information will convincea person looking at a spoon that it
is a tomato). However, conceptualknowledgecan cause individualsto seek out
perceptual dimensionsthat make senseof conceptualdistinctions.This effect of
conceptual knowledge on perception is aptly demonstratedin Heise's ongoing
dissertation research. In that research,children are presentedwith real world
correlatesof the picture stimuli usedby Gelman and Markman (1986). One triad
consists of a ball of yellow wool yarn, a ball of yellow acrylic yarn, and raw
unwoven wool. The nvo balls of yarn look very much alike. However when told
that the wool yarn and the raw wool are both wool, 5 year old children (and
adults) often look more closelyand say such things as "Yes, these two are less
shiny, that's how you can tell"). This example is telling on two counts. First, it
shows how a conceptual distinction may invite the search for a perceptual
distinction. Second,it shows how the human cognitivesystemattempts to keep
conceptualand perceptualstructuresin line with each other. Conceptualgrowth
does not causeus to abandonperceptionbut it may causeus to take another look.
Although explicit conceptualknowledgemay sometimesbe a causalforce
in perceptual comparison,the causalrelation bemveenperceptionand conception
goes mostly one way --- with perception the cause and conception the effect.
Perceptual structure ries our conceptual beliefs to reality; it is the mapping
function that takes our representationsand beliefs and gives them meaning.
Perceptual structure is the data that our conceptualtheoriesare about and just
as the data are the final arbiter of truth in science so is perception the
unassailableconstrainerof conceptsin cognitivedevelopment.
Structure
andConceptual
Perceptual
Similarity
263
Perceptualstructure is what conceptualbeliefs are about. The
and correlationsbetween perceptual
atheoretical,uninterpretedassociations
featuresmakes a bumpy terrain of the perceptuallandscape. Our explicit
conceptualbeliefs try to make conscioussenseof this terrain. Becauseour
senseof theworld,perceptualandconceptual
beliefsare aboutmaking
conceptual
structurewill generallyagree.
However,our explicitconceptualbeliefs,evenat maturity,probablydo
not do a very good job of makingexplicitthe implicit interrelationsand corelationsthat structureand organizeperception.Perceptualknowledgeis deep,
and probablymuchsmarterthan conceptual
highlyembedded
highlyentrenched,
knowledge(seeGentner& Rattermann,1990,for a similarargument)this is so
despitethe fact that the knowledgeembodiedin perceptionis severelylimited in
domains.The evidenceon howpeoplereadily
acrosscognitive
its transportabilify
the physicsis but one exampleof the
catchthrownballswithoutunderstanding
smartbut weldednatureof perceptualknowledge(see,e.g.,Bingham,Schmidt,
E. Rosenblum,1989). We expectthe samecontrastbenveen"smartperception"
and "lesssmart"conceptionis true for categories.It is perceptualknowledgethat
is deep and it is the conceptualstructurethat is superficialand inaccurateas it
dependson culturaland perhapsmistakenscientificbeliefs(seeJeyifous,1986;
Lakoff, L987;Putnam,1975for examples).
C. How Conceptual Structure Depends on Perceptual Structurc
In his recentbook,Keil (1989)examinesBoyd'sproposalabout causal
as a model for human conceptualstructure. The idea is that
homeostasis
are notjust clustersof correlatedproperties.Instead,thereare selsof
categories
clustered. The correlated
causallyimportantpropertiesthat are contingently
The lawsof physics,
biology,and
propertiesare not the resultsof happenstance.
behavior(the laws scientissseekto discover)are suchthat propertiesco-occur
(and vary together)becauseof causal"homeostasis."The presenceof some
of others.
propertiesfavorsor causesthe presence
structure
How is this structurein the world relatedto the psychological
of categories?It is informativehereto contrastour positionwith Keil'sproposal.
Following Quine, Keil considersan initial "originalsimilariry"state that is
by the
primitivesare represented
represented
as in Figure7.tLA. The perceptual
circles,triangles,squares,and ellipses. These primitivesare connectedby
associations
that are formedwhenspecificfeaturesco-occurin the
pretheoretical
accordingto Keil, is that causalbeliefs
world. What happenswith development,
264
LindaB. Smithand DianaFleise
are overlaidon this perceptualstructure.Thesebeliefsare shownby the heavier
arrows in Figure 7.778. (Keil suggesrs
rhar 7.118 may in facr more closely
representthe startingpoint for development
than 7.114.sincesome domainspecificcausal theoriesmay be "innate".) With increasingage, theoriesare
elaboratedas in 7.11C. Early in development,
Keil suggests,
causalbeliefsmay
interpret the perceptualassociations
benveenprimitivesbut with development,
theybecomemore systematic,
elaborate,
and differentiated.
