Page 1
FitnessBeatsTruthintheEvolutionofPerception1
2
ChetanPrakasha*,KyleD.Stephensb,DonaldD.Hoffmanb,3
ManishSinghc,ChrisFieldsd4
aDepartmentofMathematics,CaliforniaStateUniversity,SanBernardino,CA92407;5
bDepartmentofCognitiveSciences,UniversityofCalifornia,Irvine6
cDepartmentofPsychologyandCenterforCognitiveScience,RutgersUniversity,NewBrunswick,7
d243WestSpainStreet,Sonoma,CA954768
9
10
Abstract11
Doesnaturalselectionfavorveridicalperceptions—thosewhichaccurately,thoughperhaps12
notexhaustively,depictobjectivereality?Prominentvisionscientistsandevolutionary13
theoristsclaimthatitdoes.Hereweformalizethisclaimusingthetoolsofevolutionary14
gametheoryandBayesiandecisiontheory.Wethenpresentandprovea"Fitness-Beats-15
Truth(FBT)Theorem"whichshowsthattheclaimisfalse.Wefindthatincreasingthe16
complexityofobjectivereality,orperceptualsystems,orthetemporaldynamicsoffitness17
functions,increasestheselectionpressuresagainstveridicalperceptions.Weillustratethe18
FBTTheoremwithaspecificexampleinwhichveridicalperceptionminimizesexpected19
fitnesspayoffs.WeconcludethattheFBTTheoremsupportsthe"interfacetheoryof20
perception,"whichproposesthatoursenseshaveevolvedtohideobjectiverealityand21
guideadaptivebehavior.Italsosupportstheassertionofsomeproponentsofembodied22
*[email protected] ;+1(909)537-5390
Page 2
cognitionthat“representingtheanimal-independentworldisnotwhataction-oriented23
representationsaresupposedtodo;theyaresupposedtoguideaction”(Chemero,2009).24
25
Keywords:ReplicatorDynamics;InterfaceTheoryofPerception;EvolutionaryGame26
Theory;UniversalDarwinism;Sensation27
28
29
1.Introduction30
Itisstandardintheperceptualandcognitivesciencestoassumethatmoreaccurate31
perceptionsarefitterperceptionsand,therefore,thatnaturalselectiondrivesperceptionto32
beincreasinglyveridical,i.e.toreflecttheobjectiveworldinanincreasinglyaccurate33
manner.Thisassumptionformsthejustificationfortheprevalentviewthathuman34
perceptionis,forthemostpart,veridical.Forexample,inhisclassicbookVision,Marr35
(1982)arguedthat:36
“We...verydefinitelydocomputeexplicitpropertiesoftherealvisiblesurfacesout37
there,andoneinterestingaspectoftheevolutionofvisualsystemsisthegradual38
movementtowardthedifficulttaskofrepresentingprogressivelymoreobjective39
aspectsofthevisualworld”.(p.340)40
Similarly,inhisbookVisionScience,Palmer(1999)statesthat:41
“Evolutionarilyspeaking,visualperceptionisusefulonlyifitisreasonably42
accurate...Indeed,visionisusefulpreciselybecauseitissoaccurate.Byandlarge,43
whatyouseeiswhatyouget.Whenthisistrue,wehavewhatiscalledveridical44
Page 3
perception...perceptionthatisconsistentwiththeactualstateofaffairsinthe45
environment.Thisisalmostalwaysthecasewithvision.”46
Indiscussingperceptionwithinanevolutionarycontext,GeislerandDiehl(2003)similarly47
assumethat:48
“Ingeneral,(perceptual)estimatesthatarenearerthetruthhavegreaterutility49
thanthosethatarewideoffthemark.”50
Intheirmorerecentbookonhumanandmachinevision,Pizloetal.(2014)gosofarasto51
saythat:52
“…veridicalityisanessentialcharacteristicofperceptionandcognition.Itis53
absolutelyessential.Perceptionandcognitionwithoutveridicalitywouldbelike54
physicswithouttheconservationlaws.”(p.227,emphasistheirs.)55
Ifhumanperceptionisinfactveridical,itfollowsthattheobjectiveworldsharesthe56
attributesofourperceptualexperience.Ourperceivedworldisthree-dimensional,andis57
inhabitedbyobjectsofvariousshapes,colors,andmotions.Perceptualandcognitive58
scientiststhustypicallyassumethattheobjectiveworldissoinhabited.Inotherwords,59
theyassumethatthevocabularyofourperceptualrepresentationsisthecorrectvocabulary60
fordescribingtheobjectiveworldand,moreover,thatthespecificattributesweperceive61
typicallyreflecttheactualattributesoftheobjectiveworld.Theseassumptionsare62
embodiedwithinthestandardBayesianframeworkforvisualperception,whichwe63
considerinthenextsection.64
Someproponentsofembodiedcognitionrejecttheclaimthatperceptionisnormally65
veridical.Forinstance,Chemero(2009)arguesthat“…perceptualsystemsevolvedtoguide66
behavior.Neitherhumansnorbeetleshaveaction-orientedrepresentationsthatrepresent67
Page 4
theanimal-independentworldexactlycorrectly.Indeed,representingtheanimal-68
independentworldisnotwhataction-orientedrepresentationsaresupposedtodo;they69
aresupposedtoguideaction.Sothesetofhumanaffordances,thatis,action-oriented70
representeds,isjustastightlygearedtohumanneedsandsensorimotorcapacitiesasthose71
ofothertypesofanimal.Thisleavesuswithamultiplicityofconflictingsensorimotor72
systems,eachofwhichisappropriateforguidingtheadaptivebehaviorofanimalswhose73
systemstheyare.”TheFBTTheorem,whichwepresentbelow,supportsChemero’sclaim.It74
issupported,inturn,byspecificexamplesofnon-veridicalperceptions,suchasthose75
discussedbyLoomis(2004)andKoenderinket.Al.(2010).76
77
2.ThestandardBayesianframeworkforvisualperception78
Thestandardapproachtovisualperceptiontreatsitasaproblemofinverseoptics:The79
“objectiveworld”—takentobe3Dscenesconsistingofobjects,surfaces,andlightsources—80
projects2Dimagesontotheretinas.Givenaretinalimage,thevisualsystem’sgoalistoinfer81
the3Dscenethatismostlikelytohaveprojectedit(e.g.Adelson&Pentland,1996;Feldman,82
2013;Knill&Richards,1996;Mamassian,Landy,&Maloney,2002;Shepard,1994;Yuille&83
Bulthoff,1996).Sincea2Dimagedoesnotuniquelyspecifya3Dscene,theonlywaytoinfer84
a3Dsceneistobringadditionalassumptionsor“biases”tobearontheproblem—basedon85
priorexperience(whetherphylogeneticorontogenetic).Forexample,ininferring3Dshape86
fromimageshading,thevisualsystemappearstomaketheassumptionthatthelightsource87
