From: AAAI Technical Report FS-94-05. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved. Introductory AI for Whom? Presenting AI to the Non-Scientist Rebecca E. Skinner, Visiting Scholar, KnowledgeSystems Laboratory, Computer Science Department, Stanford University 701 Welch Road, Building C, Palo Alto, CA94304. 415 725 3855. fax 415 725 5850. skinner@hpp.stanford.edu Abstract ArtificialIntelligence, as a set of toolsfordescribing, analyzing andcreatingartifacts, has muchto offer the humanitiesand sciences.Thesekleinmaybe appliedto far moreinformal human artifacts suchas culture,as a means to better understand the non-engineered complex artifacts whichsurroundus. This paper describes "AI for Everyone",a coursefor advanced undergraduates whichis intendedto introducevariousideas in knowledge representation madAIto the non-scientisL Theauthor is currentlyproposing that this coursebe offeredat Stanford University during 1994-1995.Thecourse introduces the conceptsof (a) designof complex artifacts and(b) internal representationderivedfromcognitivescience,as seenin the progress of AI research. The class focuses on symbolic representation techniques, withtwosessionsonreiterationsof other kindsof intelligence(Nouvelle AI, Connectionism). Also included aresessionsoncritiquesof symbolic representation and on potential meansto apply KRconceptsto fields outside computing. Notoclmicalbackground in computing or cognitive s~enc~ is assumed. InWoduction Theproblemas posed assumesthat Introductory AI as a set of programmingtechniques should be taught to computerscientists, andonly to computerscientists. The author questions this premise. She suggeststhat instead certain of the analytical approachesdevelopedby 33 be introducedto social science andhumanitiesstudents. The sciences of the complex,of which33 is the pre-eminent representativeaddressingcognitivecomplexity,shouldbe taught to far more people. A well-defined subset of existing KRconceptsshouldbe incorporatedinto the base of knowledge whichdefines a civilized, educatedperson. The state of knowledgeconcerningthe sciences of the complexin the educated public and in the fields of humanities and social sciences is very poor (I). AI, cognitivescience, and their commercial applications have beenpresentedto the public, includingto highlyeducated non-technicalprofessionals, by the Dreyfusbrothers and businessjournalistsrespectively.Thelatter present33as a simplistic, inhumane,and failed attempt to replace human intelligence with that of machines ("the Frankensteinmodel").Thefield’s leaders are cognizantof this unsavorypublic imageof 33 as either "impossibleor invisible" (see reports on AAAIBoardmeetingsand the AA33’s1993round of ballots for Councilors). However, not muchis beingdoneaboutthe situation. Tothose in the thickof 33research,this institutionalintellectualpassivity maynot be either obviousor importanLWhyshould the 38 non-specialist knowanything about this field ? AI softwareimplementations are typically highlytransparent, so there is relatively little needfor the averagePCuser to understand AI. Accordingto this interpretation, the public’s ignoranceabout 33 is irritating but innocuous. Theauthorasserts that this is a gravemisreading of the situation. ActuallyAI has muchto offer the humanities andsocial sciences. The field concentrates on the developmentof useful software (and hardware) tools based on two distinct exemplarsof intelligence: "33 as empirical inquiry", entailing the reiteration of logical cognizantintellectual processes, and the recreation of intelligence fromthe bottomup (epitomizedby Brooks’as "elephantsdon’t play chess"). However,both ’AI as complexity’ and ’AI as simplicity’ orientations overlook the very general definition of the term "tool": "an implementor machine usedto do workor performa task". Ideally the pragmatic softwaretool orientation of the current AI worldcouldbe complementedby a wider understanding of very basic conceptsof 33, in a non-technicalform. In a sense such understanding is part of oneof the basic imperativesof 33 as empiricalinquiryandtool creation, or the creationof a Knowledge Medium(Stefik, 1986). Whatis a Knowledge Medium if it is not a meansfor conveyingthought ? The latter is of course, doneby softwareandmaterial tools, but it is also donebythinking,andis part of the collective store of wealthavailableto any givenculture. Thesimple potential usageof knowledge representationas a meansto analyzingproblemsin a moreor less formalwayoffers a meansto help peoplethink moreclearly. Thedescriptive andanalytical potencyof the basic conceptsand issues of AI (in symbolic and connectionist forms alike) hold regardless of whoseNCAIdemois more impressive. In this light, the exclusivefocuson softwaretools amidstthe field’s wealth of ideas about knowledge,intelligence, information,social interactions(collaboration,interaction) is curious rather than natural. Theintroversion Of 33 amongacademic disciplines seems incongruent and inappropriate. RogerSchankhas defined AI as an algorithmic approach to various domains: " AI should in principle be a contribution to a great manyfields of study... 