MIT Deep Learning 6.S191

MIT's introductory course on deep learning methods with applications to computer vision, natural language processing, biology, and more! HOME(current) OVERVIEW SCHEDULE TEAM F.A.Q. CourseDescription MIT'sintroductorycourseondeeplearningmethodswithapplicationstocomputervision,naturallanguageprocessing,biology,andmore!StudentswillgainfoundationalknowledgeofdeeplearningalgorithmsandgetpracticalexperienceinbuildingneuralnetworksinTensorFlow.Courseconcludeswithaprojectproposalcompetitionwithfeedbackfromstaffandpanelofindustrysponsors.Prerequisitesassumecalculus(i.e.takingderivatives)andlinearalgebra(i.e.matrixmultiplication),we'lltrytoexplaineverythingelsealongtheway!ExperienceinPythonishelpfulbutnotnecessary.ThisclassistaughtduringMIT'sIAPtermbycurrentMITPhDresearchers.Listenersarewelcome! TimeandLocation MonJan24-FriJan28,2022 1:00pm-4:00pmESTEveryday 1:00pm-1:45pm:LecturePart1 1:45pm-2:30pm:LecturePart2 2:30pm-4:00pm:SoftwareLabs ClasseswilltakeplaceonlinevirtuallyviaMITCanvasandpublicallyopen-sourcedeveryweek,startingMarch9th,2022. CourseSchedule LecturesoccuredduringMITIAP2022.Newlectures,slides,andlabswillbeopen-sourcedeveryweekstartingMarch11at10AMET! IntrotoDeepLearning Lecture1 Mar.11,2022 [Slides][Video] DeepSequenceModeling Lecture2 Mar.18,2022 [Slides][Video] IntrotoTensorFlow;MusicGeneration SoftwareLab1 [Code] DeepComputerVision Lecture3 Mar.25,2022 [Slides][Video] DeepGenerativeModeling Lecture4 Apr.1,2022 [Slides][Video] De-biasingFacialRecognitionSystems SoftwareLab2 [Paper][Code] DeepReinforcementLearning Lecture5 Apr.8,2022 [Slides][Video] LimitationsandNewFrontiers Lecture6 Apr.15,2022 [Slides][Video] LearningEnd-to-EndSelf-DrivingControl SoftwareLab3 [Code] AutonomousDrivingwithLiDAR Lecture7 Apr.22,2022 [Info][Video] SpeechRecognition Lecture10 Apr29,2022 [Info][Video] FinalProject Workonfinalprojects AI4Science Lecture9 May13,2022 [Info][Video] UncertaintyinDeepLearning Lecture8 May28,2022 [Info][Video] ProjectCompetition Projectpitchesandfinalawards! FrequentlyAskedQuestions Foranyotherquestionspleasereachouttothecoursestaffatintrotodeeplearning-staff@mit.edu. Arelistenersallowedtoattend? Alllistenersarewelcometoattend! IfyouareanMITstudent,pleaseformallyregisterasalisteneronWebsis. Clickhereforinstructions IfyouarenotanMITstudent,youcanstillattendthecoursewithoutregistering. Whatisthegradingpolicy? 6.S191isofferedasa6unitscourseandgradedP/D/Fbasedoncompletionofprojectproposalassignment.Listenersalsowelcome! HowcanIregister? IfyouareanMITstudentpleaseregisterhere.Youcandothisbyclicking"createnewform"andselecting"AddDrop".Enterthesubjectinformation(6.S191)and6unitswhenprompted.Youcanalsospecifyifyouwanttoberegisteredasalistenerorregularstudentthere. AllotherMITaffiliates(postdocs,faculty,staff,etc)areverywelcometoattendandparticipateinthecourse.Pleasesignupforthemailinglisttoreceiveupdates. AftertheMITcourse,thecoursewillbemadepublicallyavailabletonon-MITaffiliatesaswell.Again,pleasesignupforthemailinglisttoreceiveupdateswhenthisoccurs. Whatpre-requisitesarerequired? Weareexpectingveryelementaryknowledgeoflinearalgebraandcalculus.Howtomultiplymatrices,takederivativesandapplythechainrule.FamiliarityinPythonisabigplusaswell. Thecoursewillbebeginnerfriendlysincewehavemanyregisteredstudentsfromoutsideofcomputerscience. IsthereamailinglistIcanjoin? IfyouwouldliketoreceivecourserelatedupdatesandlecturematerialspleasesubscribetoourYouTubechannelandsignupforourmailinglist. Arethecoursematerialsopen-source? AllcoursematerialsavailableonlineforfreebutarecopyrightedandlicensedundertheMITlicense.Ifyouareaninstructorandwouldliketouseanymaterialsfromthiscourse(slides,labs,code),youmustaddthefollowingreferencetoeachslide: © AlexanderAminiandAvaSoleimanyMIT6.S191:IntroductiontoDeepLearningIntroToDeepLearning.com HowdoIreferencethesecoursematerials? AllcoursematerialsarecopyrightedandlicensedundertheMITlicense.Ifyouareaninstructorandwouldliketouseanymaterialsfromthiscourse(slides,labs,code),youmustaddthefollowingreferencetoeachslide: © AlexanderAminiandAvaSoleimanyMIT6.S191:IntroductiontoDeepLearningIntroToDeepLearning.com HowcanIhelpteachthisclass? IfyouareanMITstudent,postdoc,faculty,oraffiliateandwouldliketobecomeinvolvedwiththiscoursepleaseemailintrotodeeplearning-staff@mit.edu.Wearealwaysacceptingnewapplicationstojointhecoursestaff. HowcanIbecomeacoursesponsor? ThisclasswouldnotbepossiblewithoutouramazingsponsorsandhasbeensponsoredbyGoogle,IBM,NVIDIA,ErnstandYoung,LambdaLabs,TencentAI,Microsoft,Amazon,andOnepanel.Ifyouareinterestinginbecominginvolvedinthiscourseasasponsorpleasecontactusatintrotodeeplearning-staff@mit.edu. Wherearethepastcoursewebsites? Toviewarchivedversionsofthiswebsitefrompastyearspleaseclickherefor2021,2020,2019,2018,and2017. 6.S191Team AlexanderAmini LeadOrganizerInstructor AvaSoleimany LeadOrganizerInstructor TeachingAssistants CarmenMartinAlonso ShinjiniGhosh JodyMou SubhaPushpita ChristabelSitienei SamSledzieski NadaTarkhan Tsun-Hsuan(Johnson)Wang YudiXie FuningYang Wearealwaysacceptingnewapplicationstojointhecoursestaff.IfyouareinterestedinbecomingaTA,[email protected] Sponsors Thisclassanddeliverywouldnotbepossiblewithoutouramazingsponsors!Ifyouareinterestinginbecominginvolvedinthiscourseasasponsorpleasecontactusatintrotodeeplearning-staff@mit.edu Copyright©MIT6.S191.bannerimage;pagetemplate × BuildingLiDAR&PerceptionSoftwaretoMeetAutomotiveStandardsforSafeAutomatedDriving OmerKeilaf,CEO,Innoviz;AmirDay,HeadofCVandDeepLearning,Innoviz TalkAbstract Massproductionisamajorchallengeforautomakerssetondeliveringsafeautomateddriving.Forexample,theymustconsiderimpairedvisioncausedbyweather,lighting,andotheradverseroadconditions,includingslopesandcurvaturesfrompotholesorhills.InnovizCEOOmerKeilafwillofferanintroductiontoautomotiverequirementsthatcarmakersaskforinordertomeettheirtargets.AmirDay,Innovizdirectorofcomputervision,willthenmatchtheserequirementswithaccompanyingperceptionsoftwarechallengesandshareexperiencesfromhowInnovizsolvesthem. SpeakerBio OmerKeilafhasspentthegreaterpartofthepasttwodecadesdrivingcuttingedgetechnologiesfrominceptiontocommercialization.AsCo-FounderandCEOofInnovizTechnologies,OmerisworkingtirelesslytobringsafeAVstoconsumermarketsbydeliveringhigh-performance,automotive-gradeLiDARandperceptionsoftwaretoautomakersatmassmarketprices.Lastyear,KeliaftookhiscompanypublicthroughaSPACmergerwithCollectiveGrowthCorporation,raisingover$370million.PriortofoundingInnoviz,OmerservedasanofficerintheeliteIDFintelligenceUnit81,andheldseniorleadershiprolesatvariouscompanies.OmerholdsaBScandMScinElectricalEngineeringandanMBA,allfromTelAvivUniversity. AmirDayistheDirectorforComputerVisionatInnovizTechnologies.Withmorethanadecadeofexperienceexecutingcomplex,multi-disciplinarytechnologyprojectsfromconceptiontoproduction,Amir'steamdevelopedInnoviz'sperceptionsoftware(InnovizAPP)fromscratch,focusingondeeplearningandsoftwareinfrastructureasameanstointerpretInnoviz'srichpointclouddata.BeforeworkingatInnoviz,AmirservedintheIDF'sMilitaryIntelligenceCorps,leadingateamof100+engineerstodevelopnew,novelsystemssolutions.Today,Amirfocuseson3DComputerVision,DeepLearningandDataPipelineforautonomousvehiclesperception. × HowRev.comharnesseshuman-in-the-loopanddeeplearningtobuildtheworld'sbestEnglishspeechrecognitionengine MiguelJette,HeadofAIR&D;andJenniferDrexler,SeniorSpeechScientist TalkAbstract