《中國移動研究院：2022年6G無線內生AI架構與技術白皮書（英文版）（36頁）.pdf》由會員分享，可在線閱讀，更多相關《中國移動研究院：2022年6G無線內生AI架構與技術白皮書（英文版）（36頁）.pdf（36頁珍藏版）》請在三個皮匠報告上搜索。

1、（2022）6G Native AI Architecture andTechnologies White PaperChina Mobile Research Institute（CMRI）Contents1Preface.12Driving Forces.22.1Challenges of 5G intelligent networks.22.2New scenarios of 6G ubiquitous intelligence.33Definition and Scope.43.1Definition of 6G Native AI.43.2Scope of 6G Native AI.

2、44New Idea.54.1Quality of AI service(QoAIS).54.2AI lifecycle orchestration management.94.3Deep integration of AI computing and communication.105New Architecture.125.1Data plane.145.2Smart plane.165.3Extended control and user planes.186New technologies.206.1AI model selection and finetuning.206.2AI m

3、odel training with terminal and network collaboration.226.3AI model inference with terminal and network collaboration.246.4AI performance pre-validation based on digital twin.267Summary and Outlook.27Abbreviations.30Authors.30Reference.31China Mobile6G Native AI Architecture and Technologies white p

4、aper11 PrefaceArtificial intelligence(AI)has developed rapidly in the past decade,which has surpassed theHuman Intelligence in modeling the nonlinear laws of big data samples and online accuratedecision-making in interaction with the environment,and has achieved great success in the fieldsof compute

5、r vision,natural language processing,and robot control.The reason for the rapiddevelopment of AI is,on one hand,the breakthrough of artificial intelligence algorithmsrepresented by deep learning and reinforcement learning;on the other hand,the rapid decline onthe cost and popularization of artificia

6、l intelligence computing power represented by GPUs.Since 5G,AI has gradually been widely used in mobile communication networks,such asnetwork configuration optimization at the network management level,resource schedulingoptimization at the network element level,and even the physical layer of the air

7、 interface.Inaddition,there are also more and more AI applications on the terminal side.Towards the future,6G network needs to facilitate the digitalization and intelligentization of thousands of industries,and it needs to provide intelligent services with less latency and better performance than cl

8、oudintelligence.For operators,network operation costs need to be greatly reduced,and networkoperation and maintenance needs to evolve from local intelligent scenarios to high-level networkautonomy.At present,AI applications are mainly based on centralized cloud resources.Cloud serversaggregate large

9、 amounts of data,utilize centralized computing power to preprocess them,and trainand validate AI models.However,transmitting a large amount of raw data in the network will notonly put enormous pressure on network transmission bandwidth and performance indicators(suchas latency),but also bring great

10、challenges to data privacy protection.Besides,due to the lack ofcomputing power,algorithms and data,there is still much room for improvement in the intelligentapplications on the terminal side.In the face of the above challenges,it is necessary to introduce native AI capabilities into thenetwork,aba

11、ndon the patched mode of AI applications,and realize the deep integration ofcommunication connection,computing,data and AI models at the network architecture level,where the distributed computing power and data in the network are fully utilized for thecoordination mechanisms between multiple nodes a

12、nd between terminals and the network,andrealization of the integration of distributed and centralized processing.In this way,not only dataprivacy can be protected,the efficiency of data processing,the timeliness of decision-making andreasoning,and utilization efficiency of network nodes can also be

13、improved.This white paper firstintroduces the driving forces and application scenarios of native intelligence.The demand fornative AI support by 6G network is derived from the current status of intelligent networkapplications,the requirements on high-level network autonomy,ubiquitous intelligence,hi

14、ghvalue network services,extreme service experience,and network safety and trustworthiness.Then,the paper elaborates on the definition and scope of native AI,and proposes the deep integration ofAI computing power,data,algorithms and network connections.Besides,the new concepts of 6Gnative AI are int

15、roduced including AI service quality(QoAIS),orchestration of AI workflows ofits full life cycle,computing and communication integration,and integration of native AI andChina Mobile6G native AI architecture and technical white paper2digital twins.The new architecture driven by native AI is proposed a

16、nd described in detail,including data plane,smart plane and extended control plane and user plane,and new technologiesare introduced including AI model orchestration,distributed model training,distributed modelinference,pre-validation and optimization of digital twins.Finally,the future research dir

17、ectionsare prospected.2 Driving ForcesThe application of artificial intelligence technology in 5G networks has promoted theintelligent development of mobile communication networks and vertical industries,but theapplication mode of patching and plug-in hinders the effectiveness of AI applications.At

18、the sametime,the application and exploration of artificial intelligence in all walks of life has put forwardrequirements for new basic capabilities of future networks.To realize the vision of ubiquitousintelligence,6G networks need to build native AI capabilities.2.1 Challenges of 5G intelligent net

19、worksIn the 5G era,intelligent network practices requires the integration of AI technologies withthe hardware,software,systems,and processes of 5G communication networks,and AI is used tohelp communication networks realize intelligent planning,construction,maintenance andoptimization,so as to improv

20、e quality and efficiency and reduce cost.The utilization of AIpromotes the technological and systematic transformation of the network itself,enables agilebusiness innovation,and promotes the construction of intelligent networks,including cloudnetwork equipment intelligence,network operation intellig

21、ence,and network service intelligence.In 5G network,AI is mainly used for optimization of communication connections and serviceprocesses.Although the service cloud has been introduced,network and cloud is loosely coupledsince the 5G architecture,protocol functions and processes have been finalized,a

22、nd onlyincremental iterations can be made to the existing architecture solutions.The challenges of 5G intelligent network practices based on patched-on and plug-in AI are asfollows:-The lack of a unified framework leads to the lack of effective verification and guaranteemethods for AI performance.Th

23、e verification of AI application effects is carried outafter the fact,so the overall end-to-end process is long and complex,and theintermediate process generally requires a lot of human intervention.The relative largeimpact on the network makes it difficult to quickly promote AI in the process ofapp

24、lying it to the existing network.-The plug-in mode is difficult to achieve a fully automatic closed-loop of pre-validation,online evaluation and optimization.AI model training usually requires the preparationof a large amount of training data.In the plug-in mode,it is difficult to collect and labeld

25、ata on the existing network in a centralized manner,and the transmission and storageChina Mobile6G native AI architecture and technical white paper3overhead is also large,resulting in a long iteration cycle of the AI model,high trainingoverhead,slow convergence,and poor model generalization.-In the

26、plug-in mode,computing power,data,models and network connections belong todifferent technical systems,and no standardized interfaces and interaction rules aredefined between them.Cross-system collaboration is carried out on the managementplane,leading to extra latency of seconds and even minutes and

27、 the unguaranteedquality of service.2.2 New scenarios of 6G ubiquitous intelligenceNative AI refers to the network supporting AI through native design patterns at thearchitectural level,rather than patched-on or plug-in patterns.The driving forces of the nativedesign pattern are as follows:-The netw

28、ork provides ubiquitous AI services:To realize the vision of ubiquitousintelligence,the 6G network needs to help the digitalization and intelligentation ofthousands of industries and realize the on-demand supply of intelligent capabilities anytime,anywhere.Compared with cloud service providers,6G ne

29、tworks need toprovide intelligent capability services with higher real-time performance and betterperformance and at the same time provide federal intelligence between industries torealize cross-domain intelligent integration and sharing.On the other hand,due to thelarge amount of data in the termin

30、al,the computing power of the terminal is also gettingstronger and stronger.Considering the data privacy requirements,the native intelligentcollaborative network and the computing power,communication connection andalgorithm model of the terminal and other resources are required,such as computingpowe

31、r offloading.,model orchestration,etc.,to provide 2C customers with the ultimatebusiness experience and high-value new services.-AI provides services for the network:6G networks need to achieve a high level ofautonomy,security and trustworthiness.At present,the level of network autonomy isnot high(t

32、he level of autonomous driving network is about 2.2),and it is necessary tointroduce native AI capabilities in the network to support the perception and realizationof the intentions of operators and users,and realize the self-design,self-implementation,self-optimization,and self-development of the n

