《用于实时数据驱动型网络和物联网洞察的下一代人工智能.pdf》由会员分享,可在线阅读,更多相关《用于实时数据驱动型网络和物联网洞察的下一代人工智能.pdf(65页珍藏版)》请在三个皮匠报告上搜索。
1、#CiscoLive#CiscoLiveHugo Latapie,Principal EngineerBRKETI-1000Next Generation AI for Real-Time Data-Driven Network and IoT Insights 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveEnter your personal notes hereCisco Webex App 3Questions?Use Cisco Webex App to chat with the
2、 speaker after the sessionFind this session in the Cisco Live Mobile AppClick“Join the Discussion”Install the Webex App or go directly to the Webex spaceEnter messages/questions in the Webex spaceHowWebex spaces will be moderated by the speaker until June 9,2023.12343https:/ 2023 Cisco and/or its af
3、filiates.All rights reserved.Cisco PublicBRKETI-1000Agenda 2023 Cisco and/or its affiliates.All rights reserved.Cisco PublicIntroduction to Hybrid AIOverview of Hybrid AINetworking ApplicationsVideo ApplicationsIntroduction of Open-source project Deep Vision which utilizes self-supervised learning f
4、or video analysisConclusionQ&A sessionBRKETI-10004 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveWhat is Hybrid AI?Lets discuss how a picture like this processed via deep learning versus hybrid AI.BRKETI-10005 2023 Cisco and/or its affiliates.All rights reserved.Cisco Pu
5、blic#CiscoLiveObject Detection on COCO test-dev6BRKETI-1000 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLive7BRKETI-1000Training usually takes around 100K labeled images per class.Deep learning does a good job learning a hierarchical representation of the image 2023 Cisco
6、and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveHow Deep Learning FailsAs this saliency map shows,deep learning algorithms will make the final panda/no-panda decision based on just the few pixels highlighted here?BRKETI-10008 2023 Cisco and/or its affiliates.All rights reserved.Cisco
7、 Public#CiscoLiveAI 2.0 ApproachSensorDog Agent“panda”“panda”95%confidence 95%confidence Because:appearance,behavior,Because:appearance,behavior,location,location,EyeEyeSnoutSnoutBranchBranchTreeTreeBlackBlackWhiteWhiteColorColorFaceFaceisisisishas has GPSGPSBeijingBeijingBjzooBjzooGiant PandaGiant
8、Pandahas in famous for climbsMammalMammal39.939012,116.339539SensorDogSensorDog vs DLvs DLBRKETI-10009 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveKnowledge RepresentationCritically important to distinguish between distinctions and similarities.Amoeba Amoeba Amoeba Pre
9、datorPredatorPredatorFoodFoodFoodBRKETI-100010 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveSelf-wonder?PandaPanda!MammalMammalTreeTreeGibbonGibbonGiant PandaGiant PandaBeijingBeijingColorColorWhiteWhiteBlackBlackBjzooBjzooGPSGPSBranchBranchSnoutSnoutEyeEyePanda Image+N
10、oise39.939012,116.339539Self-wonder?GibbonGibbon!Less Data&More KnowledgeIncreased Abstraction Experience,memory&ontologies,contextualizationBRKETI-100011 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveSummary of problemNeed model of the world Need concept of experience b
11、ased truth Knowledge representation should cover abstractionNeed self-supervised cumulative learningNeed machine generated ontologiesBRKETI-100012 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveWhat about large language models like ChatGPT?13BRKETI-1000Hybrid AI 2023 Cisc
12、o and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveHybrid AI Overview Neurosymbolic ApproachLeveraging efficient learning by reasoningConceptual worldConceptual worldIdeas,maps,concepts,metaconcepts,Meta-metaconcepts,Physical worldPhysical worldsensor data,text,video,Cognitive Cogniti
13、ve SynergySynergyUnsupervised LearningUnsupervised LearningSymbolic LearningSubsSubsymbolicymbolic LearningLearningL L2.1 2.1 -L L2.2.L L1 1L L0 0Subsymbolic LearningWith a model of knowledgemodel of knowledge supporting abstractionabstraction,distinctionsdistinctions,similaritiessimilarities,struct
14、urestructureBRKETI-100015 2023 Cisco and/or its affiliates.All rights reserved.Cisco PublicBRKETI-100016 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveReferencesBlogs&NewsAI for Traffic Monitoring in Sydney Australia,https:/news- AGI Matters Today,https:/ Innovation Habi