ismorelikelytobeoverhead(e.g.Kleffner&Ramachandran,1992).Similarly,ininferring88
3Dshapefrom2Dcontours,itappearstousetheassumptionthat3Dobjectsaremaximally89
compactandsymmetric(e.g.Lietal.,2013).90
Page 5
Formally,givenanimage𝑥",thevisualsystemaimstofindthe“best”(generallytakento91
mean“mostprobable”)sceneinterpretationintheworld.Inprobabilisticterms,itmust92
comparetheposteriorprobabilityℙ 𝑤 𝑥" ofvarioussceneinterpretations𝑤,giventhe93
image𝑥".ByBayes’Rule,theposteriorprobabilityisgivenby:94
ℙ 𝑤 𝑥" = ℙ 𝑥" 𝑤 ∙ℙ(()ℙ(*+)
95
Sincethedenominatortermℙ(𝑥")doesnotdependon𝑤,itplaysnoessentialrolein96
comparingtherelativeposteriorprobabilitiesofdifferentscenesinterpretationsw.The97
posteriorprobabilityisthusproportionaltotheproductoftwoterms:Thefirstisthe98
likelihoodℙ 𝑥" 𝑤 ofanycandidatesceneinterpretationw;thisistheprobabilitythatthe99
candidatescenewcouldhaveprojected(orgenerated)thegivenimage𝑥".Becauseany2D100
imageistypicallyconsistentwithmanydifferent3Dscenes,thelikelihoodwilloftenbe101
equallyhighforanumberofcandidatescenes.Thesecondtermisthepriorprobability102
ℙ(𝑤)ofasceneinterpretation;thisistheprobabilitythatthesystemimplicitlyassignsto103
differentcandidatescenes,evenpriortoobservinganyimage.Forexample,thevisual104
systemmayimplicitlyassignhigherpriorprobabilitiestosceneswherethelightsourceis105
overhead,ortoscenesthatcontaincompactobjectswithcertainsymmetries.Thus,when106
multiplesceneshaveequallyhighlikelihoods(i.e.areequallyconsistentwiththeimage),107
thepriorcanserveasadisambiguatingfactor.108
ApplicationofBayes’Ruleyieldsaprobabilitydistributiononthespaceofcandidate109
scenes—theposteriordistribution.Astandardwaytopickasingle“best”interpretation110
fromthisdistributionistochoosetheworldscenethathasthemaximalposterior111
probability—onethat,statisticallyspeaking,hasthehighestprobabilityofbeingthe112
“correct”one,giventheimage𝑥".Thisisthemaximum-a-posterioriorMAPestimate.More113
generally,thestrategyoneadoptsforpickingthe“best”answerfromtheposterior114
Page 6
distributiondependsonthechoiceofaloss(orgain)function,whichdescribesthe115
consequencesofmaking“errors,”i.e.pickinganinterpretationthatdeviatesfromthe“true”116
(butunknown)worldstatebyvaryingextents.TheMAPstrategyfollowsunderaDirac-117
deltalossfunction—nolossforthe“correct”answer(or“nearlycorrect”withinsome118
tolerance),andequallossforeverythingelse.Otherlossfunctions(suchasthesquared-119
errorloss)yieldotherchoicestrategies(suchasthemeanoftheposteriordistribution;see120
e.g.Mamassianetal.,2002).ButwefocusontheMAPestimateherebecause,inawell-121
definedsense,ityieldsthehighestprobabilityofpickingthe“true”sceneinterpretation122
withinthisframework.123
ThisstandardBayesianapproachembodiesthe“veridicality”or“truth”approachtovisual124
perception.Bythiswedonotmean,ofcourse,thattheBayesianobserveralwaysgetsthe125
“correct”interpretation.Giventheinductivenatureoftheproblem,thatwouldbea126
mathematicalimpossibility.Itisneverthelesstruethat:127
(i) ThespaceofhypothesesorinterpretationsfromwhichtheBayesianobserver128
choosesisassumedtocorrespondtotheobjectiveworld.Thatis,thevocabulary129
ofperceptualexperiencesisassumedtotherightvocabularyfordescribing130
objectivereality.131
(ii) Giventhissetup,theMAPstrategymaximizes(statisticallyspeaking)the132
probabilityofpickingthe“true”worldstate.133
134
3.EvolutionandFitness135
TheBayesianframework,summarizedabove,focusesonestimatingtheworldstatethathas136
thehighestprobabilityofbeingthe“true”one,givensomesensoryinputs.Thisestimation137
Page 7
involvesnonotionofevolutionaryfitness.2Inordertobringevolutionandfitnessintothe138
picture,wethinkoforganismsasgatheringfitnesspointsastheyinteractwiththeir139
environment.ThuseachelementwoftheworldWhasassociatedwithitafitnessvalue.In140
general,however,thefitnessvaluedependsnotonlyontheworld,butalsoontheorganism141
oinquestion(e.g.,lionvs.rabbit),itsstates(e.g.,hungryvs.satiated),andtheactionclassa142
inquestion(e.g.,feedingvs.mating).Givensuchafitnesslandscape,naturalselectionfavors143
perceptionsandchoicesthatyieldmorefitnesspoints.144
Wemaythusdefineaglobalfitnessfunctionasa(non-negative)real-valuedfunctionf(w,o,145
s,a)ofthesefourvariables.However,oncewefixanorganism,itsstateandagivenaction146
class,i.e.,oncewefixo,sanda,aspecificfitnessfunctionissimplya(non-negative)real-147
valuedfunction𝑓:𝑊 → [0,∞)definedontheworldW.148
Inordertocomparethefitnessofdifferentperceptualand/orchoicestrategies,onepits149
themagainstoneanotherinanevolutionaryresourcegame(forsimulationsexemplifying150
theresultsofthispaper,see,e.g.,Mark,Marion,&Hoffman,2010;Marion,2013;andMark,151
2013).Inatypicalgame,twoorganismsemployingdifferentstrategiescompetefor152
availableterritories,eachwithacertainnumberofresources.Thefirstplayerobservesthe153
availableterritories,chooseswhatitestimatestobeitsoptimalone,andreceivesthefitness154
payoffforthatterritory.Thesecondplayerthenchoosesitsoptimalterritoryfromthe155
remainingavailableones.Thetwoorganismsthustaketurninpickingterritories,seeking156
tomaximizetheirfitnesspayoffs.157
Inthiscase,thequantityofresourcesinanygiventerritoryistherelevantworldattribute.158
Thatis,Wishereinterpretedasdepictingdifferentquantitiesofsomerelevantresource.1592Asnotedabove,Bayesianapproachesofteninvolvealoss(orgain)function.However,thisisquitedistinctfromafitnessfunction,asdefinedbelow.Specifically,lossfunctionsarefunctionsoftwovariablesl(x,x*),wherex*isthe“true”worldstate,andxisahypotheticalestimatearrivedatbytheobserver.Afitnessfunctionis,however,notafunctionoftheobserver’sestimatex.
