Philosophical Aspects of AI

Arya MirΤεχνίτη Νοημοσύνη και Ρομποτική

26 Νοε 2011 (πριν από 5 χρόνια και 8 μήνες)

897 εμφανίσεις

Philosophical questions are never far below the surface in AI! For example: What is intelligence? What is artificial intelligence? Can a machine think? (cf. Alan Turing) Is the brain a computer? (cf. John Searle) How can thoughts be about something? What is ‘free will’? Can a computer exercise free will (like Hal in 2001: A Space Odyssey)? What is consciousness? Is consciousness essential to thought? Is machine consciousness possible?

COMP2039ArtificialIntelligence:
PhilosophicalAspectsofAI
BobDamper
April24,2006
COMP2039ArtificialIntelligence:PhilosophicalAspectsofAI–p.1/19
IntroductoryRemarks
Philosophicalquestionsareneverfarbelowthe
surfaceinAI!Forexample:
Whatisintelligence?Whatisartificialintelligence?
Canamachinethink?(cf.AlanTuring)
Isthebrainacomputer?(cf.JohnSearle)
Howcanthoughtsbeaboutsomething?
Whatis‘freewill’?Canacomputerexercisefree
will(likeHalin2001:ASpaceOdyssey)?
Whatisconsciousness?Isconsciousness
essentialtothought?Ismachineconsciousness
possible?
Recommendedreading...B.JackCopeland,Artificial
Intelligence:APhilosophicalIntroduction,Blackwell
Scientific,Oxford,UK,1993.
COMP2039ArtificialIntelligence:PhilosophicalAspectsofAI–p.2/19
Searle’sFamousChineseRoom
TheChineseroomargument(CRA)isacelebrated
thoughtexperimentduetoJohnSearle(1980).
Itisintendedtoshowthefutilityofattemptsby
computerscientiststoimplementwhatSearlecalls
‘strongAI’.
StrongAIclaimsthatintelligence/thinkingismerelythe
manipulationofformalsymbols(cf.thephysical
symbolsystemhypothesisofAllenNewell).
Putotherwise,“theappropriatelyprogrammed
computerliterallyisamind”.
This,claimsSearle,isnonsense,astheCRAis
designedtoshow.
In1991,GomilladescribedtheliteratureontheCRA
as“nearlyinfinite”.
COMP2039ArtificialIntelligence:PhilosophicalAspectsofAI–p.3/19
OutlineoftheCRA
Searle,hiddeninaroom,acceptsquestionsinChinese
writtenonan‘input’cardandpostedintohisroom.
SearleknowsnoChinese.
Heconsultsamanualcontainingrules(inEnglish)
tellinghimpreciselywhatsquigglesandsquogglesto
writeonan‘output’cardwhichhepostsback.
Byvirtueofthemanual,thesesquigglesand
squogglesconstitutetheanswertoanyinputquestion.
Hence,Searle-in-the-roompassestheTuringtest,but
heknowsnoChinese!
SoanAIprogram(correspondingtothemanual)couldpass
theTuringtest,buthavenounderstanding,contrarytothe
tenetsofstrongAI.
COMP2039ArtificialIntelligence:PhilosophicalAspectsofAI–p.4/19
RepliestotheCRA
TheCRAhasproducedintensedebate/controversy.
Almostnoonethinksitisright!Buttherearemany
verydifferentreasonsgivenforthinkingitiswrong.
ThreeofthesewerediscussedbySearleintheoriginal
(1980)article:
Systemsreply:intelligenceresidesinthetotal
systemnotjustinSearlehimself.
Robotreply:replacethedisembodiedAIprogram
byarobotwithsensors,effectors,etc.Thisis
‘intelligent’inthewaythatahumanis.
Brainsimulatorreply:insteadoftheAIprogram,
simulatetheactualneuronsofaChineseperson
answeringquestionsinChinese.Thiswouldbean
‘intelligent’computerprogram.
COMP2039ArtificialIntelligence:PhilosophicalAspectsofAI–p.5/19
MoreRepliestotheCRA
Searleidentifiessomeotherreplies:
Combinationreply:somekindofcombinationofthe
systems,robotandbrainsimulatorreplies.
Otherminds:howdoweknowthatSearleunderstands
Englishwhenheclaimsto?Onlybecausehumans
naturallyascribeintelligence(‘intentionality’)toother
humans.SoifweascribeintentionalitytoSearle,we
mustdothesameforthecomputerprogram.
Manymansions:therearemanypossiblekindsof
‘computer’and‘computation’.Inthefuture,wemay
haveunimaginedformsof‘computer’whichwould
displayartificialintelligence.
Allofthese“missthepoint”accordingtoSearle.
COMP2039ArtificialIntelligence:PhilosophicalAspectsofAI–p.6/19
TheSystemsReply
ThemostpopularargumentagainsttheCRA.
Searle:imaginethatthenon-Chinese-speakinghuman‘internalizes’
everything.Thereisthennothingbutthehumanwhostilldoesnot
understandChinese.
AIpractitioner:becausethesystem–orthehuman–asawhole
doesnotunderstand,itdoesnotfollowthatpartofthe
system/humancannotunderstand.
Searle:butwhateversubsystemofthewholeisprocessing
Chinese,itisstilldoinguninterpretedsymbolmanipulationandso
doesnotunderstand.
AIpractitioner:thatbegsthequestionbyassumingthetruthof
theCRA!
Searle:thesystemsresponseitselfbegsthequestioninthefirst
placebyassumingwithoutargumentthatthesystemunderstands!
COMP2039ArtificialIntelligence:PhilosophicalAspectsofAI–p.7/19
TheRobotReply
Searle:theadditionofsensors,effectors,etc.changesnothingby
wayofunderstanding.The‘computerinsidetherobot’continuesto
manipulateuninterpretedsymbolsandstilldoesnotunderstand
Chinese.
AIpractitioner:asbefore,thisbegsthequestionbyassumingthe
truthoftheCRA!
Searle:asbefore,therobotresponsebegsthequestionby
assumingthattherobotunderstands!
AIpractitioner:butiftherobot’ssensors,effectorsandbodyare
giventheexternalappearanceofahuman,itbecomes
indistinguishablefromahumanChineseunderstander–andthere
canbenovaliddistinctiondrawnbetweenindistinguishables!
