《信息科學(xué)類專業(yè)英語》課件第19章

Lesson19SmartRooms

(第十九課智能房間)

Vocabulary(詞匯)ImportantSentences(重點(diǎn)句)QuestionsandAnswers(問答)Problems(問題)

Increatingcomputersystemsthatcanidentifypeopleandinterprettheiractions,researchershavecomeonestepclosertobuildinghelpfulhomeandworkenvironments

Imagineahousethatalwaysknowswhereyourkidsareandtellsyouiftheyaregettingintotrouble.Oranofficethatseeswhenyouarehavinganimportantmeetingandshieldsyoufrominterruptions.Oracarthatsenseswhenyouaretiredandwarnsyoutopullover.Scientistshavelongtriedtodesigncomputersystemsthatcouldaccomplishsuchfeats.Despitetheirefforts,modernmachinesarestillnomatchforbaby-sittersorsecretaries.Buttheycouldbe.

Theproblem,inmyopinion,isthatourcurrentcomputersarebothdeafandblind:theyexperiencetheworldonlybywayofakeyboardandamouse.Evenmultimediamachines,thosethathandleaudiovisualsignalsaswellastext,simplytransportstringsofdata.Theydonotunderstandthemeaningbehindthecharacters,soundsandpicturestheyconvey.Ibelievecomputersmustbeabletoseeandhearwhatwedobeforetheycanprovetrulyhelpful.Whatismore,theymustbeabletorecognizewhoweareand,asmuchasanotherpersonorevenadogwould,makesenseofwhatwearethinking.

Tothatend,mygroupattheMediaLaboratoryattheMassachusettsInstituteofTechnologyhasrecentlydevelopedafamilyofcomputersystemsforrecognizingfaces,expressionsandgestures.Thetechnologyhasenabledustobuildenvironmentsthatbehavesomewhatlikethehouse,officeandcardescribedabove.Theseareas,whichwecallsmartrooms,arefurnishedwithcamerasandmicrophonesthatrelaytheirrecordingstoanearbynetworkofcomputers.Thecomputersassesswhatpeopleinthesmartroomaresayinganddoing.Thankstothisconnection,visitorscanusetheiractions,voicesandexpressions-insteadofkeyboards,sensorsorgoggles-tocontrolcomputerprograms,browsemultimediainformationorventureintorealmsofvirtualreality.[1]Thekeyideaisthatbecausethesmartroomknowssomethingaboutthepeopleinit,itcanreactintelligentlytothem.WorkingtogetherwithPattieMaesandme,graduatestudentsTrevorDarrellandBruceM.Blumbergconstructedthefirstsmartroomin1991atM.I.T.Theinitiativequicklygrewintoacollaborativeexperimentandnowinvolvesfivesuchrooms,alllinkedbytelephonelines,aroundtheworld:threeinBoston,oneinJapanandoneintheU.K.(InstallationsarealsoplannedforParis,NewYorkCityandDallas.)

Eachroomcontainsseveralmachines,nonemorepowerfulthanapersonalcomputer.Theseunitstackledifferentproblems.Forinstance,ifasmartroommustanalyzeimages,soundsandgestures,weequipitwiththreecomputers,oneforeachtypeofinterpretation.Ifgreatercapabilitiesareneeded,weaddmoremachines.Althoughthemodulestakeondifferenttasks,theyallrelyonthesamestatisticalmethod,knownasmaximumlikelihoodanalysis:thecomputerscompareincominginformationwithmodelstheyhavestoredinmemory.[2]Theycalculatethechancethateachstoredmodeldescribestheobservedinputandultimatelypicktheclosestmatch.Bymakingsuchcomparisons,oursmart-roommachinescananswerarangeofquestionsabouttheirusers,includingwhotheyareandsometimesevenwhattheywant.1Where?

