Artificial Intelligence, Machine Learning, Deep Learning .
Artificial Intelligence, MachineLearning, Deep Learning, Cisco?Tomáš OndovčíkSystems Architect29.10.2019
Part 1:Introduction toAI / Machine Learning, 2019 Cisco and/or its affiliates. All rights reserved.
An Introduction to ML/AILip ReadingDeep BlueWatsonCounting peopleRecognitionNLP TranslationDeep MindSelf-drivingPlaying MusicWaveNet (TTS)NetworksAnd Many more: CRM, Healthcare, Personal Assistants, . 2019 Cisco and/or its affiliates. All rights reserved.
McCullogh & Pitts - 1943Von Neumann used this logicwhen designing the“universal computer”Binary threshold neuronIf sum threshold, output 1Output 0 otherwiseWeighted inputs 2019 Cisco and/or its affiliates. All rights reserved.
In the 1950s and 1960s,Principles of Neurodynamics were examinedand Symbolic ML expandedWhen I show theseshapes to the cameraThis IBM 704computer can say “it’sa triangle” 2019 Cisco and/or its affiliates. All rights reserved.
Why Now?Computing powerData “tsunami”Solve above equation for coefficients:Insatiable desire for efficiency / productivity 2019 Cisco and/or its affiliates. All rights reserved.Availability of tools
The Emergence of Large Data Sets and MLData is the rocket fuel of Machine Learning – Andrew Ng Open data sets have been a crucial factor in thesuccess of ML http://yann.lecun.com/exdb/mnist/ http://image-net.org/ https://bis.lexisnexis.com https://catalog.data.gov/dataset Allows for direct comparison of learning andinference algorithmsThe result has been improvement of error rates Video analytics (facial recognition) NLP/Voice recognition 2019 Cisco and/or its affiliates. All rights reserved.Graph courtesy Andrew Ng
MNIST – Handwritten Digit Data Set Training set of 60,000 examples (6000 per digit) Test set of 10,000 examples (1000 per digit) Each character is a 28x28 pixel box 2019 Cisco and/or its affiliates. All rights reserved.
Understanding the RolesData AcquisitionDataScientistData ManipulationAnalytics / BI SoftwarePackagesData MovementSelection of the Rightanalytics ToolsMachine Learning MethodsSQL DevelopersConnecting Data Insights tothe BusinessVisual Analytics ToolsUsers of Data Mining ToolsNetwork Engineers build the platform supporting the business 2019 Cisco and/or its affiliates. All rights reserved.DataAnalyst
Artificial Intelligence, Machine Learning, and DeepLearning (AI/ML/DL)Artificial IntelligenceTechnique where computer canmimic human behaviorMachine LearningSubset of AI techniques which usealgorithms to enable machines to learn fromdataDeep LearningSubset of ML techniques which uses multilayer neural network to learn 2019 Cisco and/or its affiliates. All rights DeepLearning
It is a Complex LandscapeSupervisedLearningDevelop gDiscover patterns and anomaliesbased on input data 2019 Cisco and/or its affiliates. All rights MatrixNeuralNetworksHiddenMarkov
Machine Learning: The Main MethodsSupervised LearningLearning with a labeled training setimage by Jeff DeanUnsupervised LearningDiscovering patterns in unlabeled dataInference (Statistical Learning)How confident are you in the result? 2019 Cisco and/or its affiliates. All rights reserved.
Reinforcement Learning A semi-supervised learning model No training data or correct/incorrect guidance needs to be given. Involves behavioural psychology In summary, a lot of trial and errorAgentState StAction AtReward RtEnvironment 2019 Cisco and/or its affiliates. All rights reserved.
Supervised Learning Part 1Linear Regression 2019 Cisco and/or its affiliates. All rights reserved.
Supervised Learning: Linear Regression (Single Variable)Data sample from across company“How much is ourheadcount going tocost us in 5 years?” 2019 Cisco and/or its affiliates. All rights reserved.
