Google Research

Our teams aspire to make discoveries that impact everyone, and core to our approach is sharing our research and tools to fuel progress in the field. Our ... Advancingthestateoftheart GoogleResearchtackleschallengesthatdefinethetechnologyoftodayandtomorrow. Ourapproach Ourteamsaspiretomakediscoveriesthatimpacteveryone,andcoretoourapproachissharingourresearchandtoolstofuelprogressinthefield. Ourresearcherspublishregularlyinacademicjournals,releaseprojectsasopensource,andapplyresearchtoGoogleproducts. Seeourresearchphilosophy Exploreasampleofourresearch ResearchersatGoogleareworkinginmanydomains. SeesomeofourlatestresearchdevelopmentsfromtheGoogleAIblogandelsewhere. ResolvingHigh-EnergyImpactsonQuantumProcessors HighlightedResearch ResolvingHigh-EnergyImpactsonQuantumProcessors CanRobotsFollowInstructionsforNewTasks? Robotics CanRobotsFollowInstructionsforNewTasks? LaMDA:TowardsSafe,Grounded,andHigh-QualityDialogModelsforEverything NaturalLanguageProcessing LaMDA:TowardsSafe,Grounded,andHigh-QualityDialogModelsforEverything DoesYourMedicalImageClassifierKnowWhatItDoesn’tKnow? MachineIntelligence DoesYourMedicalImageClassifierKnowWhatItDoesn’tKnow? SeparatingBirdsongintheWildforClassification MachinePerception SeparatingBirdsongintheWildforClassification AccurateAlphaMattingforPortraitModeSelfiesonPixel6 MachinePerception AccurateAlphaMattingforPortraitModeSelfiesonPixel6 IntroducingStylEx:ANewApproachforVisualExplanationofClassifiers MachinePerception IntroducingStylEx:ANewApproachforVisualExplanationofClassifiers   WereimaginetechnologyacrossallareasofComputerScienceresearch. Learnhowwechallengeconventions. Seeourresearchareas Publications Wepublishhundredsofresearchpaperseachyearandpresentourworkinawiderangeofvenues. Seesomeofourmostrecentresearch. OnFacilityLocationProblemintheLocalDifferentialPrivacyModel PreviewAbstract Westudytheuncapacitatedfacilitylocation(UFL)problemundertheconstraintsimposedbythelocaldifferentialprivacy(LDP).Recently,Guptaetal.(2009)andEsencayietal.(2019)proposedlowerandupperboundsfortheUFLproblemonthecentraldifferentialprivacy(DP)modelwhereacuratorfirstcollectsalldatabeforebeingprocessed.Inthispaper,wefocusontheLDPmodel,where... Viewdetails OnFacilityLocationProblemintheLocalDifferentialPrivacyModel ChenglinFan,DiWang,MarcoGaboradi,ShiLi,VincentPierreCohen-addad,YunusEsencayi 25thInternationalConferenceonArtificialIntelligenceandStatistics(AISTATS2022)(2022)(toappear) ArchitecturesforProtectingCloudDataPlanes PreviewAbstract Thispaperexploresthreeapproachesforprotectingcloudapplicationdataplanestopreventunauthorizedaccesstotheapplicationanditsdataandtopreventunwanteddataexfiltration.Throughanexplorationofvariousconcretesecurityarchitectures,wefocuson(1)CloudSecurityPerimeterstoprovideaboundaryarounddataandinfrastructureinthecloudthatprovidesalineofdefense... Viewdetails ArchitecturesforProtectingCloudDataPlanes GrantDasher,InesEnvid,BradCalder Google(2022) DeepNullmodelsnon-linearcovariateeffectstoimprovephenotypicpredictionandassociationpower PreviewAbstract Genome-wideassociationstudies(GWAS)examinetheassociationbetweengenotypeandphenotypewhileadjustingforasetofcovariates.Althoughthecovariatesmayhavenon-linearorinteractiveeffects,duetothechallengeofspecifyingthemodel,GWASoftenneglectsuchterms.HereweintroduceDeepNull,amethodthatidentifiesandadjustsfornon-linearandinteractivecovariateeffects... Viewdetails DeepNullmodelsnon-linearcovariateeffectstoimprovephenotypicpredictionandassociationpower AndrewWalkerCarroll,BabakAlipanahi,CoryMcLean,FarhadHormozdiari,NickFurlotte,TaedongYun,ThomasColthurst,ZacharyRyanMccaw (2022) ExploringtheLimitsofLargeScalePre-training PreviewAbstract Recentdevelopmentsinlarge-scalemachinelearninghavecreatedatemptingpicturesuggestingthatbyscalingupdata,modelsizeandtrainingtimeproperly,onecanobtainamodelthatcanbeusedsuccessfullyinfew-shotsettingsinalldownstreamtasks.Inthiswork,weinvestigatethispremiseempiricallyandprovideastrongcaseagainstit.Inparticular,weconsiderimagerecognition... Viewdetails ExploringtheLimitsofLargeScalePre-training