33、etwork.Evolution,and ultimatelyachieve a high level of network autonomy.In addition,in the future,the network willcarry more diversified services,serve more application scenarios,and carry more typesof data.Therefore,the network will face a large number of new and complex attackmethods.The security

34、capabilities based on native AI are embedded in all aspects of the6G network to realize autonomous detection of threats,autonomous defense or assist indefense.It can be seen from the above driving force analysis that in addition to meeting basiccommunication needs,6G networks also need to consider t

35、he integration of computing,data,models/algorithms,etc.,that is,6G needs to be designed through native AI at the architecturalChina Mobile6G native AI architecture and technical white paper4level to meet the Diversified new business scenarios and network autonomous optimizationrequirements for netwo

36、rk AI include AI applied to network optimization and user experience(such as air interfaces rewritten with AI),as well as various AI services required by third parties.3 Definition and ScopeDeep integration with AI will be considered in 6G design stage,i.e.,6G native AI.Unlike 5G,which treats AI fun

37、ctions as added-on features,6G native AI will exercise an end-to-endorchestration and control of computing power,data,and models.The key ingredients,such asconnections,computing,data and AI algorithms/models are inherently integrated and meanwhilethe on-demand orchestrating those key ingredients int

38、o wireless,transmission,core network,etc.are supported,which provides the inherent intelligence capabilities required for high-levelnetwork autonomy and diversified business needs.That is the native AI capability of 6G,whichwill make network intelligence more efficient and perform better.At the same

39、 time,the networkintelligence will also be expanded accordingly,not only accelerating the continuous optimizationof network performance,but also providing intelligent service capabilities,which enables thedigital and intelligent transformation of various industries.Network intelligence will continue

40、 toevolve in the 6G era,promoting a truly intelligent native network.3.1 Definition of 6G Native AI6G network native AI is operating within the 6G network architecture,which provides datacollection,data preprocessing,model training,model inference,model evaluation and other entirelife cycle AI workf

41、lows.The key ingredients of AI services,e.g.,computing power,data,algorithms,connections,and network functions,protocols,procedures are deeply integrated intothe initial design of 6G network.6G network native AI aims at providing real-time and efficientintelligent services and capabilities for high-

42、level network autonomy,ubiquitous intelligence forindustrial users,ultimate service experience for subscribers,and native network security.3.2 Scope of 6G NativeAIThe existing mobile communication network is mainly connection-oriented data transmission,which requires the transmission link guarantee

43、based on QoS(Quality of Service)(such as datarate,delay,etc.).However,the native intelligence is required to implement end-to-end control andorchestration of computing power,models and data.Thus,there are enormous gap towards thenetwork design and implementation and operation.Therefore,it is critica

44、l to consider the abovedistinct requirements at the beginning of 6G network design.On one hand,new concepts such asqualityassuranceofAI-basedservice,end-to-endorchestrationandcomputingandcommunication integration should be introduced,on the other hand,new architecture designs suchChina Mobile6G nati

45、ve AI architecture and technical white paper5as AI data plane,AI intelligent plane,extended control plane and user plane should also beconsidered.The quality of AI service assessment and assurance should be built for native AI,and thenbased on quality of AI service the end-to-end AI life-cycle orche

46、stration will be implemented,including computing power,AI models,data and connections.Native AI requires deep integration of computing and communication.Considering that thecapabilities of native AI are distributed into a large number of network nodes,which are usuallyrestricted in data collection,c

47、omputing power,bandwidth,and delay,and thus it is vitallyimportant to adopt the co-design of computing and communication resources coordination.Inaddition,it is necessary to rethink the network architecture,protocols and functions,which areadaptable to air interface transmission and the performance

48、optimization of native AI.4 New IdeaIt is quite complicated to integrate AI with the traditional connection oriented network at thebeginning of 6G network design,which requires cross-domain expertise.It is vital to think out ofbox of the traditional design paradigm and incorporate fresh AI ingredien

49、ts and concepts.Webelieve that the assessment and assurance of AI service quality,the orchestration and managementof AI life cycle,and the deep integration of AI computing and communication will become thebasic concepts of native AI systems.Facing various industries and scenarios,there are diversifi

50、ed demands for 6G native AInetwork.The first question,we need to answer,is how to translate user demands into network AIservice capabilities?We propose the concept of AI service quality,namely QoAIS(Quality of AIService),and believe that the network should provide an assessment and assurance for QoA

51、IS.Next,how to evaluate and continuously satisfy QoAIS and implement QoAIS assurance requiresthe involvement of the management plane,control plane,and user plane.From a managementplane perspective,we propose the orchestration and management of the AI life cycle workflow,i.e.,the semi-static allocati

52、on of network resource,such as computing power,data,algorithms,andconnections,to satisfy QoAIS requirements;from control and user plane perspective,it isimportant to allocate real-time network resource to continuously satisfy QoAIS,in which the deepintegration of AI computing and communication is th

53、e key.4.1 Quality ofAI service(QoAIS)QoAIS is a set of metrics profiles of AI service quality assessment and the correspondingassurance mechanisms.1.6G networks will provide inherent AI capabilities,which can serve avariety of intelligent applications,namely AIaaS.Considering that the requirements o

54、f differentintelligent scenarios towards the quality of AI services are expected to be highly diverse,and thus,a set of indicators is required to express user-level needs and network orchestration and control(including AI models/algorithms,computing power,data,connections,etc.)in aquantitativeorhier

55、archical way.China Mobile6G native AI architecture and technical white paper6The native 6G AI services can be categorized into the following types,i.e.,AI data,AItraining,AI inference,AI validation and etc.Each type of AI service requires a different set ofQoAIS.The QoS of the communication service

56、in traditional communication networks mainlyconsiders the performance related to the connectivity,such as delay and data rate(MBR,GBR,etc.).Besides that,the 6G network will introduce a variety of resource dimensions for AI serviceorchestration and control,such as distributed heterogeneous computing

57、resources,storageresources,data resources,and AI models/algorithms.Therefore,6G native AI service qualityshould be evaluated from multiple dimensions of network resources,such as connection,computing,algorithm,and data.At the same time,with the implementation of the carbonneutrality and carbon peak

58、policies,the global industry of intelligent applications will pay moreattention to data security and privacy,and network automation.In the future,performance-relatedKPIs will no longer be the only indicators to be highlighted,and the requirements for security,privacy,autonomy and resource overhead w

59、ill gradually play more important roles and becomecrucial dimensions for evaluating AI service quality.Therefore,from the initial design,the QoAISindicators needs to consider performance,overhead,security,privacy and autonomy and otheraspects as well.Table 4.1-1：QoAIS indicators of AI training servi

60、ceTypes of AIServicesEvaluationdimensionsQoAIS indicatorsAI trainingperformancePerformancebounds,trainingtime,generalization,reusability,robustness,interpretability,consistencybetween loss function and optimization objective,fairnessoverhead*Storageoverhead,computingoverhead,transmissionoverhead,ene

61、rgy consumptionSafety*Storage security,computing security,transmission securityprivacy*Data privacy level,algorithm privacy levelautonomyFully autonomous,partially manually controllable,andfully manually controllableNote*:Common evaluation indicators between different types of AI servicesAmong them,

62、the performance bounds are the upper and lower bounds for evaluating themodel performance,such as model accuracy,recall rate.Generalization refers to the ability of apre-trained model adaptive to make predictions on new data.Reusability is the ability of a modelto continue to function in case of app

63、lication scenarios change.Robustness refers to the capabilitythat the model can still work even if the input data is distorted,attacked or uncertain.Interpretability refers to the degree to which a models internal mechanisms can be understood.Consistency between the loss function and the optimizatio

64、n goal refers to the degree ofconsistency between the design of the loss function and the optimization goal during the modelChina Mobile6G native AI architecture and technical white paper7training process,for example,whether the number of variables considered in the loss functioncompletely covers th

65、e optimization goals.Autonomy refers to the requirements for autonomousoperation and manual intervention in the workflow of AI data/training/validation/inferenceservices,reflecting the degree of automation of AI services.Autonomy is divided into three levels:complete autonomy(full automated AI servi