15、t,https:/ Beyond Deep Learning,https:/ Metamodel and Framework For AGI(2020),https:/arxiv.org/pdf/2008.12879.pdfA reasoning-based model for anomaly detection in the Smart City domain(2019),https:/cis.temple.edu/tagit/events/papers/Hammer.pdfKeynotesShare Your Science:Bringing Intelligence to the Net
16、work with High Performance Computing,NVIDIA Developer,2016,https:/ by Reasoning-Smart Cities,AGI Conference 2019,https:/ AGI in the Wild,AGI Conference 2020,https:/ 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveBRKETI-100019Kronos 2023 Cisco and/or its affiliates.All rig
17、hts reserved.Cisco Public#CiscoLiveBRKETI-100021Sensordog 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveBRKETI-100023Aggression Model Demo 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveTracking based velocity and acceleration analysis25 keypo
18、ints velocity+25 keypoints accelerationAggression modelSelf-supervised learningtsTrackerBRKETI-100026Aggression Model DemoSelf supervised human behavior learning and anomaly detection 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveSelf-supervised LearningSensor Dog can le
19、arn normal and abnormal behaviors by analyzing the peoples behaviors in a scene.It uses time series analysis techniques,not an ML/DL model or computer vision algorithms to learn normal/abnormal behaviors.Why?There too many behaviors to learn:fall,fight,heart attack,theft Occlusions can cause a track
20、er to lose an object;e.g.,two people fighting.BRKETI-100029 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveSelf-supervised LearningTracklet:A set of bounding boxes that are associated via tracking the same object over time.If a tracklet changes its behavior,its called a r
21、egime change.The behavior can be based on a tracklets speed,coordinates,or width/height ratio,etcBRKETI-100030Regime changes in abnormal cases become erraticAfter the student falls,self-supervised module picks up abnormal number of regime changes in other students tracklets.Regime changes in abnorma
22、l cases become erraticAfter the fight starts,self-supervised module picks up abnormal number of regime changes for other people in the scene.People as a Sensor 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveBRKETI-100034Cisco Deep Vision Open Source Project 2023 Cisco and
23、/or its affiliates.All rights reserved.Cisco Public#CiscoLiveCisco Deep Vision targets several key problems including improving analytics accuracy on spatio-temporal data,self-supervised learning,lifelong learning and creating explainable models.Deep Vision is a serverless edge platform for explaina
24、ble perception challenges that is focused on enabling the development and deployment of new computer vision and multi-modal spatio-temporal algorithms.SummaryBRKETI-100036 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveWhat is Cisco Deep Vision?Explainable perception star
25、ts with a 4D space-time assumption about the worldLeverage algorithms for identifying interesting,surprising,and anomalous eventsCategorize behaviors,events,and objects via similarity metrics and reasoning.Focusing on making sense of perception is the metalearning objective which enables self-superv
26、ised learning.Supports cumulative and generational learning via machine generated ontologies.BRKETI-100037 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveFocusing on making sense of perception is the key to self-supervised learning.By understanding the world around them,a
27、lgorithms can learn without the need for labels or ground truth.BRKETI-100038 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveoObjective is the development of trustworthy computer vision systems oScalable and modular serverless open-source framework for bringing together:S
28、tate-of-the-art object detectors,trackers,behavior detectors,which are changing daily New types of reasoning engines 3D semantics projects Multi-modal data stream analytics Neurosymbolic integration projects Knowledge representation projects Self supervised learning projects Hyperdimensional Computi
29、ng oReasoning on machine generated ontologies leveraging explainable perceptionoExplainable perception challengesBRKETI-100039 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveBRKETI-100040 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveVideo/Sen