A
B
Figure7.11.TakenfromKeil (1989).A: Networksof features
linkedby associations;
B,C:
The emergence
of theorvindicatedbv arrows.
Perceptual
Similarity
andConceptual
Structure
265
We concurwith the main pointsof thisoutlineof a theory. Yet we find
its portrayalof perceptual
structuredisquieting.
Again,staticperceptualstructure
providesat besta minorconstraint
on conceptual
structure.In Keil'sdescription,
perceptualstructureonlyweaklyreflects(in is associative
connections)
the causal
homeostasisthat existsbetweenpropertiesin theworld. The causalhomeostasis
CONTEXT
THE PERCEPTUAL
LANDSCAPE
C n-r
Figure 7.12. Perceptual landscapesthat vary with context and causal belief systems
(arrows) that attempt to explainthem.
266
LrndaB. Smithand DianaHeise
that Boyd writes about finds its psychological
home in Keil's view in people's
elaboratedand systematic
causalbeliefs.
We offer as an alternativeto Keil the outlineof a theoryof perceptual
context Ct, is shownthe perceivedsimilarityof objectsalong someset of and
conceptualstructure
depictedin Figure7.12. Whatthisfigureattemptsto depict
is the changingperceptuallandscape
with changingcontextand people'stheories
aboutthe perceivedsimilarities.Thedynamicperceptual
Iandscape
is represented
by the contextdependent
andvariously
distortedgrids. Conceptual
theoriesare
representedby the arrowsthat point out relationsbetweendistortionsin the
perceptuallandscape
andbetweendistortions
andcontext.So,nextto dimensions
in that context. The perceivedsimilaritiesin the spaceare not uniformbecause
of learnedcorrelations,
context-dimension
associations,
and language-dimension
associations.We assumeall theseassociations
are acquiredfrom beingin the
world. ContextC, showsthe distortionsof this samesimilaity spacein another
contextthat activatesanotherset of associations
that shift the dimensionweights
and perceivedsimilarityin otherdirections.Thus,in contrastto Keil, we do not
presentone perceptuallandscape
because
thereis no onelandscape.Thereis no
staticoriginalsimilarity. Insteadthe perceptuallandscape
emergesin particular
contextswith particularcoalitionsof features.Morever,becauseof the natureof
perceptualsimilarity and the context-specific
shifts in feature weights,these
dynamicperceptuallandscapes
reflectthe causalstructureof the physicaland
socialworld. Boyd'scausalhomeostasis
findsits psychological
homein thistheory
in the bumpsand holesand ridgesof perceivedsimilarity.
People'sexplicitcausalbeliefsabout categoriesand relationsbetween
objectsare graphicallyillustratedby the arrows.By our view,peoplebuildcausal
--- incorporating
connections
betweenpointsin the perceptuallandscape
cultural
--and scientificknowledge in order to interpretand makesenseof the complex
perceptualterrain. Somecausalbeliefsmaybe localandaboutinterdependencies
in a singlecontextually
determinedterrain. Other naivetheoriesmaybe grander
and try to capturethe dynamicsof the changingemphasisand de-emphasis
of
featuresacrosscontexts.It is unlikelythat thesenaivetheoriesadequately
reflect
or explainthe rich structureof the perceptual
landscape
or the causalhomeostasis
in the world.
The differencebetweenour representation
and Keil's, then,is that the
perceptualstructureis richerand far morecomplicated.Perceptionprovidesthe
nutrientson whichconceptual
structuregrows.Perceptionis the realityon which
the conceptualsystemoperates. Indeed,it is becausewe are perceptuallyin
Structure
Similaritvand Concentual
Perceotual
267
structure
contactwith the complexworld that we needan elaboratedconceptual
that simplifiesthat complexity.