66、ce,without any manual intervention in the wholeprocess),partial manual intervention(some workflows of AI service are automated,while othersrequire manual assistance),all manual controllable(All aspects of the AI service workflows arehandled manually).In addition to the evaluation dimensions shown in

67、 the table above,QoAIS can also includeapplication specific performance indicators.Taking channel state information compression as anexample 2,normalized mean square error(NMSE)or cosine similarity can be selected as theKPIs for channel recovery accuracy,or link-level/system-level indicators(such as

68、 block error bitrate or throughput,etc.)as KPIs reflecting the impact of channel feedback accuracy on systemperformance.In addition,QoAIS can also include the availability of AI services,the response timeof AI services(from the user initiating the request to the first response message of the AI serv

69、ice)and other general evaluation indicators not related to the specific type of AI service.QoAIS is a key input for the network native AI orchestration and control.The top-levelQoAIS will be broken down by network AI management and orchestration system,and then bemapped to the specific QoS requireme

70、nts for data,algorithms,computing,connections,and etc.Figure 4.1-1:QoS indicators decomposed into QoS indicators in each resource dimensionThe above figure shows the mapping between each indicator of QoAIS and thecorresponding QoS metrics.The overall QoAIS indicators of AI services are decomposed in

71、toChina Mobile6G native AI architecture and technical white paper8QoAIS indicators in each indicator dimension,and further mapped to QoS indicators in eachresource dimension,which are guaranteed by the mechanisms from the management plane,thecontrol plane and the user planes perspective.The QoS indi

72、cators in each resource dimension inthe figure can be divided into quantitative indicators and leveled indicators(such as security level,privacy level,and autonomy level).For the former category of indicators,the quantificationschemes for some indicators are mature or relatively easy to formulate(su

73、ch as training time,algorithm performance,calculation accuracy,various resource costs,etc.),while there arecurrently no quantitative evaluation methods for other indicators(such as model robustness,reusability,generalization and interpretability,etc.),as shown in Table 4.1-2.Therefore,it is a keyiss

74、ue to design sufficiently open and inclusive network architecture in the initial stage so that themature quantitative technologies of the above indicators can be gradually introduced.Table 4.1-2:Mapping of AI training service performance QoAIS to each resource dimensionMetricdimensionQoAISindicatorr

75、esourcedimensionQuantitativeindicatorsNoquantitativemetrics yetperformancePerformancebounds,trainingtime,generalization,reusability,robustness,interpretability,optimizationtargetmatching,fairnessdataFeatureredundancy,completeness,dataaccuracy,anddatapreparation timeSamplespacebalance,integrity,sampl

76、edistributiondynamicsalgorithmPerformanceindexbounds,training time,convergence,optimizationtargetmatching degreeRobustness,reusability,generalization,interpretability,fairnesscomputingpowerComputationalaccuracy,duration,efficiencyconnectBandwidth and jitter,delay and jitter,biterror rate and jitter,

77、reliability,etc.In terms of quality assessment and assurance mechanism,there are still some problems forthe QoS mechanism of 5G network,such as coarse service differentiation granularities,longoptimization and adjustment period,and inability of radio resource management adaptive todynamic fluctuatio

78、n of network and services.Therefore,it is also necessary to consider how todesign end-to-end efficient QoAIS mechanisms and procedures when proposing QoAIS indicatorsfor assessing and assuring AI services in 6G networks.Extended question：China Mobile6G native AI architecture and technical white pape

79、r91.After the introduction of AI services in the network,users may have different requirementsfor security and privacy.The QoS and security design are independent in traditionalcommunication networks,and how to co-design QoS and security for 6G native AI?2.Currently there are no mature quantitative

80、evaluation methods for some QoAIS indicators(such as model generalization,interpretability,and reusability 3).How to design aninclusive architecture to accommodate gradually introduced quantitative technologies forsuch indicators?4.2 AI lifecycle orchestration managementThe AI life cycle refers to t

81、he life cycle of an AI workflow in the network,that is,thegeneration,execution,monitoring,evaluation,optimization,completion,and deletion of an AIworkflow.Network Native AI Workflow refers to one or more tasks that the network needs tocomplete step by step in order to complete an AI service.Currentl

82、y,AI has a similar end-to-endworkflow in various industrial applications 4,which can be divided into four flows:datamanagement,model learning,model validation,and model deployment.Figure 4.2-1 shows acommon AI end-to-end workflow.Figure 4.2-1:Common AI end-to-end workflowAt present,in the practice o

83、f 5G network intelligence,most of the AI workflows isimplementedoffline,independentfromthenetworkoperatingenvironment,andthesilo-development model is adopted between different intelligent applications(that is,for eachintelligent application the research and development is carried out independently w

84、ithout resourcecoordination and sharing),which leads to low efficiency and high cost.The 6G network willprovide a complete operating environment for end-to-end AI workflows of various intelligentapplications.Depending on the type of AI service,the AI workflow may include different tasks.Not all ofth

85、e corresponding workflows are end-to-end.For example,the workflow of AI data service onlyincludes tasks related to data management;AI validation workflows can include both datamanagement and model validation related tasks,or only model validation tasks;AI trainingworkflow can include only model lear

86、ning,or both data management and model learning,depending on whether the data provided by users is sufficient for the quality requirements.The AIChina Mobile6G native AI architecture and technical white paper10inference workflow can only include tasks related to model deployment or can include tasks

87、related to data management and model deployment simultaneously.For an intelligent applicationrequesting multiple AI services at the same time,the corresponding workflow may last end-to-end.Figure 4.2-2 shows the relationship between the native AI workflow and AI services in the 6Gnetwork.Figure 4.2-

88、2:diagram of native AI services and workflows in 6G networksThe 6G network generates the required workflow and relative tasks for each AI service,andthen orchestrates the respective resources(computing,algorithms,data,connections,etc.)for eachtask in the workflow to ensure continuous assurance of th

89、e QoAIS.In this process,themanagement plane is responsible for collecting performance monitoring data,evaluating QoAIS,and analyzing the impact of the task design and resource management,so as to continuouslyoptimize the schemes and strategies and realize intelligent orchestration management.Extende

90、d question：1.In order to ensure the continuous satisfaction of QoAIS,is it adequate to rely solely onmanagement plane to orchestrate resources required by the workflow?Does the control planeneed to be involved?How do management and control work together?4.3 Deep integration ofAI computing and commun

91、icationIn order to ensure the continuous achievement of QoAIS,in addition to realizing intelligentorchestration and management of the AI workflows from the management plane,it is alsonecessary to realize the deep integration of AI computing and communication on the control planeand user plane.The co

92、mputing resources in the traditional communication network mainly serve thecommunication services.The computing resources are integrated in the network equipmentprocessing board,and the computing resources are allocated based on the predefined proceduresof the communication services.In contrast with

93、 communication services,AI services arecomputing intensive.In recent years,various processor architectures(GPU,NPU,DPU,TPU,etc.)have been emerging to improve computing efficiency and reduce energy consumption.The keyChina Mobile6G native AI architecture and technical white paper11requirements for co

94、mputing of native AI services in 6G networks are high computing efficiency,low energy consumption,and low latency.Although the computing efficiency of centralizedcomputing resources in the cloud is high,it is hard to meet low-latency requirements for edge AIapplications.The computing power resources

95、 of edge nodes are limited,while the scale is largeand the real-time performance is better.Therefore,it is preferable to coordinate computingresources between edge and cloud,which is expected to meet computing performancerequirements for various AI services.The edge computing capabilities have been

96、introduced in 5G MEC solutions to providelow-latency services,however the network connection and computing are loosely coupled,andthere is room for further improvement in terms of efficiency,deployment cost,security,andprivacy protection.For example,in the 5G MEC solution 5,the user plane function U

97、PF in corenetwork can be co-located with the MEC,but they are still two relatively independent systemsfrom logical architecture and control managements perspectives.When the adjustments arerequired for network connection and computing power simultaneously,it is coordinated throughthe management plan

98、e,which leads to relatively large delay.On the other hand,the computingresources deployed in the cloud,edge and device are distributed and heterogeneous.If thecoordination is adaptive to dynamic and complex radio environment in a real-time way,it iscrucial to provide real-time support from control p