30、sorData/TelemetryServerless Workflow EngineBRKETI-100041 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveBRKETI-100042 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveExplainable Perception Challenges Scenario is a typical multi-camera situation
31、where ingress/egress has camera coverage and goal is to obtain counts for on-premises people/vehicles/.Sparse multi-camera object counting challenge Scenarios where there are key moving objects in the scene with no prior trained model.example a flying drone Objective is to be able to provide analyti
32、cs for all moving objects in sceneDeep learning blindness challenge Categorize behaviors into top 10 types of behaviors Identify behaviors of interestSelf-supervised learning behavioral analytics challengesBRKETI-100043 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveUnive
33、rsity ResearchVQPy3D monocular detection/trackingVibration analyticsInteractive Video ExplorationMulti-modal computer visionHyperdimensional Computing Reasoning Engines:OpenNARS,ONA,AERA,and HyperonSemantic SlamBRKETI-100044 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLive
34、Key Takeaways Current AI models are advancing at an amazing rateHybrid AI can help build trustworthy mission critical systemsWide range of use cases from network insights to human behavior analyticsBRKETI-100045 2021 Cisco and/or its affiliates.All rights reserved.Cisco Confidential46 2021 Cisco and
35、/or its affiliates.All rights reserved.Cisco ConfidentialOur Mission Statement:Create good data for the best possible training set across data modalities by focusing on 3 key steps1.Make raw data searchable2.Support signal detection and model intent formation3.Drive accurate inferenceVision Works:Pr
36、oduct OverviewOpen-source augmentationAuto-labeling engineSynthetic data generationKey Product FeaturesKey Product FeaturesIntent searchCustomize parametersCollaborationFeature ParityDifferentiationABCDEFCore valueCore Value:AutoCore Value:Auto-Labeling Engine Labeling Engine ComponentsComponentsDat
37、a explorationData explorationSpeed to modelSpeed to modelData prep Data prep accuracyaccuracyBenefit:DiscoveryVenture into unknown,Venture into unknown,new spaces which are prohibitively expensive to explore today,or assess new business problems.Component:Multi-modal embeddings drives clustering for
38、 the unknowns(i.e.,unknown objects,behaviors,and events)Benefit:SpeedAt least 5 times faster At least 5 times faster than todays labeling approaches,with potential for even speedier iteration cycles.Component:Auto-labeling engine i.e.,Automatic“interesting”clip identification,Automatic cluster detec
39、tion and label propagationBenefit:AccuracySimilar to todays levelsSimilar to todays levelsi.e.,Approach human levels of accuracy with average of 3 humans across data in bulk.Component:Integrating symbolic and sub-symbolic spaces allows us to reason and focus on the smallest amount of relevant inform
40、ationDo you want to supercharge your data exploration,insights,and model preparation process?Come talk to us!2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveHow to learn moreFollow to stay informed on latest developments including blogs,papers,open source,and more.Schedule
41、 time in the Whisper Suites to discuss Edge AIBRKETI-100049Cisco Research Overview 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveGoal:To conduct and foster researchin technology areas of strategic interest to Cisco and generate business,technology and societal impactBRKE
42、TI-100051 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveSponsored Research Partners100+Universities 25+CountriesUniversity GiftProgramMIT MLAMIT MediaLabStanford ICMECMU CylabPurdue CERIASQEDCCQNUIDPUniversity/Industry ConsortiumsUniversity EngagementsBRKETI-100052 2023
43、Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveEthical AI(5)Ethical AI(5)Supply Chain(4)Supply Chain(4)Bias detection/mitigation,ethical design,privacy-preserving AI/ML,AI for ethicsFuture of Work Future of Work(10)(10)Productivity,worker wellness,smart homeEdge Computing(18)E