Beforeasmartroomcanbegintofigureoutwhatpeoplearedoing,itneedstolocatethem.SograduatestudentsChristopherR.Wren,AliAzarbayejaniandDarrellandIdevelopedasystemcalledPersonFinder(Pfinderforshort)thatcantrackonepersonasheorshemovesaroundintheroom.Asdoourothersystems,Pfinderadoptsthemaximumlikelihoodapproach.First,itmodelsthepersonthecamerarecordsasaconnectedsetofblobs-twoforthehands,twoforthefeetandoneeachforthehead,shirtandpants.Itdescribeseachblobintwoways:asadistributionofvaluesfortheblob’scolorandplacement,andasaso-calledsupportmap,essentiallyalistindicatingwhichimagepixelsbelongtotheblob(pixelsare“pictureelements,”similartothedotsthatmakeupatelevisionimage).[3]Next,Pfindercreatestexturedsurfacestomodelthebackgroundscene.Eachpointononeofthesesurfacescorrelatestoanaveragecolorvalueandadistributionaroundthatmean.Wheneverthecamerainthesmartroompicksupanewpictureinthevideostream,Pfindercomparesthatimagewiththemodelsithasmadeandwithotherreferencesaswell.Tostart,thesystemguesseswhattheblobmodelshouldlooklikeinthenewimage.If,forexample,aperson’supperbodywasmovingtotherightatonemeterpersecondatenthofasecondago,thenPfinderwillassumethatthecenteroftheupperbodyblobhasmovedatenthofametertotheright.Suchestimatesarealsocheckedagainsttypicalpatternsofmovementthatwehavederivedfromtestingthesystemonthousandsofpeople.Forinstance,weknowthatblobscorrespondingtothetorsomustmoveslowly,whereasthoserelatingtohandsandfeetgenerallymovemuchfaster.

Predictionsfinished,Pfinderthenmeasuresthechancethateachpixelinthenewimagebelongstoeachblob.Itdoessobysubtractingthepixel’scolorandbrightnessvaluesfromeachblob’smeancolorandbrightnessvalues.Itcomparestheresultwitheachblob’sdistributiontodeterminehowlikelyitisthatthedifferencehappenedbychance.If,forexample,thebrightnessdifferencebetweenapixelandablobwere10percent,andtheblob’sstatisticssaidthatsuchadifferencehappenedonly1percentofthetime,thechancethatthepixelbelongedtotheblobwouldbeamereonein100.[4]Shadowspresentaminorprobleminthattheycausebrightnessdifferencesthathavenothingtodowiththeprobabilitythatsomepixelbelongstosomeblob.SoPfindersearchesoutshadows,areasthataredarkerthanexpected,andevensouttheircolorhueandsaturationusingthearea’soverallbrightness.

Pfindermustalsoovercomeslightchangesinthelightingorarrangementofobjectsintheroom,eitherofwhichmightmakeitplacecertainpixelsinthewrongmodels.Toavoidthisdifficulty,thesystemcontinuouslyupdatesthepixelsthatarevisiblebehindtheuser,averagingtheoldcolorinformationwiththenew.Inthisway,itkeepstrackofchangesthatoccur,forinstance,whentheusermovesabookandthusaltersthesceneintwoplaces:wherethebookwasandwhereitnowis.Aftercompletingthesevariouscalculationsandcompensations,Pfinderatlastassignseachpixelinthenewimagetothemodelthatmostlikelycontainsit.Finally,itupdatesthestatisticsdescribingtheblobmodelandthebackgroundscene,aswellasthoseanticipatingwhichwaytheblobswillmove.2WhoandHow?

Asidefromknowingwherepeopleare,asmartroommustalsoknowwhotheyareandwhattheyaresaying.Manyworkershaveinventedalgorithmsthatallowcomputerstounderstandspeech.Virtuallyallthosesystemsworkwellonlywhentheuserwearsamicrophoneorsitsnearone.Aroomthatinterpretedyouractionsonlywhenyoustoodinaparticularspotwouldnotseemsosmart.SograduatestudentsSumitBasuandMichaelCaseyandIlookedforanothersolution—onethatwouldletacomputerdecodeauser’sspeechasheorshemovedfreelyaboutsomeroom,eveniftheroomwerequitenoisy.OurendproducttakesadvantageofthefactthatPfinderfollowstheuser’spositionatalltimes.Borrowingthisinformation,thespeech-recognitionsystemelectronically“steers”anarrayoffixedmicrophonessothattheyreinforceonlythosesoundscomingfromthedirectionoftheuser’smouth.[5]Itisaneasyjob.Becausesoundtravelsatafixedspeed,itarrivesatdifferentlocationsatslightlydifferenttimes.Soeachsoundlocationyieldsadifferentpatternoftimedelays.Thus,ifthesystemtakestheoutputsfromafixedarrayofmicrophonesandaddsthemtotimedelaysthatcharacterizeacertainlocation,itcanreinforcethesoundfromthatlocation.Thenitneedonlycomparethesoundwiththoseofknownwordsuntilamatchisfound.