99、lane and user planes perspective.Taking mobile networks as an example,there are potential three co-design modes for thedeep integration of AI computing and communication illustrated in Figure 4.2-3.Figure 4.2-3：Three modes of AI computing and communication co-designMode 1:A new logical computing uni

100、t(NC,Computing Node)is introduced into the mobilenetwork architecture,which is independent from the base station.CRC(Computing ResourceControl)in NC interacts with RRC(Radio Resource Control)in xNB control plane throughspecified interfaces.The advantage of this mode is that it allows inter-vendor de

101、ployment betweenbase stations and computing unit vendors,however,relatively long interaction delay may beintroduced.Mode 2:The computing unit as an inner function is built in the base station.The RRC andCRC interact with each other via an internal interface.The advantage of this mode is that betterp

102、erformance may be achieved,and meanwhile,radio communication resources and computingresources are independently controlled and coordinated on demand,which ensures the scalabilityof RRC and CRC design to some extent.China Mobile6G native AI architecture and technical white paper12Mode 3:The logical c

103、omputing unit is built in the base station and RRC and CRC are mergedinto a unified resource control entity(xRC),which controls the connection and computingresources at the same time.The advantage is that the control decisions on connection andcomputing resources can be made at the same time,which l

104、eads to the best performance of thecoordinated connection and computing resource control.However,it is more complex to designsuch mechanism and meanwhile scalability issue may come into being.By implementing deep integration of computing and communication on the control plane,more efficient measures

105、 can be provided to continuously achieve the QoAIS targets.Theadvantage is that when some QoAIS indicator is deteriorating,some policy or scheme can bequickly optimized.For example,if connection bandwidth is restricted and local computing poweris sufficient,it is preferable to compress the AI data w

106、ith more sophisticated algorithms with highrecovery accuracy and lower transmission bandwidth;when the connection bandwidth is sufficientand stable,but the local computing power is limited,it is preferable to increase the localcomputing capabilities by collaborating with surrounding nodes.For user p

107、lane,the deepintegration of AI computing and communication is mainly reflected in the joint design andoptimization of AI computing protocols and communication protocols.In terms of computingprotocols,for the same AI computing task,different protocols and configuration parameters maybe required for h

108、eterogeneous computing resources,which eventually have an impact oncomputing accuracy and computing time.In terms of communication protocols,various types ofdata from AI task(such as model intermediate results,model weights,model gradients,etc.)shallbe processed optimally considering the instability

109、 of bandwidth and channel state,such as thecoding and encoding of source and channel.In addition,since the computation andcommunication of AI tasks are often sequentially processed in time domain,it provides apossibility to allow joint design and optimization.Extended question：1.How to effectively c

110、onsider co-design of computing and communication by coordinatingmanagement plane and control plane,to achieve a balanced network resource allocation,andbetter resource and energy consumption efficiency?2.How to jointly design and optimize the computing protocol and communication protocol ofAI tasks

111、from user planes perspective to meet the performance and overhead requirements atthe same time?5 New ArchitectureThe integration of AI resource ingredients into 6G network architecture design is the mostfundamental feature for 6G native AI.As the three basic ingredients of AI(data,algorithms andcomp

112、uting power)have become as fundamental resources as network connections,the design ofthe corresponding architecture,interfaces,protocols should be reflected through the entire AI lifecycle.In addition to its own internal management,control,processing and transmission,eachresource ingredient will als

113、o cooperate with others to meet QoAIS requirements.Therefore,unlikeChina Mobile6G native AI architecture and technical white paper135G network,new data plane,smart plane,and computing plane will be defined in 6G network,andtraditional control plane and user plane are expected to be extended as well.

114、The following figure5-1 shows the logical architecture of the 6G native AI network.Figure 5-1:logical architecture of the 6G native AI networkHorizontally 6G networks can be divided into resource layer,network function layer andapplication and service layer(three layers).The resource layer provides

115、radio access,computing,storage and other underlying resources,and provides corresponding support and services for theupper layers.In network function layer a specific network function is formed by combining one ormore network functions together to provide network service capabilities to application

116、and servicelayers.The application and service layer provides corresponding support for customers businessneeds.In terms of vertical logical architecture,except the communication plane that carriestraditional communication services,New data plane,computing plane and smart plane areintroduced in 6G ne

117、twork.The data plane is responsible for data collection,cleaning,processingand storage in end-to-end network,and provides data services to other layers.The computingplane provides a unified computing power warehouse,perceives computing power requirements,manages computing tasks,provides computing ro

118、uting,computing modeling,and meanwhileprovides computing services for other layers.The intelligent plane provides the operatingenvironment for full life-cycle of native AI.It invokes the services provided by data plane andcomputing plane,and provides intelligent services for other layers.The managem

119、ent planeprovides operation and maintenance of all other layers and planes.For 6G native AI,the implementation of the following new concepts is mainly reflected inthe three layers and the intelligent plane,data plane,computing plane and management plane.Itis worth mentioning that the control plane a

120、nd user plane belong to the network function layer.The control plane and user plane traditionally are oriented to support communication services.After the introduction of the data plane,computing plane and intelligent plane,some new servicedata will be generated from these planes,such as data collec

121、ted and transmitted on the data plane,input,output,and intermediate data of computing tasks on the computing plane,parameters of AImodels on the intelligent plane and etc.In order to provide support such new service,the controlplane and user plane are also extended as well.This chapter focuses on th

122、e data plane,smart plane,and extended control and user planes.China Mobile6G native AI architecture and technical white paper145.1 Data plane5G network intelligence practices 8 show that the data collection is difficult,and the dataquality is hard to guarantee.There are limited specified interfaces

123、for data collection in previousnetwork architecture and protocol design,and the data collection partly relies on implementation,such as deep packet inspection or data probing,and meanwhile it is hard to guarantee datacollection in a timely manner.There are some problems for management-based data col

124、lection,such as few types of data,long collection period(more than 15 minutes),inconsistent data formats,naming,and calculation from different vendors,and it is also difficult to open southbound networkmanagement interface.At the same time,due to the instability of data collection,the transmissionlo

125、ss,the limited storage in the network management equipment,and the difficulty in obtaininglabels,the collected data is often not of good quality,such as missing,non-labeled or label errors.Before AI model training,it takes a lot of time and labor cost to pre-process the input data.Facing the above c

126、hallenges,“data plane”7 is introduced to 6G network architecture.Thedata elements in the data plane will cover internal and external data of the network,includingservice data,user data,network data,sensing data and so on.The data services include datacollection,data preprocessing,data storage,data a

127、ccess,data sharing and collaboration,etc.Basicdata services have the following technical characteristics:support for trusted authentication,authorization,access,efficient data storage and management,on-demand data collection,datapreprocessing and aggregation,open interface for external access,etc.Fi

128、gure 5.1-1 shows thelogical functional architecture of the 6G network data plane.China Mobile6G native AI architecture and technical white paper15Figure 5.1-1:6G network data plane functional architectureThe data plane architecture of the 6G network consists of a central data center and local datace

129、nters in each domain,which adopts a hybrid centralized and distributed architecture.The centraldata center stores the end-to-end global data,and orchestrates the data globally on demand;thelocal data centers within each domain store and manage the data collected from the local network,and provide da

130、ta services for various upper-layer applications.Since the data required and generated by AI services also is processed by data plane,such astraining samples,AI model parameters,model intermediate results,model gradients,inferencesamples,and inference results,etc.Various data services provided by da

131、ta plane can be invokedthroughout the entire life cycle of AI workflows.For example,trusted services provided by dataplane can be invoked to guarantee the trustworthiness requirements of AI service defined inQoAIS 9;computing and transmission overheads can be reduced by invoking on-demand datacollec

132、tion and preprocessing services.In the 6G network,trustworthiness is expected to be a key requirement for data services 10.The credibility of data services is mainly reflected in the stages of data collection,data storage,data access,data sharing and collaboration.Regarding data collection,data priv

133、acy,fairness,reproducibility and robustness are major considerations.Data privacy is mainly guaranteed bysome data processing technologies,such as debias sampling and annotation,data sources tracing,China Mobile6G native AI architecture and technical white paper16data anonymousness and differential