Searle:thisistheothermindsresponse,whichIhavealready
answered.
COMP2039ArtificialIntelligence:PhilosophicalAspectsofAI–p.8/19
TheBrainSimulatorReply
Searle:bythisargument,proponentsofstrongAIhaveabandoned
thesymbolsystemhypothesiswhichstatesthatonedoesn’thaveto
knowhowthebrainworks!
AIpractitioner:wearenotsayingthatyouhavetoknowhowthe
brainworks!AIisimplementation-independent.Thepointofthe
brainsimulationisthatthisisacomputerprogramthatyouhaveto
admitisintelligent.
Searle:insteadofsimulatingneuralcircuitry,supposeIsimulateda
circuitofwaterpipeswiththesamefunction.TheCRAstillapplies.
Thesimulationsimulatesthewrongthingaboutthebrain–itsformal
symbolmanipulationproperties.Itmissestheimportantthing–its
causalproperties.
AIpractitioner:whatare“itscausalproperties”?Youdon’ttellus
andyoudon’tseemtoknow.
COMP2039ArtificialIntelligence:PhilosophicalAspectsofAI–p.9/19
SomeMoreProsandCons–I
Thebiologicalobjection
Searle:IamabletounderstandEnglish...becauseIamacertain
sortoforganismwithacertainbiological(i.e.chemicalandphysical
structure).
AIpractitioner:thisdoesnotmeanthatsomeotherartifactmight
alsohavetheright(butdifferent)structure–asSearlehimself
agrees!
Searle’sintuitions
AIpractitioner:theCRAisn’tanargument–it’sSearle’sintuitions
dressedupasargument.Myintuitionsaredifferent!StrongAIis
possible!
Searle:it’safact,notanintuition,thatIdon’tunderstandChinese.
COMP2039ArtificialIntelligence:PhilosophicalAspectsofAI–p.10/19
SomeMoreProsandCons–II
TheimplausibilityofaChineseunderstandingprogram
AprogramtounderstandChineseiswaybeyondthecapabilitiesof
currentAI.Wedon’tevenknowifit’spossible.
Copeland(p.127)refersto“thebuilt-inabsurdityofSearle’s
scenario”.
Sowhatcanwevalidlyinferfromanabsurdity?
TheArithmeticRoom
Supposewesubstitutedquestionsinarithmeticforeveryday
questionsexpressedinChinese.
Thissimplerscenarioisamenabletoexperimentaltestofthe
hand-implementor’sunderstanding(orlackofit).
Butmaybecomputerarithmeticisn’t‘intelligent’?
COMP2039ArtificialIntelligence:PhilosophicalAspectsofAI–p.11/19
SimulationandReality
“Noonesupposesthatasimulationofastormwillleaveus
wet...Why...wouldanyonesupposeacomputersimulation
ofmentalprocessesactuallyhadmentalprocesses?”(Searle,
1984)
Copelandsaystherearetwodifferentkindsofsimulation:
Simulation
1
:lackssomeessentialdetailsofwhatis
beingsimulated,e.g.,computersimulationofarain
storm.
Simulation
2
:hasalltheessentialdetailsbuthas
merelybeenproducedbyartificialmeans,e.g.,
high-qualitysyntheticspeech.
Searleassumestheonlykindissimulation
1
butstrongAI
aimsforsimulation
2
ofthebrain.
COMP2039ArtificialIntelligence:PhilosophicalAspectsofAI–p.12/19
TheSymbolGroundingProblem
TheCRAisoneversionofthesymbolgrounding
problem(Harnad1990).
“Howcanthesemanticinterpretationofaformal
symbolsystembemadeintrinsictothesystem,rather
thanjustparasiticonthemeaningsinourheads?”
TheSGPconsistsinfindinga‘mapping’betweenthe
otherwisemeaninglesssymbolsmanipulatedbya
computerandtheexternalworld.
AccordingtoSearle,thisisexactlywhatismissing
whenhehand-implementstheAIprogram.
Usually,thesymbolsemployedinacomputerprogram
areinterpretedbyahuman,buttheremaynotbeone
inanautonomoussystem–andthereisn’tonein
cognition!
COMP2039ArtificialIntelligence:PhilosophicalAspectsofAI–p.13/19
TheCRAasaThoughtExperiment
TheCRAisaverytypical‘thoughtexperiment’.
(Whatotherthoughtexperimentdoyouknow?What
arethesimilaritiestotheCRA?Whatarethe
differences?)
IntheterminologyofJ.R.Brown,itisa‘destructive’
thoughtexperiment,designedtodisproveatheory.
The‘theory’isstrongAI.
Thoughtexperimentsworkbyinvokingintuitions.
Theyarethereforealwayspresentedinnarrativeform
–asa’story’–ineverydaylanguage.
Buteverydaylanguageisvagueandambiguous.
Soisthereanadvantagetoseekingaformallogical
schemafortheCRA?Willthishelpusresolvethe
debate?
COMP2039ArtificialIntelligence:PhilosophicalAspectsofAI–p.14/19
ALogicalStructurefortheCRA?
PeijnenburgandAtkinson(2003)offerthefollowingsimple
(modustollens)structureofadestructivethought
experimentintermsofpropositionallogic:
(T∧E)⇒S
¬S
E⊢¬T
Here,Tisatheory,EisathoughtexperimentandSis
somesituationwhicheveryoneknowsisnotthecase.
TisstrongAIandEistheCRAitself.Thepremiseisthat
TandEtogetherimplyS,namelythatSearleunderstands
Chinese.But,sotheargumentgoes,everyoneknowsthat
SearledoesnotunderstandChinese,¬S.Hence,¬Tand
strongAIisfalse.
COMP2039ArtificialIntelligence:PhilosophicalAspectsofAI–p.15/19
UsingModalLogic
Perhapspropositionallogicisnotpowerfulenoughto
capturethefullcomplexityoftheCRA.
Whataboutmodallogic?
Amodalisanexpression(like‘necessarily’or
‘possibly’)thatisusedtoqualifythetruthofa
judgement.
AsBunzl(1996)says“‘questionsofpossibilityand
impossibilityenteratthegroundfloor’wherethought
experimentsareconcerned”.
Conventionally,themodalities‘itisnecessarythat’and
‘itispossiblethat’aresymbolisedas
and
,
respec.,relatedas:
A
def=¬
¬A
COMP2039ArtificialIntelligence:PhilosophicalAspectsofAI–p.16/19
Sorensen’s‘NecessityRefuter’
Agenerallogicalstructurefora(destructive)thought
experiment:
1.S:Modalsourcestatement
2.S⇒
I:Modalextractor
3.(I∧C)