Asmartroommustalsoknowwhoisspeakinginitortoit.Toactwithseemingintelligence,itisabsolutelyvitalthatasystemknowitsusers’identity.Whogivesacommandisoftenasimportantasthecommanditself.Thefastestwaytoidentifysomeonemaywellbetorecognizehisorherface.Sowedevelopedasystemforourroomstodojustthat.Toemploythemaximumlikelihoodapproach,thissystemfirstneededtobuildmodelsofallthefacesit"knew."WorkingwithM.I.T.graduatestudentsMatthewA.TurkandBabackMoghaddam,wefoundthatitwasimportanttofocusonthosefeaturesthatmostefficientlydescribedanentiresetoffaces.Weusedamathematicaltechniquecalledeigenvectoranalysistodescribethosesets,dubbingtheresults“eigenfaces”.Tomodelaface,thesystemdeterminedhowsimilarthatfacewastoeacheigenface.

Thestrategyhasworkedwell.Whenthecameradetectsaperson,theidentifyingsystemextractshisorherface-locatedbyPfinder-fromthesurroundingsceneandnormalizesitscontrast.Thesystemthenmodelsthefaceintermsofwhatsimilaritiesitbearstotheeigenfaces.Next,itcomparesthemodelwiththoseofknownpeople.Ifanyofthesimilarityscoresareclose,thesystemassumesthatithasidentifiedtheuser.Usingthismethod,oursmartroomshaveaccuratelyrecognizedindividualfaces99percentofthetimeamidgroupsofseveralhundred.

Facialexpressionisalmostasimportantasidentity.Ateachingprogram,forexample,shouldknowifitsstudentslookbored.Soonceoursmartroomhasfoundandidentifiedsomeone’sface,itanalyzestheexpression.Yetanothercomputercomparesthefacialmotionthecamerarecordswithmapsdepictingthefacialmotionsinvolvedinmakingvariousexpressions.Eachexpression,infact,involvesauniquecollectionofmusclemovements.Whenyousmile,youcurlthecornersofyourmouthandliftcertainpartsofyourforehead;whenyoufakeasmile,though,youmoveonlyyourmouth.InexperimentsconductedbyscientistIrfanA.Essaandme,oursystemhascorrectlyjudgedexpressions-amongasmallgroupofsubjects-98percentofthetime.3What?

Recognizingaperson’sface,expressionandspeechisjustthefirststep.Forhouses,officesorcarstohelpus,theymustbeabletoputthesebasicperceptionsincontext.Thesamemotions,afterall,canbeinterpretedquitedifferentlydependingonwhatthepersonmakingthemintends.Whenyoudriveacar,forexample,yousometimestakeyourfootfromthegaspedalbecauseyouwanttoslowdown.Butyoudothesamewhenyougetreadytomakeaturn.Thedifferenceisthatinpreparingforaturn,youadjustthesteeringwheelasyoumoveyourfoot.Soacomputersystemwouldneedtoconsiderhowyourmovementshadchangedovertime,incombinationwithothermovements,toknowwhatyouweredoingatanyonemoment.

Indesigningsuchasystem,weborrowedideasfromthescientistsworkingonspeechrecognition.Theymodelindividualwordsassequencesofsounds,or,astheycallthem,internalstates.Eachwordhasacharacteristicdistributionofinternalstates,whicharesometimesphonemes(thesmallestdistinguishableunitsofspeech)andsometimesjustpartsofphonemes.Acomputersystemtriestoidentifywordsbycomparingthesequenceofsoundstheycontainwithwordmodelsandthenselectingthemostlikelymatches.