134、privacy.Data fairness is mainly assessed via quantitativeindicators,such as the correlation coefficient of variables,loss function,complete Cartesianproduct,etc.The reproducibility and robustness of data collection can be guaranteed by datasource tracing.Extended question：1.How to support the openne

135、ss and use of internal data in network entities from the networkarchitecture level?2.How to support the on-demand data extraction from the network architecture level?Including the type of collected data,the amount of collected data,the collection method,thedata preprocessing method,etc.5.2 Smart pla

136、neIn the previous chapter,the new concepts involving the design of the management plane,control plane and user plane have been introduced.These new mechanisms provide a completeoperating environment for the entire life cycle of various AI workflows,in order to satisfy therequirements of QoAIS.This c

137、omplete operating environment is called as the smart plane of the6G network.Figure 5.2-1 shows the functional architecture design of the smart plane of the 6Gnetwork.Figure 5.2-1:6G network smart plane functional architectureThe smart plane of 6G native AI network has the following technical charact

138、eristics:The first technical feature is the self-generation of AI use cases.An AI use case is a one-timeAI service request made by a user to the network.An AI use case may involve one or more typesof network native AI services(such as AI training,validation,and inference services),which isregulated

139、by the AI use case description.From this description,the network can learn informationon intelligent application scenarios,input and output data,model selection,model training,modelvalidation,and decisions on model outputs.The network can generate AI use case descriptions byitself based on its own d

140、ata analysis or external instructions.The management plane is responsiblefor managing all AI use cases,scheduling and implementing AI use cases,generating theChina Mobile6G native AI architecture and technical white paper17corresponding AI services,AI workflows,and QoAIS requirements,and provisionin

141、g networkresources(including data,algorithms,computing power,connections,etc.).Figure 5.2-2 shows thelogical relationship between AI use cases,AI services,AI workflows,and AI tasks.Figure 5.2-2:The logical relationship between AI use cases,AI services,AI workflows,and AI tasksSecond,QoAIS generation

142、.QoAIS is used for quality evaluation for native AI services in thenetwork.An AI service corresponds to a set of QoAIS,and the QoAIS corresponding to an AI usecase is composed of the respective QoAIS of all AI services it contains.When the networkreceives an AI use case,it needs to know the QoAIS re

143、quirements corresponding to the use case,so that it can be decomposed into the specific requirements for the orchestration,scheduling,andcontrol of various network resources.The AI use case description can be imported externally andcan be generated internally,e.g.,AI use cases may be generated based

144、 on upper-layer intentinformation.Third,the entire life cycle of AI workflow is carried within the network.Various AIworkflows can be generated by the network management plane,including data collection,preprocessing,data expansion,and data analysis;model selection,training,parameter adjustment;model

145、 verification,integration,monitoring,and updating,etc.And then the required resources areorchestrated,monitored,and optimized to meet QoAIS requirements.In such scenarios high levelautomation is required without human intervention.Fourth,the management plane,control plane and user plane collaborate

146、with each other toensure the continuous achievement of QoAIS,which is mainly achieved through the orchestrationand control of the three key ingredients of AI(algorithm,computing power,data)and networkfunction(connection).The management plane is responsible for resource scheduling in the initialstage

147、 and non real-time resource allocation adjustment.The control plane and user plane performreal-time QoS assurance according to the dynamic changes of the network environment.Fifth,the combination of AI centralized and distributed architecture.The central AIsupercomputer has sufficient computing powe

148、r,massive storage capacity.It is suitable forintelligent applications with large scale models,high performance requirements,and non real-timerequirements.The central AI unit in each domain of the wireless,transmission,and core networkacts as a centralized AI engine in the respective domain,and is re

149、sponsible for the AI use casesthat can be completed in the local domain.The edge nodes distributed in each domain havelimited computing power and storage,and will support intelligent applications with high real-timerequirements through cooperation between network entities.When the QoAIS of an AI use

150、 case inthe local domain cannot be achieved within the single domain(such as lack of feature data ofChina Mobile6G native AI architecture and technical white paper18other domains,lack of computing resources),the use case will be escalated to the central AIsupercomputer and achieved through global re

151、source orchestration.This hybrid architecture canrelieve the performance pressure caused by a single centralized architectures.s5.3 Extended control and user planesThe control plane and user plane of the existing mobile communication network are designedto meet the quality requirements of traditiona

152、l communication services(including voice,datapacket transmission,streaming media,etc.),and their main purpose is to provide connections fordata transmission,support user mobility,and ensure service experience.In terms of resource types,dedicated computing resources are used,and the demand for comput

153、ing and storage resources isnot high.Unlike traditional communication services,AI services are data-intensive andcomputing-intensive services.New resource dimensions will be introduced for the native AIservices including heterogeneous computing and storage resources,new computing tasks,as wellas the

154、 AI data required and generated by AI services.It is necessary to design a management andcontrol mechanism for new dimension resources,and at the same time,it is important to design anefficient user plane mechanism for the input,output and in-process data of AI services,that is,AIservices will becom

155、e a special user of 6G networks.These will greatly expand the control planeand user plane in traditional mobile communication networks.We call the new control plane and user plane with extended protocols and procedures tosupport QoAIS as AI Control Plane(AI CP)and AI User Plane(AI UP)respectively.Ta

156、ble 5.3-1shows the comparison of AI CP and AI UP with CP and UP in traditional mobile communicationnetworks.Table 5.3-1:Comparison of AI CP and AI UP with CP and UP in traditional mobilecommunication networkstraditionalcommunicationbusinessNative AI ServicesConnectionConnectionComputingpowerAlgorith

157、mDataMultidimensionalresourcesNF CPcontrolmechanismsfor AIconnectionscontrolmechanismsfor AIcomputingpowercontrolmechanisms forAI algorithmself-optimizationAIon-demanddynamicdatacollectionandprocessingcontrolmechanismAI CPChina Mobile6G native AI architecture and technical white paper19NF UPtranspor

158、tmechanismsfor AI dataexecutionmechanismsfor AIcomputingtasksprocess flow forAI algorithmself-optimizationprocessingmechanismsfor the dataAI needsAI UPThe reason why the traditional control plane and user plane are often end-to-end sincetraditional communication services usually involve terminals,wi

159、reless,transmission and corenetwork domains.However,the workflows of AI services no longer last end-to-end.Instead,thecontrol plane of AI is responsible for control multiple dimensional resources to complete aspecific task rather than end-to-end communication,and similarly the user plane of AI iscom

160、posed of data processing on multiple dimensional resources.There is an enormous gap between QoAIS and QoS of communication services;for example,AI data may include training samples,inference results,model parameters,intermediatecalculation results of training/inference,model gradients,etc.And the tr

161、ansmission mode,datatype,data volume of AI data is quite different from those of data from communication services.Inaddition,the impact caused by radio channel changes,user mobility,and user distribution aredifferent between AI services and communication services.It is not known yet whether theexist

162、ing control and transmission mechanism of communication networks still applicable.It maybe necessary to design a special connection control mechanism and data transmission protocol forAI services,or it may be feasible to take advantage of the same functional module to serve bothtraditional communica

163、tion services and AI services.Figure 5.3-1 shows two possible modesbetween AI connection and traditional connection regarding control mechanism and datatransmission protocol.Figure 5.3-1:The relationship between AI connection and traditional connection in controlmechanism and data transmission proto

164、colExtended question：China Mobile6G native AI architecture and technical white paper201.What are the pros and cons of providing AI services via the application layer?Is it capable toprovide AI services by utilizing the existing control and user plane of communicationnetwork?If not,what areas need to

165、 be improved or innovated?2.What is the relationship between connections,computing power,algorithms,and dataresources regarding the data processing and control mechanisms?6 New technologiesFor 6G native AI network,each stage of AI life cycle,including data collection,modelselection,model training,mo

166、del inference,performance evaluation and optimization,requirescorresponding technologies support.Unlike cloud AI with centralized deployment,a large scaledistributed nodes in a wireless network usually require collaboration to complete AI tasks.It isindispensable to design specific mechanisms consid

167、ering computing and communicationintegration.Depending on the specific AI task,the collaboration approach and integrationmechanism might be different.On the other hand,in order to ensure that AI training and inferencedoes not have a negative impact on network performance,it is particularly important