W:Counterfactual
4.¬
W:Absurdity
5.
C:Contentpossibility
A

Bdenotesthesubjunctiveconditional
interpretedas‘ifAwerethecase,thenBwouldbethe
case’.
1..5aredeemedtobemutuallycontradictory.
Sooneofthem(atleast)mustbefalse,butisitS?
COMP2039ArtificialIntelligence:PhilosophicalAspectsofAI–p.17/19
AppliedtotheCRA...
Statements1to5areinterpretedasfollows:
1.S:Themodalsourcestatementisthe‘theory’ofstrongAI,namely
thatexecutinganAIprogramisnecessarilyconstitutiveof
understanding.
2.S⇒
I:ThemodalsourcestatementSimpliesthelogical
necessitythatanyimplementationoftheChinese-understanding
programunderstandsChinese,inaccordancewiththetenetsof
strongAI.
3.(I∧C)

W:IfSearleweretohand-implementanysuchprogram
thenhewouldunderstandChinese–a‘weird’consequence,W.
4.¬
W:SearledoesnotunderstandChinese(thisisan‘absurdity’).
5.
C:ItispossiblethatSearlecanhand-implementtheChinese
understandingprogram.
COMP2039ArtificialIntelligence:PhilosophicalAspectsofAI–p.18/19
SoWhat?
Thisstructuregivesusasystematicwaytoclassify
objectionstotheCRA.
Forexample,thesystemsreplyisanattackonthe
counterfactual3...
ItisnotSearlewhowouldunderstandChinese,butthe
systemofwhichheisbutapart.
Alternatively(equivalently?),theattackisonthe
translationfromeverydaylanguagetoformallogical
form.
COMP2039ArtificialIntelligence:PhilosophicalAspectsofAI–p.19/19