44、dge Computing(18)Infrastructure,federated/distributed ML,CAVs,MLOps,serverless,5GAnomaly prediction,supply-demand planning,RFID-based tracking,securityHealthcare(13)Healthcare(13)AI/ML for diseases,federated learning,mental healthcare,radiology,remote health monitoringSecurity(27)Security(27)Malware
45、,pen-testing,privacy-preserving computation,SW supply chain,biometricsOthers(19)Others(19)Storage switch,network verification,indoor localization,sustainability,hardwareAI/ML/CV(13)AI/ML/CV(13)Scene prediction,multimodal sensing,image reconstruction,AIOpsNLP(11)NLP(11)NL understanding,language model
46、s,text summarizationFuture DirectionsFuture DirectionsData Management,Data Processing,Sustainability,Distributed Systems,Edge Computing,Metaverse,Blockchain,Security,AI/ML,Networking,Cloud120 Total Projects FundedBRKETI-100053 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLi
47、veOps TeamResearch TeamUniversity Research EngagementsInternal Research ProjectsThought LeadershipRamana KompellaHead of Cisco ResearchBRKETI-100054 2023 Cisco and/or its affiliates.All rights reserved.Cisco PublicVisit Outshift in the World of Solutions!Take a picture of this slide and bring it to
48、the Outshift booth in the World of Solutions.(#3307)Get your badge scanned to be entered into our daily drawing for an Apple iPad!Explore Explore 55Session ID 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveLearn more about ET&I Cloud-Native ProductsPanopticaSimplified Clo
49、ud-Native Application Security for DevSecOps,Platform,and DevOps teams.CalistiCisco Service Mesh Manager.An enterprise-ready Istio platform for DevOps and SREs that automates lifecycle management and simplifies connectivity,security,and observability for microservice-based applications.See them both
50、 in the Cisco ShowcaseApplication Performance,AppDynamics,Full Stack ObservabilityThursday,Feb 92:15 PM-3:15 PM CETWayne BrownSE Manager,AMER Elite Partners,Cisco Systems,Inc.LocationTBUOpenTelemetry is quickly becoming the de-facto vendor-neutral standard for collecting metrics,events,logs,and trac
51、es from a wide range of applications and systems.In this session,we will learn the history behind OpenTelemetry,the benefits that OpenTelemetry provides,the architecture of OpenTelemetry,and the future of the standard.Finally,we will talk about how AppDynamics is embracing the OpenTelemetry standard
52、 and supporting its developments.Do Tell About OTel:An Introduction to OpenTelemetry and How AppDynamics is Embracing It BRKAPP-1154More Learning OpportunitiesBRKETI-100056 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveMore Learning OpportunitiesCloudThursday Feb 9,2:30P
53、M 3:30PM CETGiles HeronPrincipal Engineer,Cisco SystemsLocationSession Room G102The cloud-native approach based on containerised micro-services has transformed the software landscape but is largely focussed on non-real time web-based applications,especially in the case of service meshes which use we
54、b proxies to interconnect workloads.Media Streaming Mesh uses real-time media proxies to observe,route,encrypt and protect north-south and east-west media traffic.Youll leave this session with an understanding of the Media Streaming Mesh architecture,of some of its key use cases,and how you can appl
55、y it to your own media workflows.Real-Time Media in a Cloud Native WorldBRKETI-2006Hybrid Cloud,CloudThursday Feb 9,4:00PM 5:00PM CETPeter BoschDistinguished Engineer,Cisco SystemsLocationSession Room G109Application data represents the core to digital enterprises:data is handled by applications tha
56、t are hosted on-prem,in-cloud,in containers and virtual machines,by API gateways,stored on in-VM databases and cloud storage resources.Losing,not knowing where it resides,not knowing if data is handled in a compliant manner,or not knowing someone copied the data can be a disaster to a digital enterp
57、rise.This talk presents Ciscos approach to data security and compliance.We present how Cisco tracks data in-flight and at-rest,and how to turn such data into information.We show where,in the application,information is vulnerable and the application is not compliant to the various compliancy standard
58、s.Data Security and Compliance in Cloud Native andOn-Prem ApplicationsBRKETI-2414Hybrid Cloud,CloudFriday Feb 10,9:15AM 10:45AM CETShannon McFarlandDistinguished Engineer,Cisco SystemsLocationSession Room E105The need to have an end-to-end view of a complex microservices environment and the underlyi