168、 to setup adigital twin network for a physical network.New techniques related to model orchestration,training,and inference will be introduced in the first three sections,and then the interactionbetween native AI and digital twin will be described in the last section.6.1 AI model selection and finet

169、uningThe AI model orchestration is one of the key technologies for the orchestration andmanagement of the entire AI life cycle.During the entire life cycle management of AI models,including model training,model validation,model inference,and model transfer,it is important toselect the appropriate ba

170、seline model considering model complexity,inference overhead,retraining overhead,especially in the case of the dynamic changes in computing power andcommunication resources for wireless network.Through theoretical analysis and experiments,it is found that the depth and complexity of theAI model have

171、 a significant impact on the performance of the model.The more complex the AImodel,the greater the probability of a more optimal solution tends to be 11.But the morecomplex the model,the more overhead required to retraining.In order to reduce training overhead(such as training data collection,comput

172、ing power),the common practice is to introduce modelretraining,that is,select a baseline model with good performance well trained on a large data setof the source domain,and then use the target domain data to retrain the baseline model to learn thedistribution bias between the source and target doma

173、ins.A key problem that needs to be solved for model retraining is to choose a suitable baselinemodel structure and weights.To apply AI model selection and retraining to wirelesscommunication systems,the following factors need to be considered:China Mobile6G native AI architecture and technical white

174、 paper211.Due to the difference caused by nonlinear power amplifier of transceivers in real system,complex air interface fluctuation,user distribution and movement and other factors,it isvitally important to introduce retraining to improve the adaptability of AI model.However,it is difficult to accu

175、mulate adequate channel state information,and thus it isnecessary to utilize small samples to retrain the baseline model to achieve better transferperformance in the target domain.2.It is also necessary to consider retraining the baseline model on distributed nodes,oreven on terminals.Due to limited

176、 computing resources and power consumption,it isimportant to choose a suitable baseline model while ensuring low retraining overhead.The performance and cost comparison are listed in Table 6.1-1 under three training setting.Among them,basic learning is based on random weights of a basic model;transf

177、er learning uses alarge amount of data in the source domain to train the weights of the basic model;similarly,meta-learning 12 also uses the training samples in the source domain,and it aims at learningmeta-knowledge,i.e.transferable characteristics,rather than simply fitting the training samples.Ta

178、ble 6.1-1:Comparison of performance and cost of under three training settingbasic learningtransfer learningmeta-learningsourcedomainmodelBasic modelrandom weightsLearning model structureand weights based on largeamounts of dataLearning from largeamounts of dataMeta knowledgesourcedomainmodeltraining

179、overheadwithoutLarger overhead(1x)Max(10 times)target domainperformanceSmall samples:lowperformanceLarge samples：highperformanceLarge distributiondifferences:mediumperformanceSmall distributiondifferences:highperformanceHigh performancefor small samplesRetrainingoverheadhigh costlow overheadlow over

180、headFrom the above table,it is important to extend the model orchestration function of AIretraining to implement the selection of AI baseline models.Specifically,in order to enableoptimal selection of the AI baseline model training method,the data distribution differencebetween the source domain and

181、 the target domain should be evaluated first.For example,if thereis a large distribution difference and sufficient training resources,it is preferable to considermeta-learning to train the baseline model on the source domain dataset.Moreover,the datadistribution of the source domain dataset can be a

182、nalyzed to arrange the transfer feature dataset tospeed up the baseline model training.If the distribution differences are small and trainingresources are relatively limited,it is preferable to consider transfer learning to train the baselineChina Mobile6G native AI architecture and technical white

183、paper22model.In addition,the performance of the baseline model with the best performance in the sourcedomain is not necessarily the best candidate in the target domain.We can also leverage the datacharacteristics and distribution of the target domain to enable optimal selection of the baselinemodel.

184、6.2 AImodeltrainingwithterminalandnetworkcollaborationDistributed AI model training refers to utilizing distributed computing resources deployed onthe cloud,edge,and terminals to perform AI model training,which can improve computingresource utilization,enhance model performance,and protect data priv

185、acy.As described inSection 4.3,with the deep integration of computing and communication,the coordination ofdistributed heterogeneous computing resources is considered to adapt to the dynamic and complexcommunication environment in a real-time way,which leads to real-time support from the controlplan

186、e and user plane,see chapters 5.3 Extended Control and User Plane.For data-driven AI model training,traditional methods include model training based oncentralized computing power and data,or model training based on distributed parallel computing.The latter is usually processed in a computer cluster,

187、i.e.,splitting data or model into differentcomputing nodes for parallel computing,and producing the final result on centralized computingnodes by aggregation processing 13.Because the computer cluster network condition isrelatively stable and reliable,and the training dataset distribution is usually

188、 known,the theoreticalmodeling is relatively easy,and the performance of model training is easy to guarantee.However,in the mobile communication network,there are some complex issues,e.g.,unstable channelquality,user mobility,and non-IID distribution of training data.Therefore,it is unavoidable tode

189、sign more complex distributed model training schemes in mobile network than in a computercluster.At present,many technical frameworks of distributed AI model training have been proposedin industries and academia,such as(layered)federated learning 14,group learning 15,multi-agent learning 16,and mode

190、l segmentation-based learning 1718 et al.However,mostof research is based on certain theoretical or ideal assumptions,rather than complex networkenvironments.In this case,can the performance of model training be guaranteed?Is thecommunication resource overhead and efficiency acceptable?These are iss

191、ues to be studied.It is believed that when training a distributed AI model between terminals and base stationsin a wireless network,a large volume of intermediate data will be generated during the trainingprocess,and as a consequence,the radio resources will be frequently occupied.In addition,airint

192、erface transmission delay and bit error rate will degrade the training results.In order to ensurethe convergence of the model and meanwhile improve the utilization of radio resources of the airinterface,it is a worthwhile to introduce a higher-order model learning algorithm with higherefficiency 192

193、021.Since model learning algorithms of different orders(zero-order,first-order stochastic gradient descent,second-order Newton method,etc.)have their ownChina Mobile6G native AI architecture and technical white paper23advantages and disadvantages in terms of training speed and resource overhead,it i

194、s beneficial todynamically adjust the learning algorithm according to the wireless channel state.Figure 6.2-1 is adiagram of the dynamic selection of various learning algorithms,and Figure 6.2-2 shows thefunctional interaction designed to introduce this dynamic selection mechanism.Figure 6.2-1:Schem

195、atic diagram of the dynamic selection ofvarious learning algorithmsFigure 6.2-2:Functional interaction of dynamic selection mechanisms ofvarious learning algorithmsAs mentioned in 5.3,a new control and data transmission protocol is required to beintroduced for AI connection in air interface,which ar

196、e represented by Dtrain_C and Dtrain_Urespectively in the above figure.Among them,Dtrain_C is the control function entity responsiblefor collaborate the terminals and the base station to perform AI model training.In theaforementioned scenario,this entity dynamically adjusts the model learning algori

197、thm accordingto the change of the channel condition.Dtrain_U is the service plane functional entity responsiblefor the AI model training collaboration between the terminal and the base station.It includes adedicated protocol stack required to carry information such as model parameters,gradients orgr

198、adient norms between the base station and the terminals.Since the transmission scheme of theabove-mentioned information data is different from that of traditional communication services,and the reliability requirements are also different,it is necessary to re-design the air interfacetransmission pro

199、tocol accordingly.Extended question：China Mobile6G native AI architecture and technical white paper241.The technical improvement by considering both network connection characteristics and AImodel training characteristics can theoretically improve the feasibility and performance of6G native AI networ

200、k,but it does not fundamentally change the data-driven paradigm.Is itpossible to explore a new model-driven training paradigm,so as to implement self-growth ofalgorithms/models?6.3 AImodelinferencewithterminalandnetworkcollaborationDistributed AI model collaborative inference refers to the utilizati

201、on of distributed computingresources on the cloud,edge,and terminal to perform AI model inference,which can improve theutilization of computing resources,compensate insufficient computing power on terminals,andprotect data privacy.Collaborative inference based on model partitioning is a distributed