Page 8
WecanthenconsideraperceptualmapP :W → X ,whereXisthesetofpossiblesensory160
states,togetherwithanorderingonit:Ppicksoutthe“best”elementofXinasenserelevant161
totheperceptualstrategy.Onemay,forinstance,imagineasimpleorganismwhose162
perceptualsystemhasonlyasmallnumberofdistinctsensorystates.Itsperceptualmap163
wouldthenbesomewayofmappingvariousquantitiesoftheresourcetothesmallsetof164
availablesensorystates.Asanexample,Figure1showstwopossibleperceptualmappings,165
i.e.twowaysofmappingthequantityofresources(here,rangingfrom0through100)to166
fouravailablesensorycategories(heredepictedherebythefourcolorsR,Y,G,B).167
168
Figure1.AsimpleexampleshowingtwodifferentperceptualmappingsP :W → X from169
worldstates,W=[1,100]tosensorystatesX={R,Y,G,B}.170
Inaddition,thereisafitnessfunctiononW,𝑓:𝑊 → [0,∞),whichassignsanon-negative171
fitnessvaluetoeachresourcequantity.Onecanimaginefitnessfunctionsthatare172
monotonic(e.g.fitnessmayincreaselinearlyorlogarithmicallywiththenumberof173
resources),orhighlynon-monotonic(e.g.fitnessmaypeakforacertainnumberof174
Page 9
resources,anddecreaseineitherdirection).Non-monotonicfitnessfunctions(suchasthe175
oneshowninFigure2)areinfactquitecommon:toolittlewaterandonediesofthirst,too176
muchwaterandonedrowns.Similarargumentsapplytothelevelofsalt,ortothe177
proportionofoxygenandindeedanynumberofotherresources.Indeed,giventhe178
ubiquitousneedfororganismstomainhomoeostasis,oneexpectsnon-monotonicfitness179
functionstobeprevalent.(Moreover,fromapurelymathematicalpointofview,thesetof180
monotonicfitnessfunctionsisanextremelysmallsubsetofthesetofallfunctionsona181
givendomain.Thatistosay,thereare“manymore”non-monotonicfunctionsthan182
monotonicones;hencearandomsamplingoffitnessfunctionsismuchmorelikelytoyielda183
non-monotonicone.)184
185
Figure2.Anexampleofanon-monotonicfitnessfunction𝑓:𝑊 → [0,∞).Fitnessismaximal186
foranintermediatevalueoftheresourcequantityanddecreasesineitherdirection.Given187
theubiquitousneedfororganismstomainhomoeostasis,oneexpectsthatsuchfitness188
functionsarequitecommon.189
190
4.Comparingperceptualstrategies:“Truth”vs.“Fitness-only”191
Page 10
Inthecontextoftheseevolutionarygames,inwhichperceptualstrategiescompetefor192
resourceacquisition,wetakeasfixedandknowntotheorganism:thespecificfitness193
function,itsprior(inaparticularstateandforaparticularactionclass)anditsperceptual194
map(seeFigure3).Onanygiventrial,theorganismobservesanumberofavailable195
territoriesthroughitssensorystates,sayx1,x2,…,xn.Itsgoalistopickoneofthese196
territories,seekingtomaximizeitsfitnesspayoff.Onecannowconsidertwopossible197
resourcestrategies:198
The“Truth”strategy:Foreachofthensensorystates,theorganismestimatestheworld199
stateorterritory-theBayesianMAPestimate-thathasthehighestprobabilityofbeingthe200
“true”one,giventhatsensorystate.Itthencomparesthefitnessvaluesforthoseestimated201
worldstates.Finally,itmakesitschoiceofterritorybasedonthesensorystatexithatyields202
thehighestfitness.ItschoiceisthusmediatedthroughMAPestimateoftheworldstate.203
The“Fitness-only”strategy:Inthisstrategy,theorganismmakesnoattempttoestimate204
the“true”worldstatecorrespondingtoeachsensorystate.Ratheritdirectlycomputesthe205
expectedfitnesspayoffthatwouldresultfromeachpossiblechoiceofxi.Foragivensensory206
statexi,thereisaposteriorprobabilitydistribution(given,aswiththeTruthstrategy,by207
Bayes’formula)onthepossibleworldstates,aswellasafitnessvaluecorrespondingto208
eachworldstate.Theorganismweightsthesefitnessvaluesbytheposteriorprobability209
distribution,inordertocomputetheexpectedfitnessthatwouldresultfromthechoicexi.210
Anditpickstheonewiththehighestexpectedfitness.211
Page 11
212
Figure3.Theframeworkwithinwhichwedefinethetworesourcestrategies.Weassumea213
fixedperceptualmapP :W → X aswellasafixedfitnessfunction𝑓:𝑊 → [0,∞).Givena214
choiceofavailableterritoriessensedthroughthesensorystates,sayx1,x2,…,xn,the215
organism’sgoalistopickoneofthese,seekingtomaximizeitsfitnesspayoff.216
217
5.TheoremsfromEvolutionaryGameTheory218
Inanevolutionarygamebetween the twostrategies, sayA andB,thepayoffmatrix is as219
follows:220
𝑎𝑔𝑎𝑖𝑛𝑠𝑡𝐴 𝑎𝑔𝑎𝑖𝑛𝑠𝑡𝐵𝐴𝑝𝑙𝑎𝑦𝑠 𝑎 𝑏𝐵𝑝𝑙𝑎𝑦𝑠 𝑐 𝑑
221
Herea,b,c, andd denote thevariouspayoffs to the rowplayerwhenplayingagainst the222
column player. E.g., b is the payoff to Awhen playing B.We will refer to three main223
theoremsfromevolutionarygametheoryrelevanttoouranalysis,asfollows.224
We first consider games with infinite populations. These are investigated bymeans of a225
deterministic differential equation, called the replicator equation, where time is the226
independentvariable and the relativepopulation sizes𝑥C, 𝑥Dare thedependentvariables,227
Page 12
with𝑥C + 𝑥D = 1(TaylorandJonker,1978,HofbauerandSigmund,1990,Nowak2006).In228
thiscontext,therearefourgenericbehaviorsinthelongrun:229
Theorem1. (Nowak 2006) Inagamewithan infinitepopulationof twotypes,AandB,of230
players,either231
(i) Adominates B (in thesense thatanon-zeroproportionofAplayerswilleventually232
take over the whole population), if𝑎 ≥ 𝑐 and b≥ 𝑑 (with at least one of the233
inequalitiesbeingstrict);234
(ii) BdominatesA,if𝑎 ≤ 𝑐andb≤ 𝑑(withatleastoneoftheinequalitiesbeingstrict);235
(iii) AandBcoexist,if𝑎 ≤ 𝑐andb≥ 𝑑(withatleastoneoftheinequalitiesbeingstrict),236
atastableequilibriumgivenby𝑥C∗ =JKL
JMNKOKL(and𝑥D∗ = 1 − 𝑥C∗);237
(iv) Thesystemisbistable,if𝑎 ≥ 𝑐andb≤ 𝑑(withatleastoneoftheinequalitiesbeing238
strict)andwilltendtowardseitherallAorallBfromanunstableequilibriumatthe239
samevalueof𝑥C∗ asabove.240
A fifth,non-genericpossibility is that𝑎 = 𝑐andb= 𝑑, inwhichcasewehave thatAandB241
areneutralvariantsofoneanother:anymixtureofthemisstable.242
Games with a finite population size N can be analyzed via a stochastic, as against243
deterministic, approach. The dynamics are described by a birth-death process, called the244
Moranprocess(Moran1958).Theresultsaremorenuancedthanintheinfinitepopulation245
sized case: there are now eight possible equilibrium behaviors, and they are population246
dependent,notjustpayoffdependent.247
Let𝜌CD denote the fixation probability of a single A individual in a population of N-1 B248
individuals replacing (i.e., taking over completely) that population. Similarly, let Let𝜌DC249
denotethefixationprobabilityofasingleBindividualinapopulationofN-1ofAindividuals250
Page 13
replacing(i.e.,takingovercompletely)thatpopulation.Intheabsenceofanyselection,we251
havethesituationofneutraldrift,wheretheprobabilityofeitherof theseevents is justRS.252
WesaythatselectionfavorsAreplacingBif𝜌CD >RSandthatselectionfavorsBreplacingAif253
𝜌DC >RS.254
By analyzing the probabilities of a single individual of each type interacting with an255
individualofeithertype,orofdyingoff,wecanusethepayoffmatrixabovetocomputethe256
fitness𝐹V ,when there are i entities of typeA, and the fitness𝐺V of (theN-i individuals) of257
type B. If we setℎV = 𝐹V − 𝐺V (𝑖 = 1, . . . , 𝑁), we can see thatℎR > 0implies that selection258
favorsAinvadingB,whileℎSKR > 0impliesthatselectionfavorsBinvadingA.Therearenow259
sixteen possibilities, depending upon whether selection favors A replacing B or not; B260
replacing Aor not; whether selection favors A invading Bor not; and whether selection261