33 is, potentially, the algorithmicstudy of processesin every field of inquiry. As such, the future should produce 33/anthropologists, AI/doctors, 33/political scientists andso on. It is just a matterof timeuntil 33becomes part of other fields..." (Schank1991,p7). Onthe contrary, such individuals havenot flooded AAAI on their own.It seemsmostunlikelythat they will unless the field actively undertakesto acquire a wideraudience in non-CSuniversity circles. Applyingexisting concepts to newcontexts is essential to intellectual progress (consider the importanceof analogyin KRandlearning), but is sufficiently strenuousthat peoplein AI mustwork to develop analogies betweenAI and semantically and syntactically distant fields. Shouldsuch an effort be undertaken,AI could makea significant contribution to society as an intellectual infrastructure or "publicgood" for problemanalysis and problem-solving. In a more abstract sensethan the conceptis usuallyunderstood,the rudimentsof knowledge representation could be generally introducedin highereducation.Theproliferation of AI as intellectual edifice could and should complementthe proliferationof AI as softwaretools. Onecontribution to this effort could be to introduce courses entitled "AI for Everyone". A version of the prospective course syllabus follows. The author is currently trying to arrange to teach this course in a humanities department at Stanford University. Many peopleoutsideAI are verycuriousaboutthe field, but see it as entirely inaccessible. Yet manybasic ideas in AI, such as internal representation, inferencing, heuristic vetsas algorithmicsolutions, searchstrategies, flamesand scripts, schemata, etc., maybe explained without referenceto FOPC.Boththe student andthe teacher could benefit fromsuch a course. Figuring out howto explain the basic concepts of the sciences of the complexin general and AI in specific without the assumptionsthat the student knowsFOPCand programminglanguages would be a fine excercise for AI graduate students. Perhapsteaching such a course could help to remindAI practitioners, most of whose work is effectively engineering,that they are (ostensibly)AI intellectuals well. Thefragmentationof AI as a field into discrete subdisciplines whichmakevery different assumptions aboutthe nature of ~intelligence"or ignorethe latter Big Question entirely is another salient problem. The challenge of presenting the field in simple natural language(that is, English) to a non-technicalaudience wouldbe a useful ’thoughtexperiment’as well. It might evenyield AI graduatestudents. Anhistorical approach,whichintroducesthe development of symbolic representation technology (intellectual technology,that is) in the formof the Physical Symbol Systemin mid-century,and follows its elaboration over the past several decades, could be a reasonableteaching format. This shouldincorporatea judicious allowancefor non-symbolic information processing models. Such a coursecouldintroducethe field in simplecanonicalform, commencing with a presentation of the concept of the sciences of the artificial as stated by Simonin The Sciencesof the Artificial. This shouldbe followedby a basic imroductionto knowledgerepresentation concepts, such as declarative versus procedural decompositionof problems, state space representations and operators, heuristic versusalgorithmicsolutions, framesandobjects, 39 and various forms of search. Thehistorical development fromsyntactic search proceduresto heuristic search and complementaryknowledgerepresentations should be illustrated by reference to prominent programs. The syllabus currently being developedalso includes lessons on computerlanguages,counectiouist and "Nouvelle"AI challengesto the ’Classic’ PSStradition, andthe novel developments of the recent years, including VLKB, generic task representations, and meta-tools of the knowledge sharing variety. Evena simplepresentation of canonical approachesto symbolicrepresentation and to non-symbolicinformation processing could provokea greater degreeof introspection in bright peoplewhohave beendeprivedof suchtools. This wouldalso help to place AI in the context of the sciences relating to complex artifacts. This developmenthas been extraordinarily fruitful, yet mostnon-scientistsonly see the artifacts of computingin the opaque form of PCs and operating software. Computingin general and AI in specific go much deeper than word-processing packages, spreadsheets, and computergraphics, but the outside world won’t knowit unless people from AI present this knowledge to a wideraudience. Thecreation of a syllabusfor Watching an Inl~oductionto