202、collaborative inferenceframework in wireless networks proposed by the industry in recent years.When a terminal needsto complete a model inference task and its own computing power is not sufficient,it can requestthe assistance of the computing resources from the network side to jointly complete infer

203、ence task.The decision that needs to be made is how to split the model,for example,for the following deepneural model in Figure6.3-1,the beginning two layers are split from the remaining layers,whichmeans the left part of the neural model is running on the terminal side,and the right part is running

204、on the network side.Figure 6.3-1:Diagram of end-to-end cooperative inferenceLatency and accuracy are two important performance indicators for AI inference.Thedecision on the model split point will affect the computation overhead of the terminal side and thebase station side,as well as the amount of

205、data that needs to be transmitted over the air interface.Therefore,the factors that should be taken into account may include model splitting point,terminal computing resource allocation,air interface radio resource allocation,and base stationcomputing resource allocation,etc.,and thus it is importan

206、t to coordinate scheduling ofcommunicationandcomputingresourcestimely.EspeciallywhentherearemultipleChina Mobile6G native AI architecture and technical white paper25heterogeneous computing resources on both the terminal and the base station,the allocation of thecomputing resources will have a direct

207、 impact on the inference delay.Therefore,it is necessary toconsider how to decide on the model split point,the computing resource allocation scheme on theterminal side and the base station side at the same time,and make timely adjustments as thenetwork condition changes,so as to ensure the continuou

208、s achievement of inference target.It canalso be considered to take advantage of a better decision-making scheme via reinforcementlearning,and the relevant design is shown in the following table.Table 6.3-1:Design of variables related to end-to-end collaborative inference schemebased on reinforcement

209、 learningvariablebase station sideterminal sidecondition(State)Theremainingallocatablecomputing power of various typesof computing resources in the basestation,the transmission bandwidthbetweencomputingunits(optional),theremainingallocatable uplink and downlink airinterfacechanneltransmissionresourc

210、es on the base station side,and the terminal uplink channelqualityTheremainingallocatablecomputingpower of various typesof computingresources in the terminal,the transmissionbandwidthbetweencomputingunits(optional),and the quality of the terminalsdownlink channelActionThe base station side is respon

211、siblefor calculating some modelparameters,uplink and downlinkbandwidth allocation,and basestation side computing resourceallocationThe terminal side is responsible for thecalculation of some model parameters andtheterminalsidecomputingresourceallocationRewardBase station side inference energyconsump

212、tion,etc.Inference performance indicators,terminalinference energy consumption,etc.Figure 6.3-2 shows the functional interactions for this mechanism.As mentioned in 5.3,anew control mechanism and data transmission protocol needs to be introduced for AI connection,which are represented by Dinfer_C an

213、d Dinfer_U respectively in the following figure.The controlentity Dinfer_C can dynamically decide on the model split point,the joint allocation of wirelessresources and computing resources according to the changes of radio resources and computingresources,and the data transmission entity Dinfer_U is

214、 responsible for calculating andtransmitting the intermediate results between the base station and the terminals.China Mobile6G native AI architecture and technical white paper26Figure 6.3-2:Functions and interactions of the end-to-end collaborative inference schemeExtended question：1.For AI service

215、s(such as inference)that require low latency and high reliability,how toprecisely allocate computing resources?Whether it is possible to model both heterogeneouscomputing resources and connection QoS to achieve optimal resource allocation?6.4 AI performance pre-validation based on digital twinA digi

216、tal twin network is a network composed of physical network entities and their digitaltwin entities,and real-time interaction can be performed between the physical entities and thedigital twin entities.The twin digital entities corresponding to the physical entities can beconstructed by data collecti

217、on and simulation.In this system,various network management andapplications can utilize the digital twin to efficiently analyze,diagnose,simulate and control thephysical entities based on data and models 67.The model validation in AI model life-cycle management utilizes the validation dataset tosele

218、ct the appropriate trained model,but the validation dataset and the training dataset are usuallydistributed in the same way.How to improve the generalization performance of the model in thedigital twin is one of the key technical problems to be solved.Generating samples in morescenarios in digital t

219、win can reduce the overhead of data collection,and meanwhile,improvegeneralization performance by data augmentation to diversify data samples.Conditional adversarial generative network(CGAN)22 can dynamically generateenvironmental models that conform to specific distributions due to dynamically chan

220、gingenvironmental conditions.The environmental models may include user distribution models,radiochannel models,user service models,network state models,network resource allocation models,etc.As shown in Figure 6.4-1,a CGAN is introduced into the physical network,and the randomsequence and certain se

221、mantic environmental conditions are used as input,and the Nashequilibrium is achieved through the adversarial training of the generation model and thediscriminative model.The generation model generated by the adversarial training is sent to thedigital twin,where more environmental data can be genera

222、ted by selectively changing theenvironmental conditions to conforms to specific distributions.China Mobile6G native AI architecture and technical white paper27Figure 6.4-1:Model performance pre-validation process based on digital twinIn addition to introducing a CGAN to generate diversified data sam

223、ples,in order to achievepre-validation of more scenarios,the reinforcement learning can be introduced to search datasample space.Considering that the state and action space may be very large,and the cost ofexploring all sample space is high.On one hand,a distributed search agent can be introduced to

224、speed up the search speed,and on the other hand,an efficient sample space search algorithm needsto be introduced,which may consider the feedback of the physical network.In addition,static orsemi-static environment models can be introduced in some dimensions to reduce the explorationcost of the sampl

225、e space.In addition,a two-way closed-loop optimization mechanism is introduced between thephysical network and the digital twin.The pre-validation applications in the digital twin collectabundant samples via the interaction with the pre-validation environment,and then train the AImodel to generate t

226、he required decision models for the physical network.In addition,the decisionmodel may generate deviations,and passes them to the digital twin for further correction in thepre-validation environment.sWhat needs to be further considered is that the GAN is currently mainly used in the field ofimage pr

227、ocessing,and the indicators of model convergence are used to measure the diversity andrestoration degree of image(such as IS,FID).Thus,it is necessary to design indicators suitable forthe distribution characteristics of mobile network samples.7 Summary and OutlookFrom 5G network intelligence practic

228、e,it is found that there are many shortcomings forexternal AI.However,since the 5G architecture,protocol and procedures have been finalized,wecan only make incremental improvements on the existing solutions.At the same time,in 6G era,intelligent scenarios will be more extensive.In addition to servin

229、g high-level autonomy andnetwork performance optimization itself,and providing the ultimate service experience tocustomers,it will also help the digital-intelligence transformation of thousands of industries.Theintelligent applications will be greatly enriched,and the demand for intelligent serviceC

230、hina Mobile6G native AI architecture and technical white paper28performance will be multi-dimensional.All of these require a paradigm shift from external AI tonative AI.The 6G native AI requires new concepts,new architectures,and new technologies.In termsof new concepts,this white paper proposes tha

231、t the 6G network have an evaluation system and aclosed-loop guarantee mechanism,i.e.,QoAIS.Under the guidance of QoAIS,multipledimensional resources,such as computing,data,algorithms,connections can be coordinated toserve full life cycle of AI workflows including data collection,data preprocessing,m

232、odel training,model inference,and model evaluation.At the same time,this white paper also proposes tointegrate native AI and digital twin,the AI models and workflows can be pre-verified in digitaltwin.In terms of new architecture,this white paper proposes that the 6G network will add new dataplanes

233、and smart planes,and meanwhile greatly expand the traditional control plane and userplane.Among them,the data plane will provide basic data transmission services for native AI anddigital twin,the intelligent plane will provide a complete operating environment for the full lifecycle of AI workflows,a

234、nd the extended control plane and user plane is proposed to deeplyintegrate communication and computing.In terms of new technologies,this white paper lists keytechnologies in model orchestration,training,inference,and the integration of native AI anddigital twin.At present,the industry has gradually

235、 reached a consensus on the requirements,concepts andscopes of native AI in 6G networks.The network architecture and key technologies of native AIare still under active research and discussion.Designing an intelligent native network architecturefor the rich intelligent applications in 6G era require

236、s not only a deep understanding of traditionalmobile communication networks,but also an accurate grasp of the requirements of variouspotential intelligent services in the future,as well as in-depth understanding of full life cycle AI.To this end,we jointly launched the 6GANA(6G Alliance of Network A