favorsBinvadingAornot.Ofthese,eightareruledoutbyatheoremofTaylor,Fudenberg,262
SasakiandNowak(2004).Afulldescriptionisprovidedinthatpaper,alongwithanumber263
of theorems detailing the possibilities in terms of the payoff values and population size.264
TheirTheorem6, interpretedbelowasourTheorem2, ismostrelevanttoouranalysisof265
evolutionary resource games: it gives conditions underwhich selection is independent of266
population size and is reproduced below. Interestingly, for finite populations the267
relationshipbetweenpayoffsbandcbecomesrelevant:268
Theorem2. Inagamewithafinitepopulationoftwotypesofplayers,AandB,if𝑏 > 𝑐, 𝑎 >269
𝑐and𝑏 > 𝑑,wehaveforallN,ℎV > 0∀𝑖and𝜌CD >RS> 𝜌DC:selectionfavorsA.270
Finally,wealsoconsider,withinlargefinitepopulations,thelimitofweakselection.Inorder271
tomodelthestrengthofselection,anewparameterwisintroduced.Thisparameter,lying272
between0and1,isameasureofthestrengthofselection:wewritethefitnessofAnowas273
Page 14
𝑓V = 1 − 𝑤 + 𝑤𝐹V and the fitness of B now as𝑔V = 1 − 𝑤 + 𝑤𝐺V . When𝑤 = 0, there is no274
selection:thefitnessesareequalandwehaveneutraldrift.When𝑤 = 1,wehaveselection275
atfullstrength.AnanalysisofthedynamicsoftheMoranprocessunderweakselection(i.e.,276
inthelimitas𝑤 → 0),reveals(followingNowak2006,equation7.11)that:277
Theorem3.Inagamewithafinitepopulationoftwotypesofplayers,AandB,andwithweak278
selection, 𝑎 − 𝑐 + 2 𝑏 − 𝑑 > ] OKN K(JKL)S
implies that𝜌CD >RS. Thus, if𝑎 > 𝑐 and𝑏 > 𝑑 ,279
forlargeenoughN,selectionfavorsA.3280
281
6.EvolutionaryResourceGames282
Foroursituationoftworesourcestrategies,wemaydefinethepayoffmatrixasfollows:283
a:toFitness-Onlywhenplayingagainst
Fitness-Only
b:toFitness-Onlywhenplayingagainst
Truth
c:toTruthwhenplayingagainstFitness-Only d:toTruthwhenplayingagainstTruth
284
In a game with a very large (effectively infinite) number of players, the Fitness-Only285
resource strategy dominates the Truth strategy (in the sense that Fitness-Only will286
eventually driveTruthto extinction) if the payoffs toFitness-Only as first player always287
exceedthoseofTruthasfirstplayer,regardlessofwhothesecondplayeris,i.e.if and288
andatleastoneoftheseisastrictinequality.Ifneitheroftheseinequalitiesisstrict,289
thenattheleastFitness-OnlywillneverbedominatedbyTruth.290
3ThevalueofNatwhichthishappensdependsuponthepayoffmatrix,butcanbearbitrarilylargeoverthesetofallpayoffmatricessatisfying𝑎 > 𝑐and𝑏 > 𝑑.
!a ≥c
!b ≥d
Page 15
OurmainclaiminthispaperisthattheTruthstrategy—attemptingtoinfertothe“true”291
stateoftheworldthatismostlikelycorrespondtoagivensensorystate—confersno292
evolutionaryadvantagetoanorganism.Inthenextsection,westateandproveatheorem—293
the“FitnessBeatsTruth"theorem—whichstatesthatFitness-Onlywillneverbedominated294
byTruth.Indeed,theTruthstrategywillgenerallyresultinalowerexpected-fitnesspayoff295
thantheFitness-Onlystrategy,andisthuslikelytogoextinctinanyevolutionary296
competitionagainsttheFitness-Onlystrategy.(ThestatementoftheFBTtheorem297
articulatestheprecisewayinwhichthisistrue.)Webegin,first,withanumericalexample298
thatexemplifiesthis.299
6.1NumericalExampleofFitnessBeatingTruth300
Wegiveasimpleexampletopavethewayfortheideastofollow.Supposetherearethree301
states of the world,𝑊 = {𝑤R, 𝑤], 𝑤_}and two possible sensory stimulations,𝑋 = {𝑥R, 𝑥]}.302
Each world state can give rise to a sensory stimulation according to the information303
contained in Table 1. The first two columns give the likelihood values,ℙ 𝑥 𝑤 ,for each304
sensorystimulation,givenaparticularworldstate;forinstance,ℙ(𝑥R|𝑤]) = 3/4.Thethird305
column gives the prior probabilities of the world states. The fourth column shows the306
fitnessassociatedwitheachworldstate. Ifwe thinkof theworldstatesas threedifferent307
kindsoffoodthatanorganismmighteat,thenthesevaluescorrespondtothefitnessbenefit308
anorganismwouldgetbyeatingoneofthefoods.Withthisanalogy,𝑤Rcorrespondstoan309
extremely healthful food,while𝑤]and𝑤_correspond tomoderately healthful foods,with310
𝑤]beingmorehealthfulthan𝑤_(seeTable1).Thissetupisthebackdropforasimplegame311
where observers are presented with two sensory stimulations and forced to choose312
betweenthem.313
Page 16
Likelihood:𝒙𝟏given𝒘𝒋
ℙ(𝑥R 𝑤j
Likelihood:𝒙𝟐given𝒘𝒋
ℙ(𝑥] 𝑤j
Prior
ℙ 𝑤j
Fitness
𝑓 𝑤j
𝑤R 1/4 3/4 1/7 20
𝑤] 3/4 1/4 3/7 4
𝑤_ 1/4 3/4 3/7 3
Table1:Likelihoodfunctions,priorsandfitnessforoursimpleexamplewheretheTruth314
observerminimizesexpectedfitness,whileFitness-onlyobservermaximizesit.315
Using Bayes’ theorem we have calculated (see Appendix) that for𝑥Rthe Truth (i.e. the316
maximum-a-posteriori) estimate is𝑤], and that for𝑥]thisestimate is𝑤_.Thus, if aTruth317
observer isofferedachoicebetweentwofoodstoeat,onethatgives itstimulation𝑥Rand318
onethatgivesitstimulation𝑥],itwillperceivethatithasbeenofferedachoicebetweenthe319
foods𝑤]and𝑤_. Assuming that it has been shaped by natural selection to choose, when320
possible, the foodwith greater fitness, itwill alwaysprefer𝑤]. So,whenoffered a choice321
between𝑥Rand𝑥],theTruthobserverwillalwayschoose𝑥R,withanexpectedutilityof5.322
NowsupposeaFitness-Onlyobserverisgiventhesamechoice.TheFitness-Onlyobserver323
isnotatallconcernedwithwhich“veridical”foodthesesignalsmostlikelycorrespondto,324
buthasbeenshapedbynaturalselectiontoonlycareaboutwhichstimulusyieldsahigher325
expected fitness.We have calculated (see Appendix) that the expected fitness of sensory326
stimulation𝑥Ris 5 and the expected utility of stimulation𝑥]is 6.6. Thus, when offered a327
choice between𝑥Rand𝑥], the Fitness-Only observer will always, maximizing expected328
fitness,choose𝑥].329
Page 17
Theimplicationsoftheseresultsareclear.ConsiderapopulationofTruthobservers330
competingforresourcesagainstapopulationofFitness-Onlyobservers,bothoccupyingthe331
nichedescribedbyTable1.Since,inthiscase,theTruthobserver’schoiceminimizes332
expectedutilityandtheFitness-Onlyobserver’schoicemaximizesexpectedutility,the333
Fitness-OnlypopulationwillbeexpectedtodrivethepopulationofTruthobserversto334
extinction.Seeingtruthcanminimizefitness;therebyleadingtoextinction.Thisconclusion335
isapartfromconsiderationsoftheextraenergyrequiredtokeeptrackoftruth(seeMark,336
MarionandHoffman2010fordiscussiononenergyresources).337
338
7.MathematicalBackgroundfortheMainTheorem339
Weassumethatthereisafixedpreliminarymap,p,whichassociatestoeachworldstate340
𝑤 ∈ 𝑊asensorystate𝑥 ∈ 𝑋.AndweassumeafitnessmaponW(recallFigure3).This341
placestheTruthstrategyandtheFitness-onlystrategyonacommonfootingwherethey342
canbesetindirectcompetitionagainsteachotherwithinthecontextofanevolutionary343
resourcegame.344
Webeginwithsomemathematicaldefinitionsandassumptionsregardingthesespacesand345
maps.346
Itwillsufficeforabasicunderstandingofthedevelopmentinwhatfollows,tothink347
ofWas a finite set (as in the example in 6.1).4In general, we take theworldW to be a348
compactregularBorelspacewhosecollectionofmeasurableeventsisa𝜎-algebra,denoted349
ℬ.5Weassume that< 𝑊,ℬ >comesequippedwithanaprioriprobabilitymeasure𝜇on350
4inwhichcasealltheintegralsignsbelowcanbereplacedbysummations.5Anexampleisaclosedrectangleinsomek-dimensionalEuclideanspace,suchastheunitinterval[0,1]inonedimension,ortheunitsquareintwo.
Page 18
ℬ.Wewillconsideronlythoseprobabilitymeasures𝜇thatareabsolutelycontinuouswith351
respect to the Borel measure onℬ. That is, if we writed𝑤for the uniform, or Borel,352
probability measure on W, then the a priori measure satisfies𝜇 𝑑𝑤 = 𝑔 𝑤 d𝑤.Here353
𝑔:𝑊 → ℝMis some non-negative measurable function, called the density of𝜇,satisfying354
𝑔(𝑤)d𝑤 = 1.Wewilltakeanysuchdensitytobecontinuous,sothat italwaysachieves355
itsmaximumonthecompactsetW.Thisconstitutesthestructureoftheworld:astructure356
thatappliestomostbiologicalandperceptualsituations.357
We assume that a given species interacts with its world, employing a perceptual358
mapping that “observes” theworldvia ameasurablemap𝑝:𝑊 → 𝑋.We refer to this as a359
pureperceptualmap because it involves no dispersion: eachworld state can yield only a360
singlesensorystatex.WeassumethatthesetofperceptualstatesXisafiniteset,withthe361
standarddiscrete𝜎-algebra𝒳,i.e.,itspowerset(sothatallsubsetsofXaremeasurable).In362
thegeneralcase,theperceptualmapmayhavedispersion(ornoise),andismathematically363
expressed as aMarkovian kernel𝑝:𝑊×𝒳 → 0,1 .That is, for every elementw inW, the364
kernelpassignsaprobabilitydistributiononX(henceitassignsaprobabilityvaluetoeach365
measurablesubsetofX).BecauseX is finiteandallof itssubsetsaremeasurable,herethe366
kernelmaybeviewedsimplyasassigning,foreveryelementwinW,aprobabilityvalueto367
eachelementofX.368
7.1GeneralPerceptualMappingsandBayesianInference369
We use the letterℙto indicate any relevant probability. Bayesian inference consists in a370
computation of the conditional probability measureℙ(d𝑤|𝑥)on the world, given a371
particular perception𝑥in X. The likelihood function is the probabilityℙ(𝑥|𝑤)that a372
particular world state𝑤could have given rise to the observed sensory state𝑥.Then the373
Page 19
conditionalprobabilitydistributionℙ(𝑑𝑤|𝑥)istheaposteriorprobabilitydistributionina374
(partially)continuousversionofBayesformula:375
ℙ(d𝑤|𝑥) =ℙ(𝑥|𝑤)ℙ d𝑤
ℙ 𝑥.376
Since𝜇,theprioronW,hasadensity𝑔withrespecttotheBorelmeasured𝑤,wecanrecast377
thisformulaintermsof𝑔:indeed,ℙ(d𝑤|𝑥)alsohasaconditionaldensity,𝑔(𝑤|𝑥),with378
respecttotheBorelmeasure6andweobtain379
𝑔(𝑤|𝑥) =ℙ(𝑥|𝑤)𝑔 𝑤ℙ(𝑥|𝑤′)𝑔 𝑤′
.380
We now define amaximum a posteriori estimate for𝑥in X to be any𝑤* at which this381
conditional density is maximized:𝑔 𝑤* 𝑥) =max 𝑔 𝑤 𝑥 |𝑤 ∈ 𝑊}. At least one such382
maximumwillexist,since𝑔isboundedandpiecewisecontinuous;however,therecouldbe383
multiplesuchestimatesforeach𝑥.384
Foragivensensorystate𝑥, theonlyworldstatesthatcouldhavegivenriseto it lie inthe385
fiberover𝑥,i.e., theset𝑝KR 𝑥 ⊂ 𝑊.So, foragiven𝑥,themapping𝑤 → ℙ(𝑥|𝑤)takes the386
value1onthefiber,andiszeroeverywhereelse.Thismappingmaythusbeviewedasthe387
indicatorfunctionofthisfiber.Wedenotethisindicatorfunctionby1wxy * 𝑤 .388
Forapuremappingtheconditionaldensityisjust389
𝑔(𝑤|𝑥) =𝑔 𝑤 ⋅ 1wxy * 𝑤
𝜇 𝑝KR 𝑥,390
where𝜇 𝑝KR 𝑥 istheapriorimeasureofthefiber.391
6Thatis,ℙ(d𝑤|𝑥) = 𝑔(𝑤 𝑥 d𝑤.
Page 20
In this special case of a pure mapping that has given rise to the perception𝑥,we can392
diagram the fiber overxon which this average fitness is computed.This is the shaded393
regioninfigure3below.394
395
Figure4.Theexpectedfitnessof𝑥istheaverage,usingtheposteriorprobability,overthe396
fiber𝑝KR 𝑥 .397
7.2ExpectedFitness398
Givenafitnessfunction𝑓:𝑊 → [0,∞)thatassignsanon-negativefitnessvaluetoeach399
worldstate,theexpectedfitnessofaperception𝑥is400
𝐹 𝑥 = 𝑓 𝑤 ℙ(d𝑤|𝑥) = 𝑓 𝑤 𝑔(𝑤|𝑥)d𝑤 .401
7.3TwoPerceptualStrategies.402
Wemaybuildourtwoperceptualstrategies𝑃}, 𝑃~ ,called“Truth”and“Fitness-Only”403
respectively,ascompositionsofa“sensory”map𝑝:𝑊 → 𝑋thatrecognizesterritoriesand404
!