Rev.comisoneoftheworld'slargesttranscriptioncompaniesbyvolume,processingover2500hoursofaudioandvideoeveryday.Revhasturnedtranscriptionintoa"virtuouscycle"betweenhumantranscribersandautomaticspeechrecognition(ASR)-automatically-generatedfirstdraftsmakehumansmoreefficientandhuman-editedtranscriptsareusedtotrainbetterandbetterASRmodels.Inthistalk,wewilldiscussthestate-of-the-artindeepneuralnetworkmodelsforspeechrecognitionandhowRevhasbuilttheworld'sbestEnglishASRengine.WewillcoverseveraldifferentmodelarchitecturesforASR,exploringtheimpactofmodelingchoicesonaccuracy,speed,andtheeaseofimplementingadditionalfeatureslikewordtimingsandcustomization.WewillendwithanoverviewofRev'scurrentresearcheffortsandfutureplans. SpeakerBio MiguelJettéistheHeadofAIResearchandDevelopmentatRevAI.Miguelhasover20yearsofexperienceinspeechrecognitionandmachinelearning,andholdsdegreesinMathematicsandComputerSciencefromMcGillUniversity.Miguelispassionateaboutleveragingmathematics,computerscience,statistics,andtheirintersectiontosolveimportantproblemsthatimprovethequalityofhumanlife. JenniferDrexlerisaSeniorSpeechScientistatRev,wheresheworksonproblemsinspeechrecognitionandlanguageprocessing.ShecompletedherPhDinElectricalEngineeringandComputerScienceatMIT,wheresheworkedonenablingspeechtechnologiesinlow-resourcesettings,andhasworkedonmachinetranslationintheHumanLanguageTechnologyGroupatMIT’sLincolnLabs. × DevelopingPrincipledAlgorithmsforAI4Science AnimaAnandkumar,DirectorofMLResearch,NVIDIA;ProfessoratCaltech TalkAbstract comingsoon! SpeakerBio AnimaAnandkumarisaBrenprofessorinComputingandMathematicalSciencesatCaltech,andadirectorofmachinelearningresearchatNVIDIA.AtCaltech,sheistheco-directoroftheDesign,Optimization,andLearninginitativeandco-leadstheAI4scienceinitiative.AtNVIDIA,sheisleadingaresearchgroupthatdevelopsnext-generationAIalgorithms.Amongsthermanyscientificcontributions,shespearheadedthedevelopmentoftensoralgorithms,whicharearecentraltoachievingmassiveparallelisminlarge-scaleAIapplications.ShehasbeenrecognizedbyseveralawardsincludinganAlfred.P.SloanFellowship,NSFCareerAward,andfacultyfellowshipsfromMicrosoft,Google,andAdobe.AnimareceivedherB.TechinElectricalEngineeringfromIITMadrasin2004andherPhDfromCornellUniversityin2009,andcompletedpostdoctoralresearchatMIT. × PracticalUncertaintyEstimationandOut-of-DistributionRobustnessinDeepLearning JasperSnoek,ResearchScientist,Google TalkAbstract Deeplearningmodelsarebadatsignallingfailure:Theytendtomakepredictionswithhighconfidence,andthisisproblematicinreal-worldapplicationssuchashealthcare,self-drivingcars,andnaturallanguagesystems,wherethereareconsiderablesafetyimplications,orwheretherearediscrepanciesbetweenthetrainingdataanddatathatthemodelmakespredictionson.Thereisapressingneedbothforunderstandingwhenmodelsshouldnotmakepredictionsandimprovingmodelrobustnesstonaturalchangesinthedata.Inthistalk,I'llgiveaveryabridgedversionofmyNeurIPStutorialonuncertaintyandrobustnessindeeplearningandthenintroducesomemorerecentworkdevelopedtoaddressthesechallenges. SpeakerBio JasperSnoekiscurrentlyastaffresearchscientistatGoogleBrain.Recentlyhisresearchhasfocusedonmethodsforimprovinguncertaintyandrobustnessofdeeplearningmethods.HisinterestsspanavarietyoftopicsattheintersectionofBayesianmethodsanddeeplearning.HecompletedhisPhDinmachinelearningattheUniversityofToronto.HesubsequentlyheldpostdoctoralfellowshipsattheUniversityofToronto,underGeoffreyHintonandRuslanSalakhutdinov,andattheHarvardCenterforResearchonComputationandSociety,underRyanAdams.Jasperco-foundedthemachinelearningstartupWhetlab,whichwasacquiredbyTwitter.HehasservedasanAreaChairforNeurIPS,ICMLandICLR,andorganizedavarietyofworkshopsatICMLandNeurIPS.