Wegeneralizedthisapproachinthehopeofdeterminingpeople’sintentionsfromtheirmovements.Wedevisedacomputersystemthatcantell,forexample,whetherapersonwithonearmextendedispointingormerelystretching.Thesystemrecognizestheactioninvolvedinpointingbyreferringtoamodelhavingthreeinternalstates:raisethehand,holditsteadyandreturnitquickly.Thesystemseesstretchingasonecontinuousmovement.Sobyobservingtheseinternalstates-characterizedbytheaccelerationofthehandandthedirectionofitsmovement-oursystemworksoutwhatsomeoneisdoing.

Todate,wehavebuiltseveraldifferentsystemsforinterpretinghumanactionsinthisway.Thesimplestallowpeopletousetheirbodytocontrolvirtualenvironments.OnesuchapplicationistheArtificialLifeInteractiveVideoEnvironment(ALIVE),ajointprojectofMaes’sgroupandmyown.ALIVEutilizesthesmartroom’sdescriptionoftheuser’sshapetoplaceavideomodeloftheuserintoavirtual-realityscene,wherecomputer-generatedlife-formsreside.Thesevirtualcrittersanalyzeinformationaboutauser’sgestures,soundsandpositionstodecidehowtointeractwithhimorher.Silasthevirtualdog,forexample,playsfetch.

Whenasmart-roomusermimicsthemotionsinvolvedinpickingupandthrowingSilas’svirtualball,thedogseesthevideoimageintheALIVEenvironmentdothesameandgetsreadytochaseafteritstoy.Silasalsositsandrollsoveroncommand.Thesmartroom’soutputcanbeputtoworkinanevenmoredirectmanner.Theuser’sbodypositioncanbemappedintoacontrolspaceofsortssothathisorhersoundsandgestureschangetheoperatingmodeofacomputerprogram.Gameplayers,forexample,haveusedthisinterface,insteadofajoystickortrackball,tonavigatethree-dimensionalvirtualenvironments.Ifopponentsappearontheleft,theplayerneedonlyturntothelefttofacethem;tofireaweapon,theplayerneedonlysay,“Bang.”4Why?

Virtual-realitygamesaside,manymorepracticalapplicationsofsmart-roomtechnologyexist.ConsiderAmericanSignLanguage(ASL),asetofsophisticatedhandgesturesusedbydeafandmutepeople.Becausethegesturesarequitecomplex,theyofferagoodtestofourroom’sabilities.Hence,graduatestudentThadStarnerandIsetouttobuildasystemforinterpretingASL.Wefirstbuiltmodelsforeachsign,observingmanyexamplesofthehandmotionsinvolved,asdescribedbyPfinder.WefoundthatifwecomparedthesemodelswithPfinder’smodelsofanactualuserwhileheorshewassigning,wecouldtranslatea40-wordsubsetofASLinrealtimewithanaccuracyrateof99.2percent.Ifwecanincreasethesizeofthevocabularythatoursystemunderstands—anditseemsverylikelythatwewillbeabletodoso—itmaybepossibletocreateinterfacesfordeafpeopleasreliableasthespeech-recognitionsystemsthatarenowbeingintroducedforpeoplewhocanhear.[6]Automobiledrivers,too,standtobenefitfromsmart-roomtechnology.InmanypartsoftheU.S.,theaverageworkerspends10hoursaweekinacar.Morethan40,000motoristsdieintrafficaccidentseachyear,themajorityofwhichcanbeattributedtodrivererror.SotogetherwithAndyLiu,ascientistatNissanCambridgeBasicResearch,wehavebeenbuildingasmart-roomversionofacarinterior.Theultimategoalistodevelopavehiclethatcanmonitorwhatthedriverisdoingandprovideusefulfeedback,suchasroaddirections,operatinginstructionsandeventravelwarnings.Tocompileasetofdrivingmodels—includingwhatactionspeopletookwhentheywerepassing,following,turning,stopping,acceleratingorchanginglanes—weobservedthehandandlegmotionsofmanydriversastheysteeredtheirwaythroughasimulatedcourse.Weusedtheresultingmodelstoclassifyatestdriver’sactionasquicklyaspossible.Surprisingly,thesystemcoulddeterminewhatthedriverwasdoingalmostassoonastheactionhadstarted.Itclassifiedactionswithanaccuracyof86percentwithin0.5secondofthestartofanaction.Giventwoseconds,theaccuracyroseto97percent.