W p-1(x)
!.!X! X
Page 21
“ordering”maps𝑑}, 𝑑~: 𝑋 → 𝑋,where𝑃} = 𝑑} ∘ 𝑝and𝑃~ = 𝑑~ ∘ 𝑝.Thatis,themap𝑑} re-405
namestheelementsofXbyre-orderingthem,sothatthebestone,intermsofitsBayesian406
MAPestimate,isnowthefirst,𝑥R,thesecondbestis𝑥]etc.Themap𝑑~ ,ontheotherhand,407
re-orderstheelementsofXsothatthebestone,intermsofitsexpectedfitnessestimate,is408
𝑥R,thesecondbestis𝑥]etc.Theorganismpicks𝑥Rifitcan,𝑥]otherwise.409
Wecannowassertourmaintheorem,invariouscontextsofevolutionarygames:with410
infinitepopulations,finitepopulationswithfullselection,andsufficientlylargefinite411
populationswithweakselection.412
413
8.Results414
8.1The“FitnessBeatsTruth”Theorem415
Thefollowingtheoremappliestoinfinitepopulations,ortolargefinitepopulations416
includingthosewithweakselection:417
Theorem4:Overallpossiblefitnessfunctionsandapriorimeasures,theprobabilitythatthe418
Fitness-onlyperceptualstrategystrictlydominatestheTruthstrategyisatleast( 𝑋 −419
3)/( 𝑋 − 1),where 𝑋 𝑖𝑠𝑡ℎ𝑒𝑠𝑖𝑧𝑒𝑜𝑓𝑡ℎ𝑒𝑝𝑒𝑟𝑐𝑒𝑝𝑡𝑢𝑎𝑙𝑠𝑝𝑎𝑐𝑒.Asthissizeincreases,this420
probabilitybecomesarbitrarilycloseto1:inthelimit,Fitness-onlywillgenericallystrictly421
dominateTruth,sodrivingthelattertoextinction.422
Proof:Foranygiven𝑥,theBayesianMAPestimateisaworldpoint𝑤* (itisthe𝑤*suchthat423
𝑔 𝑤* 𝑥) =max 𝑔(𝑤|𝑥 𝑤 ∈ 𝑊 ). This point has fitness𝑓 𝑤* ;let𝑥� be that𝑥forwhich424
thecorresponding𝑓 𝑤* ismaximized.Thenthis𝑥� is,ifavailable,ischosenbyTruthand425
𝐹(𝑥�),itsexpectedfitness,isthepayofftoTruth.426
Page 22
On the other hand, the fitness payoff to the Fitness-only strategy is, by definition the427
maximumexpectedfitness𝐹(𝑥�)overallfibers,soclearly,𝐹(𝑥�) ≤ 𝐹(𝑥�).428
Asdefined earlier, our evolutionary gamehas as payoffs,a:toFitness-onlywhenplaying429
against Fitness-only; b: to Fitness-only when playing against Truth; c: to Truth when430
playingagainstFitness-only;d:toTruthwhenplayingagainstTruth.431
Weneedtoestimatetheprobability that and Weassumethat ifbothstrategies432
arethesame,theneachhasanevenchanceofpickingitsbestterritoryfirst.Thusif,inany433
givenplay of the game, two competing strategies both take aparticular territory as their434
mostfavoredone,theneachstrategyhasanevenchanceofpickingthatterritoryandthen435
theotherstrategypicksitsnext-bestchoiceofterritory.436
If Fitness-only meets Fitness-only, then each has an even chance of choosing its best437
territory, say𝑥�; thesecond tochoose thenchooses its secondbest territory, say𝑥��. Since438
eachplayerhasanequalchanceofbeingfirst,wehave439
𝑎 = 𝐹 𝑥� + 𝐹 𝑥�� /2.440
IfTruthmeetsFitness-only,itschoicewillbe𝑥� ,aslongasthisvaluediffersfrom𝑥� .Inthis441
instance, we have𝑎 > 𝑐. If, however,𝑥� = 𝑥�,half the timeTruth will choose𝑥� and the442
otherhalf𝑥�� , where𝑥�� isthesecondbestoftheoptimalterritoriesfor𝑻𝒓𝒖𝒕𝒉.Hence443
𝑐 =𝐹(𝑥�),ifdifferentbestterritories𝐹(𝑥�) + 𝐹 𝑥��
2,ifsamebestterritories
444
andsince𝐹 𝑥� ≤ 𝐹 𝑥� and𝐹(𝑥�� ) ≤ 𝐹 𝑥�� weget 445
WhathappenswhenFitness-onlymeetsTruth?IfFitness-onlygoesfirst,thepayoffwillbe446
𝑏 = 𝐹(𝑥�).ThesameistrueifTruthgoesfirstandthetwobestterritoriesaredifferent.If,447
!a ≥c !!b ≥d.
!!a ≥c.
Page 23
however,thetwobestterritoriesarethesame,thenthepayofftoFitness-onlyisitssecond-448
bestoutcome:449
𝑏 =𝐹(𝑥�),ifdifferentbestterritories𝐹 𝑥�� ,ifsamebestterritories
450
Finally,whenTruthmeetsTruth,wehavethat451
𝑑 =𝐹 𝑥� + 𝐹 𝑥��
2.452
Soitisclearthat𝑏 ≥ 𝑑,aslongasthetwobestterritoriesaredifferent.Iftheyarethesame,453
this may or may not be true: it depends on the relative size of the average d and𝐹 𝑥�� 454
(which,inthisinstance,alsoliesinbetween𝐹 𝑥�� and𝐹(𝑥�) = 𝐹 𝑥� ).455
Now,apriori,thereisnocanonicalrelationbetweenthefunctionsfandg,bothofwhichcan456
beprettymucharbitrary(infact,fneednotevenbecontinuousanywhere,andcouldhave457
big jumps as well as bands of similar value separated from each other in W). Also,458
generically themaximum foreachstrategywillbeuniqueandalso theexpected fitnesses459
forthedifferentterritorieswillallbedistinct.460
Thus,generically,𝐹(𝑥�)and𝐹 𝑥�� willbedifferentfromandindeedstrictlylessthan𝐹(𝑥�)461
(and also𝐹 𝑥�� < 𝐹 𝑥�� ). The only impediment to the domination of Fitness-Only can462
come fromthesituationwhere thebest territories forbothstrategiesare thesame.LetX463
havesize 𝑋 = 𝑛.Thereare𝑛waysthetwostrategiescanoutputthesameterritory,outof464
the𝑛! [2! 𝑛 − 2 !]waysofpairingterritories.Thus,acrossallpossibilitiesfor fandg, the465
probabilitythatrandomlychosenfitnessandapriorimeasureswouldresultinchoosingthe466
sameterritoryforbothstrategies,i.e.,that𝐹 𝑥� = 𝐹(𝑥�),willhappenwithaprobabilityof467
𝑛𝑛!2! 𝑛 − 2 !
=2
𝑛 − 1468
Page 24
Finally, theprobabilityofthetwofibersbeingdifferent isthecomplement:1 − ] KR
= K_ KR
.469
470
8.1DynamicFitnessFunctions471
Apossibleobjectiontotheapplicabilityof thistheoremisthat itseemstoassumeastatic472
fitnessfunction,whereasrealisticscenariosmayinvolvechanging,orevenrapidlychanging,473
fitnessfunctions.However,aclosescrutinyoftheproofofthetheoremrevealsthatatany474
moment, the fitness functionatthattime being the same for both strategies, the relative475
payoffsremaininthesamegenericrelationasatanyothermoment.Hencethetheoremalso476
appliestodynamicallychangingfitnessfunctions.477
478
9.Discussion479
AswenotedintheIntroduction,itisstandardintheliteraturetoassumethatmoreaccurate480
perceptionsarefitterperceptionsandthat,therefore,naturalselectiondrivesperceptionto481
increasingveridicality—i.e.tocorrespondincreasinglytothe“true”stateoftheobjective482
world.Thisassumptioninformstheprevalentviewthathumanperceptionis,forthemost483
part,veridical.484
Ourmainmessageinthispaperhasbeenthat,contrarytothisprevalentview,attemptingto485
estimatethe“true”stateoftheworldcorrespondingtoagivenasensorystate,confersno486
evolutionarybenefitwhatsoever.Ratherastrategythatsimplyseekstomaximizeexpected-487
fitnesspayoff,withnoattempttoestimatethe“true”worldstate,doesconsistentlybetter488
(intheprecisesensearticulatedinthestatementofthe“FitnessBeatsTruth”Theorem).489
Indeed,this“Fitness-only”strategydoesnotestimateanysingleworldstate;itsimply490
!
