面向边缘智能的协同推理综述王睿1,2齐建鹏1陈亮1杨龙11（北京科技大学计算机与通信工程学院北京100083）2（北京科技大学顺德研究生院广东佛山528300）（wangrui@ustb.edu.cn）SurveyofCollaborativeInferenceforEdgeIntelligenceWangRui1,2,QiJianpeng1,ChenLiang1,andYangLong11（SchoolofComputerandCommunicationEngineering,UniversityofScienceandTechnologyBeijing,Beijing100083）2（ShundeGraduateSchoolofUniversityofScienceandTechnologyBeijing,Foshan,Guangdong528300）AbstractAtpresent,thecontinuouschangeofinformationtechnologyalongwiththedramaticexplosionofdataquantitymakesthecloudcomputingsolutionsfacemanyproblemssuchashighlatency,limitedbandwidth,highcarbonfootprint,highmaintenancecost,andprivacyconcerns.Inrecentyears,theemergenceandrapiddevelopmentofedgecomputinghaseffectivelyalleviatedsuchdilemmas,sinkinguserdemandprocessingtotheedgeandavoidingtheflowofmassivedatainthenetwork.Asatypicalscenarioofedgecomputing,edgeintelligenceisgainingincreasingattention,inwhichoneofthemostimportantstagesistheinferencephase.Duetothegenerallowperformanceofresourcesinedgecomputing,collaborativeinferencethroughresourcesisbecomingahottopic.Byanalyzingthetrendsofedgeintelligencedevelopment,weconcludethatcollaborativeinferenceattheedgeisstillintheincreasingphaseandhasnotyetenteredastablephase.Wedivideedge-edgecollaborativeinferenceintotwoparts:Intelligentmethodsandcollaborativeinferencearchitecture,basedonathoroughinvestigationofedgecollaborativeinference.Theinvolvedkeytechnologiesaresummarizedverticallyandorganizedfromtheperspectiveofdynamicscenarios.Eachkeytechnologyisanalyzedinmoredetail,andthedifferentkeytechnologiesarecomparedhorizontallyandanalyzedontheapplicationscenarios.Finally,weproposeseveraldirectionsthatdeservefurtherstudyingincollaborativeedgeinferenceindynamicscenarios.Keywordsedgecomputing；edgeintelligence；machinelearning；edgecollaborativeinference；dynamicscenario摘要近年来，信息技术的不断变革伴随数据量的急剧爆发，使主流的云计算解决方案面临实时性差、带宽受限、高能耗、维护费用高、隐私安...