V. STRUCTURE AND PROCESS
The chaptersof this book haveall grappledwith the questionof the
relationbetweenstructureandprocess.Although,wehavenot directlyconfronted
structureand processmay be the root of the
the issue,the distinctionbenrueen
perception-conception
interactions.
The idea of the
problem in understanding
structure-process
distinctionis that cognitionis composed
of nruoparts: structure
whichis the represented
whichoperateson the structures.
knowledge
and process
In this view of cognition,structuresare the nounsof thoughts,the things,and
processesare the verbs,the actions.The problemwith this metaphoris that
perceptualand conceptualstructuresare designatedas thingsas if that theywere
when thinking
staticand unchanging.This metaphoris particularlytroublesome
developmentis
aboutchange
Smith
aboutdevelopmentbecause
(see,Smith,1.990;
& Sera,1990).
How might cognitivedevelopmentgo forward if knowledgeis a set of
"things"-- if we acquirepiecesof knowledgelike new piecesof clothing? One
possibilityis that childrenjust keep addingnew piecesto their knowledgeset.
Another possibilityis that childrenabandonor ("trade up") their use of one
structurein favor of another. For example,they might abandonperceptual
structurein favor of conceptualstructure.We havearguedthat this particular
solutionto developmental
changes
in categorization
doesnot fit the data.
Thereis an alternateto partitioningcognitioninto structureand process.
The alternateview is that there is only process.By this view, perceptualand
cognitivestructuresare not thing-likerepresentations
that residein the child but
are rather the emergentpropertiesof complexprocesses.This alternativeview
iS, of course, the one that underliesthe theoreticalendeavorsknown as
connectionism(e.g., Rumelhart& McClelland,1986) and dynamicalsystems
theory(Thelen,L989;Smith& Sera,1990). It is the assumption
that guidesour
work and forms our belief in perceptualsimilarityas dynamic. Perceptual
similarity--- perceptualstructure---is dynamicbecause
perceptual
structureis not
a thing; it is the emergentresultof attentionalprocesses.
268
Linda B. Smith and Diana Heise
\T. CONCLUSIONS
We have arguedfor a new researchagenda-- one that does not pit
perceptionagainstconceptionbut insteadaskshow perceptualand conceptual
structureare related and how they organizetogetherto propel development
forward. We havespecificallyarguedfor greaterattentionto the dynamicsof
perceptualsimilariryand perceptualstructure. Perceptualsimilarityis a much
malignedforce in cognition.Philosophers
suchas NelsonGoodman(1972\have
dismissed
similarityas too slipperyandvariablea conceptto haveanyexplanatory
power. By our view, the powerof perceptualsimilaritystemsdirectlyfrom its
variablenature.
(systematically)
ACKNOWLEDGBMENTS
werepresented
Portions
of thischapter
underthetitle"In defense
of perceptual
similarity"
at theSociety
for Research
in ChildDevelopment,
Kansas
City,April i989. Theresearch
reportedin this chapterwassupported
by PHS grantS07RR7021L
throughIndiana
University. We thank, Patricia Bauer, Barbara Burns, SusanJones,Jean Mandler, Clay
Mash, Doug Medin, and Maria Sera,for commentson earlier versionsof this manuscript.
REFERENCES
Bertenthal, B. I., Proffitt, D. R., Kramer, S. J., & Spetner,N. B. (1987). Infants' encoding
of kinetic displaysvaryingin relativecoherence.DevelopmentalPsycholog.v,
23,
17r-r78.
Biederman,I. (1985).Humanimageunderstanding.
ComputerVision,@ph !gq,& Imase
Processinq,
32,29-73.
Billman, D. (1989). Systemsof correlationsin rule and categorylearning: Use of
structuredinput in learningsyntacticcategories. Lanquaeeand Cognitive
Processinq,
4, 127-I55.
Bingham,G. P., Schmidt,R. C. and Rosenblum,
L. D. Heftingfor a maximumdistance
perceptual
throw: A smart
mechanism.Journalof ExperimentalPsycholoqy:
Human Perceptionand Performance
, 15,507-5?3.
Bornstein,M. H. (1985). Habituationof attentionas a measureof visualinformation
processing
in humaninfants:Summary,Systematization,
and Syntheses.
In G.
(Eds.)Measurement
GottliebandN. Krasnegor
of auditionandvisionin the first
life: A methodoloqical
overview.Norwood,
N.J.; Ablex,253vearof postnatal
300.
principlesaffectlearningand transferin children.