Page 25
averagesoverallpossibleworldstatestocomputetheexpected-fitnesspayoff491
correspondingtoanygivensensorystate(thisisanalogoustoamodel-averagingstrategyin492
modelselection).Andyet,asthetheoremshows,inanevolutionarycompetition,this493
strategyislikelytodrivethe“truth”strategytoextinction.494
Atfirstglance,thisexpected-fitnessstrategy,basedonaveragingoverallpossibleworld495
states,mayseemimplausible:Afterall,inourownperceptualexperience,weperceive496
thingstobeoneparticularway;wecertainlydon’texperienceasuperpositionor“smear”497
resultingfromaveragingovervariouswaysthattheworldcouldbe.Whilethisis498
undoubtedlytrue,oneshouldnotethatthisisafactaboutperceptualexperience,and499
providesnosupportwhatsoeverforastrategythatinvolvesestimatingthe“true”stateof500
theworld.Inwhatfollows,wesketchoutamorecompleteanswertotheseeming501
implausibilityofaveraging,basedonourInterfaceTheoryofPerception(Hoffman,Singh,&502
Prakash,2015).503
Forthepurposeofthecurrentanalysis,itwasessentialtoplacethetwostrategiestobe504
compared—“Truth”and“Fitness-only”—withinacommonframeworkinvolvingBayesian505
inferencefromthespaceofsensorystates,X,totheworld,W(recallFigure3).Thisallowed506
ustoplacethetwostrategiesonthesamefooting,sotheycouldcompetedirectlyagainst507
eachother.However,thisresultstronglysupportsourbeliefthattheveryideaofperception508
asprobabilisticinferencetostatesoftheobjectiveworldismisguided.Perceptionisindeed509
fruitfullymodeledasprobabilisticinference,buttheinferencehappensinaspaceof510
perceptualrepresentations,andnotinanobjectiveworld.511
Theseideasarepartoflargertheory,theInterfaceTheoryofPerception,thatwehave512
describedindetailelsewhere(Hoffman,2009;Hoffman&Prakash,2014;Hoffman&Singh,513
2012;Hoffman,Singh,&Prakash,2015;seealsoKoenderink,2011;2013;2014;von514
Page 26
Uexkull,1934).Forthepurposesofthecurrentdiscussion,thekeypointisthatthestandard515
Bayesianframeworkforvisualperceptionconflatestheinterpretationspace(orthespace516
ofperceptualhypothesesfromwhichthevisualsystemmuchchoose)withtheobjective517
world.Thisisamistake;itisessentiallytheassumptionthatthelanguageofourperceptual518
representationisthecorrectlanguagefordescribingobjectivereality—ratherthansimplya519
species-specificinterfacethathasbeenshapedbynaturalselection.InourITPframework,520
theprobabilisticinferencethatresultsinperceptualexperiencetakesplaceinaspaceof521
perceptualrepresentations,say,X1,thatmayhavenoisomorphicorevenhomomorphic522
relationwhatsoevertoW.TheextendedframeworkofthisComputationalEvolutionary523
PerceptionissketchedinFigure5(seeHoffman&Singh,2012;Hoffman,Singh,&Prakash,524
2015;Singh&Hoffman,2013).525
526
Figure5.TheframeworkofComputationalEvolutionaryPerceptioninwhichperceptual527
inferencestakeplaceinaspaceofrepresentationsX1thatisnotisomorphicor528
homomorphictoW.ThemorecomplexrepresentationalformatofX1evolvesbecauseit529
permitsahigher-capacitychannelP1 :W → X1 forexpectedfitness,therebyallowingthe530
organismtochooseandactmoreeffectivelyintheenvironment(i.e.inwaysthatresultin531
higherexpected-fitnesspayoffs).532
Page 27
533
Thus,thereasonwegenerallyperceiveasingleinterpretationisbecausetheprobabilistic534
inferenceintheperceptualspaceX1generallyresultsinauniqueinterpretation.Butthe535
perceptualspaceX1isnottheobjectiveworld,norisithomomorphictoit.Itissimplya536
representationalformatthathasbeencraftedbynaturalselectioninordertosupportmore537
effectiveinteractionswiththeenvironment(inthesenseofresultinginhigherexpected-538
fitnesspayoff).Inotherwords,amorecomplexorhigher-dimensionalrepresentational539
format(e.g.involving3Drepresentationsin𝑋R,inplaceof2Drepresentationsin𝑋")evolves540
becauseitpermitsahigher-capacitychannelP1 :W → X1 forexpectedfitness(seeFigure541
5).Butthisdoesnotinanywayentailthatthisrepresentationalformatsomehowmore542
closely“resembles”theobjectiveworld.Evolutioncanfashionperceptualsystemsthatare,543
inthissense,ignorantoftheobjectiveworldbecausenaturalselectiondependsonlyon544
fitnessandnotonseeingthe“truth.”545
Theseconsiderationsstronglyunderminethestandardassumptionsthatseeingmore546
veridicallyenhancesfitness,andthatthereforeonecanexpectthathumanperceptionis547
largelyveridical.Ashumanobservers,wearepronetoimputingstructuretotheobjective548
worldthatisproperlypartofourownperceptualexperience.Forexample,ourperceived549
worldisthree-dimensionalandpopulatedwithobjectsofvariousshapes,colors,and550
motions,andsowetendtoconcludethattheobjectiveworldisaswell.Butif,astheFitness-551
beats-TruthTheoremshows,evolutionarypressuresdonotpushperceptioninthedirection552
ofbeingincreasinglyreflectiveofobjectivereality,thensuchimputationshavenological553
basiswhatsoever.7554
Acknowledgments5557SeealsotheInventionofSpace-TimeTheoreminHoffman,Singh,&Prakash(2015).
Page 28
Wethank,FedericoFagginforenlighteningdiscussions.Thisworkhasbeenpartially556
fundedbytheFedericoandElviaFagginFoundation.557
558
Page 29
References 559
1. AdelsonE.H.,&PentlandA.(1996).Theperceptionofshadingandreflectance. In:D560Knill andWRichards (Eds.), Perception as Bayesian Inference, pp. 409-423 (New561York:CambridgeUniversityPress).562
2. Chemero,A.(2009).RadicalEmbodiedCognitiveScience.Cambridge,MA:MITPress.5633. Dennett,D.C.(1995).Darwin'sDangerousIdea.NewYork:Simon&Schuster.5644. FeldmanJ.(2013)Tuningyourpriorstotheworld.TopicsinCognitiveScience,5,13-565
34.5665. Geisler, W. S. and Diehl, R. L. (2003). A Bayesian approach to the evolution of567
perceptualandcognitivesystems.CognitiveScience27,379-402.5686. Hofbauer,J.,andSigmund,K.(1998).Evolutionarygamesandpopulationdynamics.569
Cambridge:CambridgeUniversityPress.5707. Nowak,MartinA.(2006).EvolutionaryDynamics.HarvardUniversityPress.5718. Hoffman,D.D.(2009).Theinterfacetheoryofperception.InS.Dickinson,M.Tarr,A.572
Leonardis, B. Schiele, (Eds.) Object categorization: computer and human vision573perspectives,NewYork:CambridgeUniversityPress,pp.148–165.574