AI couldhelp to illuminateArsundeserved invisibility, as well as to introducethe field’s analytical wealththrough culture, rather than solely through software tools. If technologyis culture, then AI shouldbe high culture. Notes (1) This is not an exaggeration.Thewriter beganstudying commercial AI and KBSunder the auspices of the economicsof technological change. Whentrying to explainthe essentialsof AI andreasonfor her interest, she has been met with commentssuch as: "Whois Noam Chomsky?"; "I read in the NewRepublic that AI has failed completely"; and, most commonly:"Hubert Dreyfussays..." by distinguishedtenured professors in various social science fields at a major research university. Bibliography D. Bobrow edited volumeof Artificial Intelligence. 1991. Haugeland,John. Artificial Intelligence: TheVe~Idea. 1982. Kuhn,Thomas."A Function for ThoughtExperiments". The Essential Tension: Selected Studies in Scientific Changeand Tradition. 1977. Schank" Whatis AI anyway?". Partridge and Wilks, eds., The Foundations of Artificial Intelligence: A Sourcebook. NewYork: CambridgeUniversity Press. 1991:3-13. Stef’ik, Mark." The KnowledgeMedium".IALMagaz_i~ 1986. ProposedSyllabus AI for Everyone:AnIntroductionto the Sciences of the Complexfor Humanities and Social Science Majors Brief Description: The one-semestercourse for juniors and seniors will providean historical overviewof Artificial Intelligence with a symbolicrepresentationslant, and the Sciencesof the Artificial to humanitiesandsocial sciencemajors. Requirements: Reading, class attendance twice per week, and two examinations. Thereare no prerequisites. Week1. Inlroduction This lecture defines the conceptof the sciences of the complex,artificiality and complexity.It then explains howa cluster of disciplines has emergedto examine artifacts, or engineered systemsof both physical and ephemeralforms (i.e., software). It suggests howthis study in turn provides novel and insightful meansby which to consider the humanphenomenatraditionally studiedin the humanitiesandsocial sciences. Thelecture suggests reasons why art history and comparative literature majorsshouldcare about this discipline. Such rationales call uponpragmatism (the increasing presence of computationin daily life in industrial nations); cynicism(technology is power);historical interest (this one of the century’s most significant intellectual developments); and philosophicalinterest (the methods the sciences of the complexmaybe applied outside the usual doma~). Termsto be introduced: artifact, sciences of the complex,technology, representation;artificial-natural, AI, cognitivescience, computer,hardware,software. Reading:(someitems over the courseof the semester):S andH. Dreyfus, WhatComputersStill C~’t Do; Haugeland, J Intelli~enoe:TheVeryIdea. JoJmson-Lmd, P. ~d~ Mind. Selectedchapters. Gardner, H. The Mind’sNewScience: A Historyof the Counitive ~. McCorduck, P.M~n~ ~ Think.Stiilings, N.S. et al. ~0enitiyeScience:AnIntroduction. Simon,H. TheSciencesof the ~. C2nd ecL).1980. Turing, A.M." ComputingMachineryand Intenigence". (1950). Week 2. Histori-------------afl Precedents Artificial Intelligence and cognitive science emerged from various methodologicaland substantive theoretical roots duringthe middledecadesof the Twentiethcentury. roots includeformallogic in philosophy;operations research; andcybernetics,or the studyof automata.These necessary underpinnings entailed (respectively) improvements in formal logic (Boole, Russell and Whitehead, Frege, Godel); improvements in the representation of informationderived from electronics (Shannon’sconcept of binary-based representation of switching); and inquiries into the nature of complex 40 intelligent systems(anthropomorphic or otherwise)under the auspicesof researchinto intelligent automata(Wiener, Von Neumann, and Newell Shaw and Simon, and Turing). Given the prevailing regime of WatsonianSkinnerian behaviorism, the agenda for exploring intelligencethroughcyberneticswaspractically subrosa. Termsto be introduced: the Cognitive Revolution, automata,Behaviorism,cybernetics, operationsresearch, informationrepresentation, internal representation, Von Neamannmachine. Reading:Gardenerchapters. Heims,S. TheCyberneticsGroup.1991. Regis, E. WhoGot Einstein’s Office ? Eccentricity_ and Geniusat the Institute for Advanced Study. Week3. The 1950s: Formationof the AI Research Programs The formation of agendas for studying intelligent information processing was slow and collaborative. Resea~hdirections whichnowseem natural were by no meansso before 1956. (Alan Newell:"Deepscientific ideas are exceedinglysimple. Others usually see themas trivial"). Directlyinspiredby the Cyberneticmodellingof information processing, several strands of cognitive science and AI coalesced during this period. These include the Chomskianlinguistic agenda, the most directly