Brown,A. L. (1990). Domain-specific
Science,
14,107-134.
Coqnitive
Perceptual
Similarityand Conceptual
Structure
269
transformations
in
of equivalence
Bruner,J. S. &-Olver, R. R. (1963). Development
children.Monoeraphof theSocietyfor Research
in ChildDevelopment,
No. 28.
of causalreasoning.
Bullock,M., Gelman,R., & Baillargeon,
R. (1982).Thedevelopment
psycholosy
In W. J. Friedman(Ed.),The developmental
of time. New York:
AcademicPress.
chansein childhood.Cambridge,
Carey,S. (1985). Conceptual
MA: MIT Press.
In P. H. Mussen(Ed.),Carmichael's
manual
Flavell,J. H. (1970).Conceptdevelopment.
of child psycholosy:Vol 1. New York: Wiley.
organize
attentionto andlearningaboutrelevantdata:
Gelman,R. (1990).Firstprinciples
Numberand the animate-inanimate
distinction
as examples.CoqnitiveScience,
t4,79-t06.
Gelman,R., Spelke,E. S.,& Meck,E. (1983).What preschoolers
know aboutanimate
and inanimateobjects.In D. RogersandJ. A. Sloboda(Eds.),The acquisition
of svmbolicskills. l-ondon: Plenum.
Gelman,S. A. (1988). The development
of inductionwithin naturalkind and artifact
categories.C-oenitive
Ps),choloqv,
20, 65-95.
and inductionin youngchildren.
Gelman,S. & Markman,E. M. (1986). Categories
Cosnition,23, 183-209.
Gelman,S.A. & Markman,E. M. (1987).Youngchildren's
inductions
from naturalkinds:
and appearances.
The role of categories
ChildDevelopment,58,1532-154I.
Gentner,D. (1989). The mechanisms
of analogical
learning. In S. Vosniadouand A.
Ortony (Eds.),Similarityand analoeicalreasoning(pp 1,99-2al).New York:
CambridgeUniversifyPress.
Gentner,D. & Rattermann,M. J. (1990).l-anguage
and the careerof similarity.In S.
Gelman(Ed.),Lansuaseandconceptual
development.
Goodman,N. (1972). Problemsand projects.Indianapolis:Bobbs-Merrill.
Goldstone,
R. L., Medin,D. L., & Gentner,D. (in press).Relationalsimilarityand the
of featuresin similarityjudgments.CoenitivePsycholog.v.
non-independence
Greer,A. E. & Sera,M. D. (1990).Artifactsvs.naturalkinds:An empiricalinvestigation.
(Unpublishedmanuscript).
Gluck, M. A. & Bower,G. H. (1988). From conditioningto categorylearning: An
adaptivenetworkmodel. Journalof Experimental
Psycholory:General,117,
227-47.
Heise,D. & Smith,L.B. (1990).Learningco-relations.
(Unpublished
manuscript).
Jeyifous,
S. (1986).Antimodemo:Semantic
conceptual
development
amongthe Yoruba.
Doctoral dissertation,
CornellUniversity.
Jones,S. S.,Smith,L.B., & I-andau,B. (i990). Objectpropertiesandknowledge
in early
lexicallearning. (Under review)
Johnson,
M. (1987).The bodyin themind: The bodilybasisof meaning,
imagination,
and
reasonins.Chicago: Universityof ChicagoPress.
Keil, F. C. (1989). Concepts,
kinds,andcoqnitivedevelopment.Cambridge,
MA: MIT
Press.
Kemler Nelson,D. G. (1990). When experimentalfindingsconflictwith everyday
observations:
Reflections
on children's
categorylearning.Child Development,
61,606-610.
270
Linda B. Smithand DianaHeise
I-akoff,G. (1972). Hedges:A studyin meaningcriteriaand the logicof fuzzyconceprs.
In Papersfrom the EighthResionalMeetingof the ChicagoLinsuisticSocietv.
Also in Journalof Philosophical
Loqic(1973)2, 458-508.
I-akoff,G. (1987). Women,fire,anddanserous
thinqs:Whatcateqories
revealaboutthe
mind. Chicago:Universityof ChicagoPress.