59、ng infrastructure is growing in demand as customers migrate from legacy and monolithic workloads to Cloud Native environments.The days of scouring through endless logs and trying to correlate those logs with events,metrics,and even traces are coming to an end.Just seeing the raw info does little to
60、help one understand what is happening.Observability is so much more than seeing the data but making sense of it and taking prescription actions based on the data.This session will go over common MELT(metrics,events,logging,and tracing)Cloud Native ObservabilityBRKCLD-2158BRKETI-100057 2023 Cisco and
61、/or its affiliates.All rights reserved.Cisco PublicComplete your Session SurveyPlease complete your session survey after each session.Your feedback is important.Complete a minimum of 4 session surveys and the Overall Conference survey(open from Thursday)to receive your Cisco Live t-shirt.All surveys
62、 can be taken in the Cisco Events Mobile App or by logging in to the Session Catalog and clicking the Attendee Dashboard”athttps:/ 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveFrom technology training and team development to Cisco certifications and learning plans,let u
63、s help you empower your business and Learning and CertificationsHere at the event?Visit us at The Learning and Certifications lounge at the World of SolutionsPay for Learning with Pay for Learning with Cisco Learning Credits Cisco Learning Credits(CLCs)are prepaid training vouchers redeemed directly
64、 with Cisco.Cisco Training BootcampsIntensive team&individual automation and technology training programsCisco Learning Partner ProgramAuthorized training partners supporting Cisco technology and career certificationsCisco Instructor-led and Virtual Instructor-led trainingAccelerated curriculum of p
65、roduct,technology,and certification coursesCisco Certifications and Specialist CertificationsAward-winning certification program empowers students and IT Professionals to advance their technical careersCisco Guided Study Groups180-day certification prep program with learning and supportCisco Continu
66、ing Education ProgramRecertification training options for Cisco certified individualsLearnCisco U.IT learning hub that guides teams and learners toward their goalsCisco Digital LearningSubscription-based product,technology,and certification trainingCisco Modeling LabsNetwork simulation platform for
67、design,testing,and troubleshootingCisco Learning Network Resource community portal for certifications and learningTrainCertifyBRKETI-100059#CiscoLive 2023 Cisco and/or its affiliates.All rights reserved.Cisco PublicAgenda60Visit the On-Demand Library for more sessions at any of the related sessions
68、at the DevNet,Capture the Flag,and Walk-in Labs zones.Visit the Cisco Showcase for related demos.Book your one-on-one Meet the Engineer meeting.Continue Your EducationBRKETI-1000 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLiveFill out your session surveys!Attendees who fi
69、ll out a minimum of four session surveys and the overall event survey will get Cisco Live-branded socks(while supplies last)!61Session IDThese points help you get on the leaderboard and increase your chances of winning daily and grand prizesAttendees will also earn 100 points in theCisco Live Challe
70、nge for every survey completed.2023 Cisco and/or its affiliates.All rights reserved.Cisco PublicContinue your educationVisit the Cisco Showcase for related demosBook your one-on-oneMeet the Engineer meetingAttend the interactive education with DevNet,Capture the Flag,and Walk-in LabsVisit the On-Dem
71、and Library for more sessions at www.CiscoL IDThank you#CiscoLive 2023 Cisco and/or its affiliates.All rights reserved.Cisco Public#CiscoLive64Gamify your Cisco Live experience!Get points Get points for attending this session!for attending this session!Open the Cisco Events App.Click on Cisco Live Challenge in the side menu.Click on View Your Badges at the top.Click the+at the bottom of the screen and scan the QR code:How:123464 2023 Cisco and/or its affiliates.All rights reserved.Cisco PublicBRKETI-1000#CiscoLive