237、I)forum inDecember 2020 with 18 members from operators,equipment vendors,Internet service providers,and universities.6GANA is positioned as a global forum,focusing on the continuous explorationand promotion of 6G AI-related technologies,standardization and industrialization.It aims atconducting join

238、t research through the entire ecosystem,including ICT(such as chip manufacturers,network infrastructure providers,mobile network operators),vertical industries,AI serviceproviders,AI solution providers,AI academic and other stakeholders.A consensus has beenformed to promote AI to become a new capabi

239、lity and service for 6G networks 23.6GANA TG2is a working group under 6GANA responsible for the study of network architecture.It willidentify the basic technical requirements of 6G network native AI,study the key enablingtechnologies,its impact on 6G network architecture,and its impact on standardiz

240、ation,and buildan overall framework for 6G native AI network.Facing this goal,TG2 members havecomprehensively collected and summarized ten key technical issues that are widely discussed bythe industries in December 2021,and formed Ten Questions about 6G Native AI NetworkArchitecture.The content of t

241、his white paper provides reference answers to some of the keytechnical questions in Ten Questions about 6G Native AI Network Architecture.Finally,we propose that all partners in the industry chain work together to innovate,focusingon the following key technical issues to conduct in-depth research an

242、d extensive discussions:China Mobile6G native AI architecture and technical white paper29-What aspects will be reflected in the quality requirements(QoAIS)of AI services fordiversified intelligent applications in the future?What new evaluation dimensions willemerge compared to traditional QoS?How to

243、 support the generation and evaluation ofthe above indicators from the network architecture?-In order to ensure the continuous achievement of QoAIS,how to integrate or collaboratedifferent resource dimensions(data,computing power,algorithms,and connections)from the management plane,control plane,and

244、 user planes perspectives?-How to support the openness of internal network data from the network architecturelevel?How to support the on-demand dynamic extraction and processing of data bynative AI from the network architecture level?-Can we reuse the control and user plane protocols of traditional

245、communication for AIservices?What needs to be improved?-What is the relationship between native AI and digital twin?How to support the deepintegration of the two from the network architecture?-If a variety of native AI technologies is used without manual intervention and results innetwork problems,h

246、ow to trace and recover the problem?-How to ensure the credibility of AI model?If a model works well in evaluation stage,some problems to cause poor performance is encountered during inference stage,howto identify the problems in time,and how to handle exceptions?China Mobile6G native AI architectur

247、e and technical white paper30AbbreviationsAIArtificial IntelligenceMLMachine LearningQoSQuality of ServiceQoAISQuality of AI ServiceAIaaSAI as a ServiceNMSENormalized mean square errorKPIKey Performance IndicatorGPUGraphics Processing UnitNPUNeural-network Processing UnitDPUData Processing UnitTPUTe

248、nsor Processing UnitMECMobile Edge ComputingUPFUser Plane FunctionCPFControl Plane FunctionRRCRadio Resource ControlCRCComputing Resource ControlxRCx Resource ControlCPControl PlaneUPUser PlaneDtrain_CDistributed Training ControlPlane UnitDtrain_UDistributed Training UserPlane UnitCGANConditional Ge

249、nerative Adversarial NetworksAuthorsThis whitepaper is co-authored by：Future Lab,CMRI：Juan Deng,Gang Li,Qingbi Zheng,Zirui Wen,Chenkang Pan,Qixing Wang,Guangyi LiuAI Center,CMRI:Guangyu Li,Haitao Cai,Yanping Liang,Peng Zhao,Li YuChina Mobile6G native AI architecture and technical white paper31Refere

250、nce1劉光毅,鄧娟,鄭青碧,李剛,孫欣,黃宇紅.6G 智慧內生：技術挑戰、架構和關鍵特征J.移動通信,2021,45(04):68-78.2Wen C K,Shih W T,Jin S.Deep learning for massive MIMO CSI feedbackJ.IEEEWireless Communications Letters,2018,7(5):748-751.3Liu H,Wang Y,Fan W,et al.Trustworthy ai:A computational perspectiveJ.arXiv preprintarXiv:2107.06641,2021.3

251、GPP TS 38.323,“NR;Packet Data Convergence Protocol(PDCP)specification.”4Ashmore R,Calinescu R,Paterson C.Assuring the machine learning lifecycle:Desiderata,methods,and challengesJ.ACM Computing Surveys(CSUR),2021,54(5):1-39.5馬 洪 源,肖 子 玉,卜 忠 貴,趙 遠.5G 邊 緣 計 算 技 術 及 應 用 展 望 J.電 信 科學,2019,35(06):114-123

252、.6Deng J,Zheng Q,Liu G,et al.A Digital Twin Approach for Self-optimization of MobileNetworksC/2021 IEEE Wireless Communications and Networking Conference Workshops(WCNCW).IEEE,2021:1-6.7Liu G,Li N,Deng J,et al.6G Mobile Network Architecture-SOLIDS:Driving Forces,Features,and Functional TopologyJ.Eng

253、ineering,2021.8中 國 移 動:中 國 移 動 自 動 駕 駛 網 絡 白 皮 書 R/OL.(2021)2021-11-28.https:/ B,Qi P,Liu B,et al.Trustworthy AI:From Principles to PracticesJ.arXiv preprintarXiv:2110.01167,2021.10 Toreini E,Aitken M,Coopamootoo K,et al.The relationship between trust in AI andtrustworthy machine learning technologi

254、esC/Proceedings of the 2020 conference onfairness,accountability,and transparency.2020:272-28311 Choromanska A,HenaffM,MathieuM,etal.ThelosssurfacesofmultilayernetworksC/Artificial intelligence and statistics.PMLR,2015:192-204.12 Finn C,Abbeel P,Levine S.Model-agnostic meta-learning for fast adaptat

255、ion of deepnetworksC/International Conference on Machine Learning.PMLR,2017:1126-1135.13 Shukur H,Zeebaree S R M,Ahmed A J,et al.A state of art survey for concurrentcomputation and clustering of parallel computing for distributed systemsJ.Journal ofApplied Science and Technology Trends,2020,1(4):148

256、-154.14 Liu L,Zhang J,Song S H,et al.Client-edge-cloud hierarchical federated learningC/ICC2020-2020 IEEE International Conference on Communications(ICC).IEEE,2020:1-6.15 Warnat-Herresthal S,Schultze H,Shastry K L,et al.Swarm Learning for decentralized andconfidential clinical machine learningJ.Natu

257、re,2021,594(7862):265-270.16 Xu X,Li R,Zhao Z,et al.Stigmergic Independent Reinforcement Learning for MultiagentCollaborationJ.IEEE Transactions on Neural Networks and Learning Systems,2021.17 Shao J,Zhang J.Bottlenet+:An end-to-end approach for feature compression indevice-edgeco-inferencesystemsC/

258、2020IEEEInternationalConferenceonCommunications Workshops(ICC Workshops).IEEE,2020:1-6.China Mobile6G native AI architecture and technical white paper3218 Teerapittayanon S,McDanel B,Kung H T.Distributed deep neural networks over the cloud,the edge and end devicesC/2017 IEEE 37th International Confe

259、rence on DistributedComputing Systems(ICDCS).IEEE,2017:328-339.19 Li T,Sahu A K,Zaheer M,et al.Feddane:A federated newton-type methodC/2019 53rdAsilomar Conference on Signals,Systems,and Computers.IEEE,2019:1227-1231.20 Hua S,Yang K,Shi Y.On-device federated learning via second-order optimization wi

260、thover-the-aircomputationC/2019IEEE90thVehicularTechnologyConference(VTC2019-Fall).IEEE,2019:1-5.21 Ghosh A,Maity R K,Mazumdar A,et al.Communication efficient distributed approximateNewton methodC/2020 IEEE International Symposium on Information Theory(ISIT).IEEE,2020:2539-2544.22 MirzaM,OsinderoS.ConditionalgenerativeadversarialnetsJ.arXivpreprintarXiv:1411.1784,2014.23 6GANAWhitepaperChina Mobile6G native AI architecture and technical white paper33Digital Twin;Ubiquitous Intelligence