I-andau,8., smith,L. B., & Jones,s. s. (1988).The Importance
of shapein EarlyLexical
LearningCoqnitiveDevelopment,
3, 299-321.
Leslie,A. (1984). Infant perceptionof a manualpick-upevent. British Journalof
Developmental
Psycholoqv,
Z, 19-32.
Leslie,A. (1988). The necessity
of illusion: Perceptionand thoughtin infancy. In L.
Weiskrantz(Ed.),Thoughtwithoutlanguage.Oxford: Clarendon.
Mackintosh,N. J. (1965). Selectiveattention in animal discriminationlearning.
Ps),choloeical
Bulletin,64,I24-I50.
Macnamara,J.(1982). Namesfor thinss:A studyof humanlearnins.Cambridge,
MA:
MIT Press.
Mandelbrot,B. B. (1983). The fractalseometryof nature. SanFrancisco:Freeman.
Mandler,J. M. (1988). How to build a baby: On the development
of an accessible
representational
system.CognitiveDevelopment,
3, 113-136.
Mandler,J. M. (1990).How to builda baby: Part2. (Unpublished
manuscript)
Mandler,J. M. & Bauer,P. J. (1988). The cradleof categorization:
Is the basiclevel
basic? CognitiveDevelopment,
3, 237-264.
Mandler,J. M., Bauer,P. J., & McDonough,
L. (in press).Separating
thesheepfrom the
goats: Differentiating
globalcategories.CoenitivePsycholo{v.
Markman,E. M. (1989).C-atesorization
and naminqin children:Problemsof induction.
C-ambridge,
MA: MIT Press.
Medin, D. & Ortony,A. (1989). Psychological
essentialism.In S. Vosniadouand A.
Ortony (Eds.),Similarit),and analosicalreasoning.New York: C-ambridge
UniversityPress.
Medin, D. L. & Schaffer,M. M. (1978). Contexttheory of classification
learning.
Psychological
Review,85,207-238.
Medin,D.L., Ahn, W., Bettger,J., Florian,J.,Goldstone,
R., l-assaline,
M., Markman,A.,
Rubinstein,
J., & Wisniewski,
E. (1990).Safetakeoffs- softlandings.Cognitive
Science,
14,169-178.
Medin, D.L., Altom, M. W., Edelson,S. M. & Freko,D. (1982). Correlatedsymptoms
and simulatedmedicalclassification.Journal of ExperimentalPsychology:
Learning,Memoryand Cognition,
8, 37-50.
Medin, D. L., Wattenmaker,W. D., & Hampson,S. E. (1987). FamilyResemblance,
conceptualcohesiveness
and categoryconstruction.Coqnitive!, 19,242-299.
Mervis,C. B. (1987).Child-basic
objectcategories
andlexicaldevelopment.
In U. Neisser
(Ed.),Concepts
andconceptualdevelopment:
Ecological
andintellectual
factors
in cateqorization.
Cambridge,
England:Cambridge
UniversityPress.
Murphy, G. L. & Medin, D. L. (1985). Role of theoriesin conceptualcoherence.
Psychological
Review,92,?39-316.
Nosofsky,R. M. (1984). Choice,similarity,and the contextof classification.
Journalof
ExperimentalPsvcholos.y:
Learning,Memory,& Coenition,70,104-II4.
Structure
and Conceptual
Similarity
Perceptual
271
Nosofsky,R. M. (1986). Atrention,similarity,and the identification-categorization
Psycholory:General,115,39-57.
relationship.Journalof Experimental
piaget,J. (lg}g). The child'sconception
of the world. New York: HarcourtBrace.
and
of meaning.In H. Putnam(Ed.),Mind,lansuage,
Putnam,H. (1975).The meaning
Press'
University
Cambridge
reality,Vol. 2. London:
MA: MIT Press'
Quine,W. V. O. (1960). Word and object. Cambridge,
and
Quine,W. V. O. (1977). Naturalkinds. In S. P. Schwartz(Ed.),Namins,necessity,
UniversityPress.
naturalkinds. Ithaca,NY: C.ornell
4,328-350'
CoenitivePsycholoKv,
Rosch,E. (1973).NaturalCategories,
Studiesin the internalstructure
Rosch,E. & Mervis,C. B. (1975).Familyresemblances:
of categories.CognitivePsycholos.v,
7, 573-605.