Wehaveshownthat,atleastinsimplesituations,itispossibletotrackpeople’smovements,identifythemandrecognizetheirexpressionsinrealtimeusingonlymodestcomputationalresources.[7]Bycombiningsuchcapabilities,wehavebuiltsmartroomsinwhich,freefromwiresorkeyboards,individualscancontrolcomputerdisplays,playwithvirtualcreaturesandevencommunicatebywayofsignlanguage.Suchperceptualintelligenceisalreadybeginningtospreadtoawidervarietyofsituations.Forinstance,wearenowbuildingprototypesofeyeglassesthatrecognizeyouracquaintancesandwhispertheirnamesinyourear.Wearealsoworkingontelevisionscreensthatknowwhenpeoplearewatchingthem.Andweplantodevelopcreditcardsthatcanrecognizetheirownersandsoknowwhentheyhavebeenstolen.

OtherresearchgroupsattheMediaLabareworkingtograntoursmartroomstheabilitytosenseattentionandemotionandtherebygainadeeperunderstandingofhumanactionsandmotivations.RosalindW.Picardhopestodeviseasystemthatcantellwhendriversorstudentsarenotpayingattention.AaronBobickiswritingsoftwaretointerpretthehumanmotionsusedinsports—imagineatelevisioncamerathatcoulddiscriminatebetweentwofootballplays,say,aquarterbacksneakandanendrun,andfollowtheaction.Assmart-roomtechnologydevelopsevenfurther,computerswillcometoseemmorelikeattentiveassistantsthaninsensibletools.Infact,itisnottoofar—fetchedtoimagineaworldinwhichthedistinctionbetweeninanimateandanimateobjectsactuallybeginstoblur.

1.?kidn.哄騙,取笑,開玩笑,小孩,小山羊v.哄騙,取笑,開玩笑,欺騙。

2.?baby-sittern.代人臨時(shí)照看嬰孩者。

3.?audiovisualn.(常用復(fù)數(shù))視聽設(shè)備,視聽教材adj.視聽的。

4.?gogglen.眼睛睜視,(復(fù)數(shù))風(fēng)鏡,護(hù)目鏡adj.睜眼的,瞪眼的;vi.眼珠轉(zhuǎn)動(dòng),瞪眼看vt.使瞪眼。

5.?tacklen.工具,滑車,用具,裝備,扭倒vt.固定,應(yīng)付(難事等),處理,抓住vi.捉住,扭住。Vocabulary

6.?statisticaladj.統(tǒng)計(jì)的,統(tǒng)計(jì)學(xué)的。

7.?blobn.一滴,水滴,斑點(diǎn)vt.濺污。

8.?texturedadj.織地粗糙的,手摸時(shí)有感覺的,有織紋的。

9.?correlatevt.使相互關(guān)聯(lián)vi.和……相關(guān)。

10.?torso未完成的(不完整的)作品,殘缺不全的東西。

11.?huen.色調(diào),樣子,顏色,色彩;叫聲,大聲叫喊,大聲反對(duì)。

12.?saturationn.飽和(狀態(tài)),浸潤(rùn),浸透,飽和度。

13.?compensationn.補(bǔ)償,賠償。

14.?anticipatevt.預(yù)期,期望,過早使用,先人一著,占先v.預(yù)訂,預(yù)見,可以預(yù)料。

15.?pedaln.踏板;?v.踩……的踏板。

16.?phonemen.[語]音位,音素。

17.?dimensionaladj.空間的。

18.?prototypen.原型。

19.?insensibleadj.無知覺的,無同情心的,硬心腸的,麻木不仁的。

20.?blurv.涂污,污損(名譽(yù)等),把(界線、視線等)弄得模糊不清,弄污n.污點(diǎn)。

[1]Thecomputersassesswhatpeopleinthesmartroomaresayinganddoing.Thankstothisconnection,visitorscanusetheiractions,voicesandexpressions-insteadofkeyboards,sensorsorgoggles-tocontrolcomputerprograms,browsemultimediainformationorventureintorealmsofvirtualreality.

