如何做科研.doc

MIT科学家的论道:如何做科研ztMASSACHUSETTSINSTITUTEOFTECHNOLOGYARTIFICIALINTELLIGENCELABORATORYAIWorkingPaper316October,1988HowtodoResearchAttheMITAILabby:awholebunchofcurrent,former,andhonoraryMITAILabgraduatestudentsDavidChapman,EditorSeptember,1988.Abstract:Thisdocumentpresumptuouslypurportstoexplainhowtodoresearch.Wegiveheuristicsthatmaybeusefulinpickingupthespecificskillsneededforresearch(reading,writing,programming)andforunderstandingandenjoyingtheprocessitself(methodology,topicandadvisorselection,andemotionalfactors).Copyright1987,1988bytheauthors.A.I.LaboratoryWorkingPapersareproducedforinternalcirculation,andmaycontaininformationthatis,forexample,toopreliminaryortoodetailedforformalpublication.Itisnotintendedthattheyshouldbeconsideredpaperstowhichreferencecanbemadeintheliterature. ReadingAIManyresearchersspendmorethanhalftheirtimereading.Youcanlearnalotmorequicklyfromotherpeoplesworkthanfromdoingyourown.ThissectiontalksaboutreadingwithinAI;sectioncoversreadingaboutothersubjects.Thetimetostartreadingisnow.Onceyoustartseriouslyworkingonyourthesisyoullhavelesstime,andyourreadingwillhavetobemorefocusedonthetopicarea.Duringyourfirsttwoyears,youllmostlybedoingclassworkandgettinguptospeedonAIingeneral.Forthisitsufficestoreadtextbooksandpublishedjournalarticles.(Later,youmayreadmostlydrafts;seesection.)Theamountofstuffyouneedtohavereadtohaveasolidgroundinginthefieldmayseemintimidating,butsinceAIisstillasmallfield,youcaninacoupleyearsreadasubstantialfractionofthesignificantpapersthathavebeenpublished.Whatsalittletrickyisfiguringoutwhichonesthoseare.Therearesomebibliographiesthatareuseful:forexample,thesyllabiofthegraduateAIcourses.ThereadinglistsfortheAIqualifyingexamsatotheruniversities-particularlyStanford-arealsouseful,andgiveyoualessparochialoutlook.Ifyouareinterestedinaspecificsubfield,gotoaseniorgradstudentinthatsubfieldandaskhimwhatarethetenmostimportantpapersandseeifhelllendyoucopiestoXerox.Recentlytherehavebeenappearingalotofgoodeditedcollectionsofpapersfromasubfield,publishedparticularlybyMorgan-Kauffman.TheAIlabhasthreeinternalpublicationseries,theWorkingPapers,Memos,andTechnicalReports,inincreasingorderofformality.Theyareavailableonracksintheeighthfloorplayroom.Gobackthroughthelastcoupleyearsofthemandsnagcopiesofanythatlookremotelyinteresting.Besidesthefactthatalotofthemaresignificantpapers,itspoliticallyveryimportanttobecurrentonwhatpeopleinyourlabaredoing.TheresawholebunchofjournalsaboutAI,andyoucouldspendallyourtimereadingthem.Fortunately,onlyafewareworthlookingat.Theprincipaljournalforcentral-systemsstuffisArtificialIntelligence,alsoreferredtoastheJournalofArtificialIntelligence,orAIJ.MostofthereallyimportantpapersinAIeventuallymakeitintoAIJ,soitsworthscanningthroughbackissueseveryyearorso;butalotofwhatitprintsisreallyboring.ComputationalIntelligenceisanewcompetitorthatsworthcheckingout.CognitiveSciencealsoprintsafairnumberofsignificantAIpapers.MachineLearningisthemainsourceonwhatitsays.IEEEPAMIisprobablythebestestablishedvisionjournal;twoorthreeinterestingpapersperissue.TheInternationalJournalofComputerVision(IJCV)isnewandsofarhasbeeninteresting.PapersinRoboticsResearcharemostlyondynamics;sometimesitalsohasalandmarkAIishroboticspaper.IEEERoboticsandAutomationhasoccasionalgoodpapers.Itsworthgoingtoyourcomputersciencelibrary(MITsisonthefirstfloorofTechSquare)everyyearorsoandflippingthroughthelastyearsworthofAItechnicalreportsfromotheruniversitiesandreadingtheonesthatlookinteresting.Readingpapersisaskillthattakespractice.Youcantaffordtoreadinfullallthepapersthatcometoyou.Therearethreephasestoreadingone.Thefirstistoseeiftheresanythingofinterestinitatall.AIpapershaveabstracts,whicharesupposedtotellyouwhatsinthem,butfrequentlydont;soyouhavetojumpabout,readingabithereorthere,tofindoutwhattheauthorsactuallydid.Thetableofcontents,conclusionsection,andintroductionaregoodplacestolook.Ifallelsefails,youmayhavetoactuallyflipthroughthewholething.Onceyouvefiguredoutwhatingeneralthepaperisaboutandwhattheclaimedcontributionis,youcandecidewhetherornottogoontothesecondphase,whichistofindthepartofthepaperthathasthegoodstuff.Mostfifteenpagepaperscouldprofitablyberewrittenasone-pagepapers;youneedtolookforthepagethathastheexcitingstuff.Oftenthisishiddensomewhereunlikely.Whattheauthorfindsinterestingabouthisworkmaynotbeinterestingtoyou,andviceversa.Finally,youmaygobackandreadthewholepaperthroughifitseemsworthwhile.Readwithaquestioninmind.HowcanIusethis?Doesthisreallydowhattheauthorclaims?Whatif.?Understandingwhatresulthasbeenpresentedisnotthesameasunderstandingthepaper.Mostoftheunderstandingisinfiguringoutthemotivations,thechoicestheauthorsmade(manyofthemimplicit),whethertheassumptionsandformalizationsarerealistic,whatdirectionstheworksuggests,theproblemslyingjustoverthehorizon,thepatternsofdifficultythatkeepcomingupintheauthorsresearchprogram,thepoliticalpointsthepapermaybeaimedat,andsoforth.Itsagoodideatotieyourreadingandprogrammingtogether.Ifyouareinterestedinanareaandreadafewpapersaboutit,tryimplementingtoyversionsoftheprogramsbeingdescribed.Thisgivesyouamoreconcreteunderstanding.MostAIlabsaresadlyinbredandinsular;peopleoftenmostlyreadandciteworkdoneonlyattheirownschool.Otherinstitutionshavedifferentwaysofthinkingaboutproblems,anditisworthreading,takingseriously,andreferencingtheirwork,evenifyouthinkyouknowwhatswrongwiththem.Oftensomeonewillhandyouabookorpaperandexclaimthatyoushouldreaditbecauseits(a)themostbrilliantthingeverwrittenand/or(b)preciselyapplicabletoyourownresearch.Usuallywhenyouactuallyreadit,youwillfinditnotparticularlybrilliantandonlyvaguelyapplicable.Thiscanbeperplexing.Istheresomethingwrongwithme?AmImissingsomething?Thetruth,mostoften,isthatreadingthebookorpaperinquestionhas,moreorlessbychance,madeyourfriendthinksomethingusefulaboutyourresearchtopicbycatalyzingalineofthoughtthatwasalreadyformingintheirhead.GettingconnectedAfterthefirstyearortwo,youllhavesomeideaofwhatsubfieldyouaregoingtobeworkingin.Atthispoint-orevenearlier-itsimportanttogetpluggedintotheSecretPaperPassingNetwork.ThisinformalorganizationiswherealltheactioninAIreallyis.Trend-settingworkeventuallyturnsintopublishedpapers-butnotuntilatleastayearafterthecoolpeopleknowallaboutit.Whichmeansthatthecoolpeoplehaveayearsheadstartonworkingwithnewideas.Howdothecoolpeoplefindoutaboutanewidea?Maybetheyhearaboutitataconference;butmuchmorelikely,theygotitthroughtheSecretPaperPassingNetwork.Hereshowitworks.JoCoolgetsagoodidea.Shethrowstogetherahalf-assedimplementationanditsortofworks,soshewritesadraftpaperaboutit.Shewantstoknowwhethertheideaisanygood,soshesendscopiestotenfriendsandasksthemforcommentsonit.Theythinkitscool,soaswellastellingJowhatswrongwithit,theylendcopiestotheirfriendstoXerox.Theirfriendslendcopiestotheirfriends,andsoon.JorevisesitabunchafewmonthslaterandsendsittoAAAI.Sixmonthslater,itfirstappearsinprintinacut-downfive-pageversion(allthattheAAAIproceedingsallow).Joeventuallygetsaroundtocleaninguptheprogramandwritesalongerrevisedversion(basedonthefeedbackontheAAAIversion)andsendsittotheAIJournal.AIJhasalmosttwoyearsturn-aroundtime,whatwithreviewsandrevisionsandpublicationdelay,soJosideafinallyappearsinajournalformthreeyearsaftershehadit-andalmostthatlongafterthecoolpeoplefirstfoundoutaboutit.Socoolpeoplehardlyeverlearnabouttheirsubfieldfrompublishedjournalarticles;thosecomeouttoolate.You,too,canbeacoolpeople.Herearesomeheuristicsforgettingconnected:TheresabunchofelectronicmailingliststhatdiscussAIsubfieldslikeconnectionismorvision.Getyourselfontheonesthatseeminteresting.Wheneveryoutalkaboutanideayouvehadwithsomeonewhoknowsthefield,theyarelikelynottogiveanevaluationofyouridea,buttosay,HaveyoureadX?Notatestquestion,butasuggestionaboutsomethingtoreadthatwillprobablyberelevant.IfyouhaventreadX,getthefullreferencefromyourinterlocutor,orbetteryet,asktoborrowandXeroxhiscopy.Whenyoureadapaperthatexcitesyou,makefivecopiesandgivethemtopeopleyouthinkwillbeinterestedinit.Theyllprobablyreturnthefavor.Thelabhasanumberofon-goinginformalpaperdiscussiongroupsonvarioussubfields.Thesemeeteveryweekortwotodiscussapaperthateveryonehasread.Somepeopledontmindifyoureadtheirdesks.Thatis,readthepapersthattheyintendtoreadsoonareheapedthereandturnoverprettyregularly.Youcanlookoverthemandseeiftheresanythingthatlooksinteresting.Besuretoaskbeforedoingthis;somepeopledomind.Trypeoplewhoseemfriendlyandconnected.Similarly,somepeopledontmindyourbrowsingtheirfilingcabinets.Therearepeopleinthelabwhoareintoscholarshipandwhosecabinetsarequitecomprehensive.Thisisoftenafasterandmorereliablewaytofindpapersthanusingtheschoollibrary.Wheneveryouwritesomethingyourself,distributecopiesofadraftofittopeoplewhoarelikelytobeinterested.(Thishasapotentialproblem:plagiarismisrareinAI,butitdoeshappen.YoucanputsomethinglikePleasedonotphotocopyorquoteonthefrontpageasapartialprophylactic.)Mostpeopledontreadmostofthepaperstheyregiven,sodonttakeitpersonallywhenonlyafewofthecopiesyoudistributecomebackwithcommentsonthem.Ifyougothroughseveraldrafts-whichforajournalarticleyoushould-fewreaderswillreadmorethanoneofthem.Youradvisorisexpectedtobeanexception.Whenyoufinishapaper,sendcopiestoeveryoneyouthinkmightbeinterested.Dontassumetheyllreaditinthejournalorproceedingsspontaneously.Internalpublicationseries(memosandtechnicalreports)areevenlesslikelytoberead.Themoredifferentpeopleyoucangetconnectedwith,thebetter.Trytoswappaperswithpeoplefromdifferentresearchgroups,differentAIlabs,differentacademicfields.Makeyourselfthebridgebetweentwogroupsofinterestingpeopleworkingonrelatedproblemswhoarenttalkingtoeachotherandsuddenlyreamsofinterestingpaperswillflowacrossyourdesk.Whenapapercitessomethingthatlooksinteresting,makeanoteofit.Keepalogofinterestingreferences.Gotothelibraryeveryonceinawhileandlookthelotofthemup.Youcanintensivelyworkbackwardthroughareferencegraphofcitationswhenyouarehotonthetrailofaninterestingtopic.Areferencegraphisawebofcitations:paperAcitespapersBandC,BcitesCandD,CcitesD,andsoon.Papersthatyounoticecitedfrequentlyarealwaysworthreading.Referencegraphshaveweirdproperties.Oneisthatoftentherearetwogroupsofpeopleworkingonthesametopicwhodontknowabouteachother.Youmayfindyourselfclosetoclosureonsearchingagraphandsuddenlyfindyourwayintoanotherwholesection.Thishappenswhentherearedifferentschoolsorapproaches.Itsveryvaluabletounderstandasmanyapproachesaspossible-oftenmoresothanunderstandingoneapproachingreaterdepth.Hangout.Talktopeople.Tellthemwhatyoureuptoandaskwhattheyredoing.(Ifyoureshyabouttalkingtootherstudentsaboutyourideas,saybecauseyoufeelyouhaventgotany,thentrytalkingtothemaboutthereallygood-orunbelievablyfoolish-stuffyouvebeenreading.Thisleadsnaturallyintothetopicofwhatonemightdonext.)Theresaninformallunchgroupthatmeetsintheseventhfloorplayroomaroundnooneveryday.Peopletendtoworknightsinourlab,andsogofordinnerinloosegroups.Inviteyourselfalong.Ifyouinteractwithoutsidersmuch-givingdemosorgoingtoconferences-getabusinesscard.Makeiteasytorememberyourname.Atsomepointyoullstartgoingtoscientificconferences.Whenyoudo,youwilldiscoverfactthatalmostallthepaperspresentedatanyconferenceareboringorsilly.(Thereareinterestingreasonsforthisthatarentrelevanthere.)Whygotothemthen?Tomeetpeopleintheworldoutsideyourlab.Outsidepeoplecanspreadthenewsaboutyourwork,inviteyoutogivetalks,tellyouabouttheatmosphereandpersonalitiesatasite,introduceyoutopeople,helpyoufindasummerjob,andsoforth.Howtomeetpeople?Walkuptosomeonewhosepaperyouveliked,sayIreallylikedyourpaper,andaskaquestion.Getsummerjobsawayatotherlabs.Thisgivesyouawholenewpoolofpeopletogetconnectedwithwhoprobablyhaveadifferentwayoflookingatthings.Onegoodwaytogetsummerjobsatotherlabsistoaskseniorgradstudentshow.Theyrelikelytohavebeenplacesthatyoudwanttogoandcanprobablyhelpyoumaketherightconnections. LearningotherfieldsItusedtobethecasethatyoucoulddoAIwithoutknowinganythingexceptAI,andsomepeoplestillseemtodothat.Butincreasingly,goodresearchrequiresthatyouknowalotaboutseveralrelatedfields.Computationalfeasibilitybyitselfdoesntprovideenoughconstraintonwhatintelligenceisabout.Otherrelatedfieldsgiveotherformsofconstraint,forexampleexperimentaldata,whichyoucangetfrompsychology.Moreimportantly,otherfieldsgiveyounewtoolsforthinkingandnewwaysoflookingatwhatintelligenceisabout.AnotherreasonforlearningotherfieldsisthatAIdoesnothaveitsownstandardsofresearchexcellence,buthasborrowedfromotherfie