P. (1976).BasiC
M.D., & BOyes-Braem,
Rosch,8., Mervis,C.8., Gray,W.D.,Johnson,
objectsin naturalcategories.CosnitivePsycholos.v,
9, 382-439to
the cognitiveunconscrous.
access
and
of
intelligence
evolution
(1976).
The
Rozin,P.
6,245-?30'
Psycholory,
Physiolosical
and
in Psychobioloq.v
Progress
ParallelDistributed
(1986).
Group
andthePDPResearch
D. E,.,McClelland,
Rumelhardt,
MA: MIT Press.
Cambridge,
Volume1: Foundations,
Processing.
generalization
for psychologicalscience.
of
law
a
universal
Toward
R. W. (1987).
Shepard,
237,1317-1323.
W@@@@@@@
.,
T. W. (1964).Attentionandthe metricstructureof the stimulusspace.Journal
Shepard,
7, 54-87.
of MathematicalPsychologv,
MA: Harvard
andconcepts.Cambridge,
Categories
(1981).
D.
L.
Smith,E. E. & Medin,
UniversityPress.
D. G. (1984). Overallsimilarityin adults'classification:
Smith,J. D. & Kemler-Nelson,
General,1I3,I37PsycholoKy:
The child in all of us. Journalof Experimental
159.
Smith,L. B. (1981). The importanceof the overaltsimilarityof objectsfor adults'and
HumanPerception
Psychology:
Journalof Experimental
classification.
children's
L, 81I'824.
and Performance,
in children and adults.
Smith, L. B. (1989). A model of perceptualclassification
-144.
Review,96, 125
Psvcholosical
analysisof the polar structureof
Smith,L. B. & Sera,M. (1990). A developmental
(Under
review)
dimensions.
and perceptual
Smith,L. B., Jones,S. S., & fandau, B. (1990). Nouns,adjectives,
(Under review).
lexicalstrateSies.
propertiesin veryyoungchildren's
Spelke,E. S. (1982). Perceptualknowledgeof objectsin infancy. In J. Mehler,MHillsdale,
on mentalrepresentation.
Garrett,andE. Walker(Eds),Perspectives
NJ: Erlbaum.
Sugarman,S. (1983). Children'searly thousht. Cambridge,England,Cambridge
UniversityPress.
148by animalls.Edeavour,23,
N. S. (1964).The learningor discriminations
Sutherland,
r52.
approaches
C-ansystems
processes:
in developmental
Thelen,E. (1989).Self-organization
The
Development:
and
Systems
(Eds.),
work; In M. Gunnar& E. Thelen
27z
Linda B. Smithand DianaHeise
vol.22,pp. 11-1i7. HillsdaleNJ:
MinnesotaSymposiaon Child Psycholoqy,
Erlbaum.
Review,84,327-352.
Tversky,A. (1977). Featuresof similarity.Psychological
of namedand pictured
Tversky,B. (1985). Developmentof taxonomicorganization
1
19.
11
1
1-1
Psycholosy
Developmental
categories
, L,
In E. HanfmannandG. Vakar,trans.
Thoushtandlanquaqe.
Vygotsky,L. S.(1934/1962).
Cambridge,MA: MIT Press.
hierarchies:
of conceptual
Waxman,S. R. (1990).Linguisticbiasesandthe establishment
5, 123-150.
children.QpsnitiveDevelopment,
Evidencefrom preschool
In
of the nonobvious.
understanding
Wellman,H. M. & Gelman,S.A. (1988).Children's
Vol. 4. Hillsdale,
of intellisence,
in thepsycholoKv
(Ed.),Advances
R. Sternberg
NJ: Erlbaum.
mentaldevelopment,2nd Ed. New York:
Werner, H. (1948). Comparativepsycholoqv
Press.
InternationalUniversities
perception
to inference: A dimensionof cognitive
From
l.
F.
(1962).
Wohlwill,
for Researchin Child DeveloPment,
the
Society
of
development.Monoqraphs
27, 87-112.
amongfeaturecategories.
of correlations
Si!!
Younger,B.A. (1990).Infants'detection
6t,
61
4-620.
Development,
amongattributes.
of correlations
Younger,B. A. & Cohen,L. B. (1983).Infantperception
54,858-867.
Child Development,