計(jì)算機(jī)來判斷智能房間中的人在說什么，做什么。由于有了這個(gè)媒介，來訪者可以使用動(dòng)作、聲音和表情，而不是鍵盤、傳感器或視鏡，來控制計(jì)算機(jī)程序、瀏覽多媒體信息，或是在虛擬現(xiàn)實(shí)世界里探險(xiǎn)。注意破折號(hào)中間的插入語“insteadofkeyboards,sensorsorgoggles”，第二句句子框架為：VisitorscanuseAtodosomething。ImportantSentences

[2]Althoughthemodulestakeondifferenttasks,theyallrelyonthesamestatisticalmethod,knownasmaximumlikelihoodanalysis:thecomputerscompareincominginformationwithmodelstheyhavestoredinmemory.

盡管這些模塊有不同的任務(wù)，但它們都依賴同樣的統(tǒng)計(jì)學(xué)方法——極大似然法，計(jì)算機(jī)把輸入的信息和存儲(chǔ)器中已有的模型進(jìn)行匹配?！癿aximumlikelihoodanalysis”可譯為“最大似然法”。

[3]Itdescribeseachblobintwoways:asadistributionofvaluesfortheblob’scolorandplacement,andasaso-calledsupportmap,essentiallyalistindicatingwhichimagepixelsbelongtotheblob(pixelsare“pictureelements”,similartothedotsthatmakeupatelevisionimage).

它用兩種方法描述每一個(gè)塊的屬性：一種方法是描述這個(gè)塊的顏色和位置的值的分布，另一種方法是所謂的支持圖，本質(zhì)上是一張表，它描述哪一個(gè)圖像像素屬于這個(gè)塊(像素是“圖像元素”，這同組成電視圖像中的點(diǎn)相似)?！癳ssentiallyalistindicatingwhichimagepixelsbelongtotheblob”為“supportmap”的同位語。

[4]If,forexample,thebrightnessdifferencebetweenapixelandablobwere10percent,andtheblob’sstatisticssaidthatsuchadifferencehappenedonly1percentofthetime,thechancethatthepixelbelongedtotheblobwouldbeamereonein100.

舉例來說，如果一個(gè)像素和一個(gè)塊之間的亮度差為10%，而且塊的統(tǒng)計(jì)數(shù)值說明了當(dāng)時(shí)這個(gè)差的發(fā)生率僅為1%，則這個(gè)像素屬于這個(gè)塊的機(jī)會(huì)只有1%。注意“If,forexample,the…”句子結(jié)構(gòu)的使用。

[5]Borrowingthisinformation,thespeech-recognitionsystemelectronically“steers”anarrayoffixedmicrophonessothattheyreinforceonlythosesoundscomingfromthedirectionoftheuser’smouth.

借助于這種信息，語音識(shí)別系統(tǒng)電子化的“控制”由固定麥克風(fēng)組成的矩陣，加強(qiáng)從用戶嘴巴方向傳來的聲音。注意理解帶引號(hào)的謂語動(dòng)詞“steer”的含義，該詞的字面含義為“引導(dǎo)，駕駛”。

[6]Ifwecanincreasethesizeofthevocabularythatoursystemunderstands—anditseemsverylikelythatwewillbeabletodoso—itmaybepossibletocreateinterfacesfordeafpeopleasreliableasthespeech-recognitionsystemsthatarenowbeingintroducedforpeoplewhocanhear.

如果我們能往系統(tǒng)里加入系統(tǒng)可識(shí)別的詞匯——很有可能我們會(huì)這樣做——我們就有可能開發(fā)出與目前為正常人開發(fā)的接口一樣可靠的聾啞人語音識(shí)別系統(tǒng)。注意破折號(hào)中間的成分為“插入語”，在分析句子結(jié)構(gòu)時(shí)可忽略其存在。

[7]Wehaveshownthat,atleastinsimplesituations,itispossibletotrackpeople’smovements,identifythemandrecognizetheirexpressionsinrealtime