argumentativewith Behaviorism;Hebb’sprotoconnectionist biological metaphor for modelling informationprocessing;and the representationofthought throughcomputationalmanipulationof symbolicsystems. All three will be sketched, and the latter examinedin somedetail. Newelland Simon’sGeneralProblemSolver experimentsat Carnegie-Mellonusheredin the modeling of intellectual processes through weaksearch. This yielded various atomistic syntactic problem-solvingand search methods(the latter issue continued into next week). Termsto be introduced: human-computer interaction, protocol, biological metaphor, toy problem, neuron, intelligence, information processing, computational metaphor, symbolicrepresentation, weakversus strong methods,deepstructure. Reading: Scheerer, E."Toward a History of Cognitive Science". International Social Science Journal 40. Sections fromGardner, McCorduck, Stillings etal. Week4. The 1960s: Early Development of Search and Problem.Solving The PSS research program yielded both significant cognitive science research and powerfulKRtechniques. Someof these, especially the GPS-basedsemantically blind search methods,are practically takenfor grantedat the present, perhapsbecauseof their application to toy problems.The sessions will introduce the concepts of internal representation,intelligence as goal-seeking, and search as movementthrough a problem space using operators. Termsto be introduced: semantics-syntax, problem space, state space representation, top-downversus bottom-up, physical symbol system, search, pruning search spaces, graphs, declarative versus procedural informmion, heuristic, algorithm,operator, operand. Reading: TBA Reading:Gardener1985and Sailings et al., selections; Selections from Bobrow,Minsky, Newell and Simon, Schank, MYCIN Project. Week5a. The Tropismto Domainsin the Late 1960s The syntactic approachto blind search in toy spaces shifted pivotally to search of moreambitious domains (areas of expertise) duringthe late 1960s. This session will describe two of the strategies of this period which movedtoward systematic representation of difficult (rather than toy) problems, such as the modeling scientific reasoningin the Dendralproject at Stanford,and the introductionof productionsystemsat CMU in the late 1960s. Termsto be introduceA:domainrepresentation, control slzncture, generateand test; productionsystems,control structure; Reading:Chapter from Newell’s Unitary Theories of ~; other TBA. Week7. RepresentingOther Formsof Intelfigence The cognitive metaphoras presented in this course certainly cannot presumeto embracethe entirety of humanexperience and reaction to the environment.Much of the latter includes emotion,physical reaction, and instant reactionsto the environment. Thelatter shouldbe thoughtof as different sorts of intelligence. "Nouvelle AI" replicates simple, rapid reactions to the physical environmentusing robotic techniques. The biological metaphor, which tries to reiterate the rudimentary intelligence captured in the neural system of living creatures (not necessarily humans),is a useful if very distinct meansto understandintelligence. Still another approach,closer to the Classical symbolicgenre, is the examinationof intelligence in a modularperspective. There are various componentsin conscious cogitation, including memory, analogical reasoning, spatial reasoning,etc. Is it true, as MarvinMinskyis purportedto havesaid, that "Consciousnessis overrated"? If consciousnessis manythings, howcan each of these be translated into usefulartifacts ? Termsto be introduced:massiveparallelism ’wetware’, ’meat machines’, emergent states, perceptrons, Connectionism, robot, knowbot, neural network, modularityof mind,faculties theory; Reading:M. Minsky and S. Papert, Percevtrons. H. Moravec,MindChildren; D. Hillis article in Wired October 1993. Minsky, The Society of Mind. Daedalus 1988Special Issue on AI article on Connectionism. Week5b. ComputerLanguages A languageis a protocol for communication,between people, agents, or other systems. Computerlanguages descn~oeentities andactivities, just as natural languages do, but in a far moreformalandrigidly definedfashion. Natural languageand computerlanguageshoweverdiffer in formality,lexicai andreferential ambiguity,syntactic idiosyncrasies,andnuancesof semantics.This lecture will explain the need for formal languages to control engineeredartifacts, and the distinctions betweenlowlevel andhigher-levelcomputational languages. Terms to be introduced: computer versus natural language;syntax, semantics,lexical ambiguity,protocol; functional versus procedural language, object, encapsulation, global versus local variable; high-level language,platform, environment,application, operating system,binary, Boolean; Reading:(selected chapters): Brookshear,J.G. Science: AnOverview.Baron, N. ComputerLant, uages: AGuidefor the Perplexed.1986. Week 6. Changing