SamiraAbnar,MostafaDehghani,BehnamNeyshabur,HanieSedghi ICLRSpotlight(2022) LaMDA:LanguageModelsforDialogApplications PreviewAbstract WepresentLaMDA:LanguageModelsforDialogApplications.LaMDAisafamilyofTransformer-basedneurallanguagemodelsspecializedfordialog,whichhaveupto137Bparametersandarepre-trainedon1.56Twordsofpublicdialogdataandwebtext.Whilemodelscalingalonecanimprovequality,itshowslessimprovementsonsafetyandfactualgrounding.Wedemonstratethatfine-tuningwith... Viewdetails LaMDA:LanguageModelsforDialogApplications AaronDanielCohen,AdamRoberts,AlejandraMolina,AlenaButryna,AliciaJin,ApoorvKulshreshtha,BenHutchinson,BenZevenbergen,BlaiseHilaryAguera-Arcas,Chung-chingChang,ClaireCui,CosmoDu,DanielDeFreitasAdiwardana,DehaoChen,Dmitry(Dima)Lepikhin,EdH.Chi,ErinHoffman-John,Heng-TzeCheng,HongraeLee,IgorKrivokon,JamesQin,JamieHall,JoeFenton,JohnnySoraker,KathyMeier-Hellstern,KristenOlson,LoraMoisAroyo,MaartenPaulBosma,MarcJosephPickett,MarceloAmorimMenegali,MarianCroak,MarkDíaz,MatthewLamm,MaximKrikun,MeredithRingelMorris,NoamShazeer,QuocV.Le,RachelBernstein,RaviRajakumar,RayKurzweil,RomalThoppilan,StevenZheng,TaylorBos,TojuDuke,TulseeDoshi,VinodkumarPrabhakaran,WillRusch,YaGuangLi,YanpingHuang,YanqiZhou,YuanzhongXu,ZhifengChen arXiv(2022) ExploitingfermionnumberinfactorizeddecompositionsoftheelectronicstructureHamiltonian PreviewAbstract Achievinganaccuratedescriptionoffermionicsystemstypicallyrequiresconsiderablymanymoreorbitalsthanfermions.PreviousresourceanalysesofquantumchemistrysimulationoftenfailedtoexploitthislowfermionicnumberinformationintheimplementationofTrotter-basedapproachandoverestimatedthequantum-computerruntimeasaresult.Theyalsodependedonnumericalproceduresthat... Viewdetails ExploitingfermionnumberinfactorizeddecompositionsoftheelectronicstructureHamiltonian SamMcArdle,EarlCampbell,YuanSu Phys.Rev.A,vol.105(2022),pp.012403 Multi-TaskLearningwithSequence-ConditionedTransporterNetworks PreviewAbstract Enablingrobotstosolvemultiplemanipulationtaskshasawiderangeofindustrialapplications.Whilelearning-basedapproachesenjoyflexibilityandgeneralizability,scalingtheseapproachestosolvesuchcompositionaltasksremainsachallenge.Inthiswork,weaimtosolvemulti-tasklearningthroughthelensofsequence-conditioningandweightedsampling.First,weproposeanewsuiteof... Viewdetails Multi-TaskLearningwithSequence-ConditionedTransporterNetworks MichaelLim,AndyZeng,BrianAndrewIchter,MaryamBandari,ErwinJohanCoumans,ClaireTomlin,StefanSchaal,AleksandraFaust InternationalConferenceonRoboticsandAutomation2022,IEEE(toappear) Transducer-BasedStreamingDeliberationForACascadedEncoderModel PreviewAbstract Previousresearchondeliberationnetworkshasachievedexcellentrecognitionquality.Theattentiondecoderbaseddeliberationmodelsoftenworksasarescorertoimprovefirst-passrecognitionresults,andoftenrequiresthefullfirst-passhypothesisforsecond-passdeliberation.Inthiswork,weproposeastreamingtransducer-baseddeliberationmodel.Thejointnetworkofatransducer... Viewdetails Transducer-BasedStreamingDeliberationForACascadedEncoderModel ArunNarayanan,KevinHu,RuomingPang,TaraNSainath,TrevorDeatrickStrohman ICASSP2022(2022)(toappear) CVSSCorpusandMassivelyMultilingualSpeech-to-SpeechTranslation PreviewAbstract WeintroduceCVSS,amassivelymultilingual-to-Englishspeech-to-speechtranslation(S2ST)corpus,coveringsentence-levelparallelS2STpairsfrom21languagesintoEnglish.CVSSisderivedfromtheCommonVoicespeechcorpusandtheCoVoST2speech-to-texttranslation(ST)corpus,bysynthesizingthetranslationtextfromCoVoST2intospeechusingstate-of-the-artTTSsystems.Twoversions... Viewdetails CVSSCorpusandMassivelyMultilingualSpeech-to-SpeechTranslation YeJia,MichelleTadmorRamanovich,QuanWang,HeigaZen(ByunghaChun) arXiv(2022) Seeourpublications Teams&people Meetthepeoplebehindourinnovation Ourteamsadvancethestateoftheartthroughresearch,systemsengineering,andcollaborationacrossGoogle. Seeourteams Seeourpeople Joinus We'realwayslookingformoretalented,passionatepeople Ourglobalreachmeansthatresearchteamsacrossthecompanytackletoughproblemstogether. Exploreopportunities Exploreouroutreachprograms