9. Hoffman, D. D., & Prakash, C. (2014). Objects of consciousness. Frontiers of575Psychology,5:577.DOI:10.3389/fpsyg.2014.00577.576
10. Hoffman, D. D., & Singh, M. (2012). Computational evolutionary perception.577Perception,41,1073–1091.578
11. Hoffman, D. D., Singh, M., & Mark, J. T. (2013). Does natural selection favor true579perceptions?Proceedings of the SPIE8651,HumanVision andElectronic Imaging580XVIII,865104.DOI:10.1117/12.2011609.581
12. Hoffman, D. D., Singh, M., & Prakash, C. (2015). Interface Theory of Perception.582PsychonomicBulletin&Review.DOI10.3758/s13423-015-0890-8583
13. Kleffner,D.,&Ramachandran,V.(1992).Ontheperceptionofshapefromshading.584PerceptionandPsychophysics,52,18–36.585
14. Knill, D. and Richards, W. (1996). Perception as Bayesian inference. Cambridge586UniversityPress,Cambridge,UK.587
15. Koenderink, J. J., van Doorn, A., de Ridder, H., Oomes, S. (2010). Visual rays are588parallel.Perception,39,1163-1171.589
16. Koenderink, J. J. (2011). Vision as a user interface. Human Vision and Electronic590Imaging XVI, Interface theory of perception 55 SPIE Vol. 7865. doi:59110.1117/12.881671.592
17. Koenderink,J.J.(2013).World,environment,umwelt,andinner-world:abiological593perspectiveonvisualawareness,HumanVisionandElectronic ImagingXVIII,SPIE594Vol.8651.doi:10.1117/12.2011874.595
18. Koenderink,J.J.(2014).Theallseeingeye?Perception,43,1–6.59619. Li,Y.,Sawada,T.,Shi,Y.,Steinman,R.M.,&Pizlo,Z.(2013).Symmetryisthesinequa597
non of shape. In S. Dickinson & Z. Pizlo (Eds.), Shape perception in human and598computervision(pp.21–40).London:Springer.599
20. Loomis, J. M. (2014). Three theories for reconciling the linearity of egocentric600distance perception with distortion of shape on the ground plane. Psychology &601Neuroscience,7,3,245-251.602
21. MamassianP,LandyM,MaloneyLT,2002“Bayesianmodelingofvisualperception”,603inProbabilisticModelsof theBrain:PerceptionandNeuralFunction,Ed.RRao,B604Olshausen,MLewicki(Cambridge,MA:MITPress)pp13–36.605
Page 30
22. Marion, B. The Impact of Utility on the Evolution of Perceptions. Dissertation,606UniversityofCalifornia,Irvine,2013.607
23. Mark, J., Marion, B. and Hoffman, D. D. (2010). Natural selection and veridical608perception.JournaloftheoreticalBiology266,504-515.609
24. Marr,D.(1982).Vision.SanFrancisco:Freeman.61025. Moran,P.A.P.1958.Randomprocessesingenetics.P.Camb.Philos.Soc.54:60–71.611
1962.Thestatisticalprocessesofevolutionarytheory.Oxford:ClarendonPress.61226. Nowak,M.A.(2006).EvolutionaryDynamics:ExploringtheEquationsofLife.Belknap613
HarvardUniversityPress,Cambridge,MA.61427. Palmer,S.(1999).VisionScience:PhotonstoPhenomenology.MITPress,Cambridge,615
MA.61628. Pizlo,Z.,Li,Y.,Sawada,T.,&Steinman,R.M.(2014).Makingamachinethatseeslike617
us.NewYork:OxfordUniversityPress.61829. Shepard, R. (1994). Perceptual-cognitive universals as reflections of the world.619
PsychonomicBulletin&Review,1(1),2-28.62030. Singh,M.,&Hoffman,D.D.(2013).Naturalselectionandshapeperception:shapeas621
aneffectivecodeforfitness.InS.DickinsonandZ.Pizlo(Eds.)Shapeperceptionin622humanandcomputervision:an interdisciplinaryperspective.NewYork:Springer,623pp.171–185.624
31. Taylor, C, Fudenberg, D., Sasaki, A. & Nowak, M. Evolutionary Game Dynamics in625FinitePopulations.BulletinofMathematicalBiology(2004)66,1621–1644626
32. Taylor, P. D., and L. B. Jonker. 1978. “Evolutionary stable strategies and game627dynamics.”Math.Biosci.40:145–156.628
33. vonUexkull, J.(1934).Astrollthroughtheworldsofanimalsandmen:Apicturebook629of invisible worlds. Reprinted in Instinctive behavior: Development of a modern630concept,C.H.Schiller(1957).NewYork:HallmarkPress.631
34. Yuille, A., & Bulthoff, H. (1996). Bayesian decision theory and psychophysics. In:632Perception as Bayesian inference, D. Knill and W. Richards, Eds., New York:633CambridgeUniversityPress.634
635
Appendix:CalculationsforthenumericalexampleinTable1.636
InthisappendixweperformtheBayesianandexpectedfitnesscalculationsusingthedata637
giveninTable1.638
TocomputetheTruthestimates,wefirstneedtheprobabilityofeachstimulationℙ(𝑥R)and639
ℙ(𝑥]).Thesecanbecomputedbymarginalizingoverthepriorsintheworldasfollows:640
ℙ(𝑥R) = 𝑝(𝑥R 𝑤R 𝜇 𝑤R + p 𝑥R 𝑤] 𝜇 𝑤] + p 𝑥R 𝑤_ 𝜇 𝑤_ = R¡. R¢+ _
¡. _¢+ R
¡. _¢= R_
]£641
ℙ(𝑥]) = 𝑝(𝑥] 𝑤R 𝜇 𝑤R + p 𝑥] 𝑤] 𝜇 𝑤] + p 𝑥] 𝑤_ 𝜇 𝑤_ = _¡. R¢+ R
¡. _¢+ _
¡. _¢= R¤
]£642
Page 31
ByBayes’Theorem,theposteriorprobabilitiesoftheworldstates,given𝑥R,are643
𝑝(𝑤R 𝑥R = 𝑝 𝑥R 𝑤R .𝜇 𝑤Rℙ 𝑥R
=14.17/1328
=113644
𝑝(𝑤] 𝑥R = 𝑝 𝑥R 𝑤] .𝜇 𝑤]ℙ 𝑥R
=34.37/1328
=913645
𝑝(𝑤_ 𝑥R = 𝑝 𝑥R 𝑤_ .𝜇 𝑤_ℙ 𝑥R
=14.37/1328
=313646
Thusthemaximumaposteriori,orTruthestimateforstimulus𝑥Ris𝑤].647
Posteriorprobabilitiesoftheworldstates,given𝑠],are:648
𝑝(𝑤R 𝑥] = 𝑝 𝑥] 𝑤R .𝜇 𝑤Rℙ 𝑥]
=34.17/1528
=15649
𝑝(𝑤] 𝑥] = 𝑝 𝑥] 𝑤] .𝜇 2ℙ 𝑥]
=14.37/1528
=15650
𝑝(𝑤_ 𝑥] = 𝑝 𝑥] 𝑤_ .𝜇 𝑤_ℙ 𝑥]
=34.37/1528
=35651
Thusthemaximumaposteriori,orTruthestimateforstimulus𝑥]is𝑤_.652
Finally, the expected-fitness values of the different sensory stimulations𝑥Rand𝑥]are,653
respectively:654
𝐹 𝑥R = 𝑝(𝑤R 𝑥R 𝑓 𝑤R + 𝑝(𝑤] 𝑥R 𝑓 𝑤] + 𝑝(𝑤_ 𝑥R 𝑓 𝑤_ =113. 20 +
913. 4 +
313. 3 = 5;655
𝐹 𝑥] = 𝑝(𝑤R 𝑥] 𝑓 𝑤R + 𝑝(𝑤] 𝑥] 𝑓 𝑤] + 𝑝(𝑤_ 𝑥] 𝑓 𝑤_ = R¤. 20 + R
¤. 4 + _
¤. 3 = 6.6.656
Thus𝑥]hasalargerexpectedfitnessthan𝑥R.657
Page 32
658
659
660
661
Highlights662
• Wemakerigorousmathematicaldefinitionsoftwoperceptualstrategiesemployable663byagivenspecies,foragivenactionclassandwithinagivenenvironment:Truth,664basedonBayesianestimationofassumedobjectivepropertiesoftheworld,and665Fitness,tunedtoanarbitraryfitnessfunction;666
• UndertheassumptionofuniversalDarwinism(Dennett,1995)wesubjectthetwo667strategiestoanevolutionarygameanalysis;668
• WeconcludethattheFitnesswillgenerallydriveTruthtoextinction,forgeneric669fitnessfunctionsandpriors;670
• ThelikelihoodofFitnessdominatingTruthexceeds1/2assoonasthesensoriumhas671morethanfiveelements,andrisesmonotonicallyto1asthesizeofthesensorium672growstowardsinfinity;673
• Thistheoremholdsinthepresenceofchangingfitnessfunctionsandforlargefinite674populations.675
676