the Scope of Knowledge Representationin the 1970S Despite the reach of GPSand similar programs,muchof its initial technical obstacles hadinvolvedvery minute manipulationof particles of thought. In the early 1970s, the technologyfor representing small-scale information manipulationhad advancedsufficiently to allow several large-scale conceptsof AI representationof intelligence. Theseinclude production systemsand rules (Newelland Simon, Buchananand Shortliffe), frames (Minsky), schemata(Bobrowand Collins), and scripts (Schank). Thesewaysof representingintelligence (usually basedon cognitive science research on people) werealso waysof creating artifacts that had certain forms of knowledge. Thus the concepts embodied both scientific and technological knowledge. Termsto be introduced:rules, objects, frames,domains, script, expert systems, knowledge base systems; conceptmlprimitive, blackboard; 41 Week8, a. Critiques of AI, KRand Technologyin General: Week 8, b. AI Views of Emotions, Music, and Creativity: a) Critiques of AI and cognitive science have been virulent, althoughmostpractitioners havepaid no notice. The argumentsinclude a) ad hominencritiques of the social skills of hackers (Turkle, Weizenbaum); assertions of a simplistically mechanistic worldview entailed in symbolicabstractions (Merchant);c) a more open phenomenological attack on the ’atomism’ingrained in the PSSmodel(Searle and Dreyfusbrothers). However, harsh machofunctionalismand ’atomism’are out of style in AI itself, suggestingfor a start that the critiques are dated. Giventhe prominenceof distributed and situated knowledgeand beterogenous representation amongthe ’newsensitive AI researchers’of both the NouvelleAI and the KRcommunities, are these critiques still valid ? b) Recent work in AI instead suggests the potential contribution of KR to art, entertainment, and developmental andclinical psychology. Terms to be introduced: Luddites; phenomenology versusfunctionalism;intentionality, fringe consciousness; Reading(a): Selections: D. Harraway;C. Merchant,The l~,~k.o.f~]alll~;S. Turkle,Ihf,.i,~f.011d.~g~; J. Searle; J. Weizenbaum,ComputerPower and HumanReason. Reading(b): Articles on AI and Music, ~ 1991; work by J. Bates; D. Gele~nter. ~e Muse ia the Machine:Computerizing the Poetry of Human Tilought. 1994. Schank,R." Creativityas a MechanicalProcess". In R. Sternberg, ed. Tbe N~of .C, IgiliX~.1988:220-238. Week9. Current Workin KR: a. CurrentWorkin Artificial Intelligence: Criticism of the first generationof KBsystemson the groundsof their "orittle andisolated’ quality havebeen taken very seriously in the KRcommunity.Current research in this communityemphasizesthe creation of shared knowledgerepresentations. Ongoingresearch in this effort focuseson the technical andrepresentational challengesof buildingintermediate"interlingua", a kind of Esperanto between disparate existing computer languages and knowledge representations in those languages.This research challenge,intensified by the end of the Cold War, forces the PSS/KRcommunityto providepractical softwarein a shorter time framethan it is used to. The concept of an "intelligent agent" (undefined)is replacingthe knowledge base systemas the predominantartifact and vehicle for AI platforms. These are less directly anthropomorphic and possibly complementary to more numerous if less lofty applications than KBSseemsto havebeen. Termsto be introduced: CYC;very large knowledge base; device modeling; agents; knowledgesharing; KnowledgeMedium; b. In-class field trip to aa AALab, TBA:HowThings Workproject at the Heuristic ProgrammingProject 7 Tools for MedicalInformatics KRat KSL-CAMIS .9 Reading TBA. Week10. Potential Contributions of Mand Cognitive Scienceto other fields; The sciences of the complex consist of ways to understand humanintelligence and meansby whichto create newartifacts. This has not beenusedto analyzethe artifacts- physicalor intellectual- of the past. Apartfrom their intrinsic interest, these conceptscouldbe usedin the humanities and social sciences in various ways. The descriptive terminology concerning various forms of internal representationsand cognitiveemulationcouldbe used for the discipline of history. This wouldpertain specificallyto intellectualhistory(of givencultures,eras, etc.), as well as to the discrete studies of women and different ethnic and economicgroups which are now prevalent. Thecontributionsof these fields to economics is cmTentlybeing explored,andcognitiveanthropologyis an established field of study. The breadth of domain, extensibility of existing concepts,andcharacteristics of search or problem-solving which demarcate a given cultural artifact may be described by KR’s interdisciplinav/terminology in a waynot nowpossible. Week11. Final Examination Acknowledgements:Thanks to Lee Brownston of the KnowledgeSystems Lab at Stanford for insightful comments.