Supervised Learning: Linear Regression (Single Variable)Bad fitGood fit“What makes one fit bad and another good?” 2019 Cisco and/or its affiliates. All rights reserved.
How Close Is Our Prediction to the Dataset?The Cost Function𝐽 (𝜃0 , 𝜃1 ) 1𝑚2𝑚𝑖 1(𝜃0 𝜃1 𝑥𝑖 𝑦𝑖 )2How far am I from the real y, if I use my random 𝜃0 and 𝜃1 and do 𝜃0 𝜃1x 2019 Cisco and/or its affiliates. All rights reserved.
Gradient DescentGetting Closer to the Answer 2019 Cisco and/or its affiliates. All rights reserved.
Gradient DescentGetting Closer to the Answer 2019 Cisco and/or its affiliates. All rights reserved.
Gradient DescentGetting Closer to the Answer 2019 Cisco and/or its affiliates. All rights reserved.
Gradient DescentGetting Closer to the Answer 2019 Cisco and/or its affiliates. All rights reserved.
Now we can Predict a Future EventNow I can make some predictions:Salary 𝜃0 𝜃1 Years“If a new person joined my team andthey have been with the company 9years, their salary will likely be:” 54.6k 2.35k x 9 75.75k 2019 Cisco and/or its affiliates. All rights reserved.
Expanding to Many More DimensionsAnd instead of 2 dimensions (x,y),You might have 1 million dimensions!(But the idea is the same) 2019 Cisco and/or its affiliates. All rights reserved.
Supervised Learning Part 2Logistic Regression(Classification) 2019 Cisco and/or its affiliates. All rights reserved.
Supervised Learning: Classification With linear regression you try to find a number(y) that predicts an outcome. There is also classification, where your lineseparates two groups E.g. Is a credit card transaction fraudulent or is itsafe? Also known as Logistic Regression 2019 Cisco and/or its affiliates. All rights reserved.
Fitting the Data The main challenge in Supervised Learningis to find the right equation and figure outif the samples represent the full population 2019 Cisco and/or its affiliates. All rights reserved.
Regression Summary: Learn, then PredictAlgorithmModelPredictionModelTraining DataLearning (controlled experiments) 2019 Cisco and/or its affiliates. All rights reserved.New DataPredict (AI in the wild)
How Accurate is Your Model?Measure Your Accuracy!Your DataModelAccuracyTraining Data99.3%Test Data98.2% 2019 Cisco and/or its affiliates. All rights reserved.
A Typical Supervised Learning ML Work nModelDeployTestModelAdjust ModelParameters 2019 Cisco and/or its affiliates. All rights reserved.EvaluateModel
Unsupervised Learning 2019 Cisco and/or its affiliates. All rights reserved.
Unsupervised Learning You do not know the right answer, and there is toomuch data for you to guess Example: you manufacture small engines Some of them will fail You want to spot the failures before they get installed onmowers, chainsaws, etc. How do you do that? 2019 Cisco and/or its affiliates. All rights reserved.
Unsupervised Learning Your engine will group engines that4 groups 4 engine categorieshave similar characteristics. In math, this is simply groupingpoints that are close to one another And will spot the outliers Those are the engines likely to fail(different from the others)Outlier 2019 Cisco and/or its affiliates. All rights reserved.
Unsupervised LearningThe math can take may forms,but a common form is K-meansWant 3 groups?Take 3 random points (3‘centroids’, 𝜇𝑐 𝑖 )Then take a known point,calculate which centroid isclosest 2019 Cisco and/or its affiliates. All rights reserved.
Unsupervised LearningK-meansWant 3 groups?Repeat for all points. That’s ourusual distance equation:𝐾𝑖 𝜇2min 𝑥 𝑖𝑖 1𝑐(𝑖)𝑐 2019 Cisco and/or its affiliates. All rights reserved.
Unsupervised LearningThen you move your ‘centroids’,𝜇𝑐 𝑖 to the center (mean x,y) ofeach group you formed 2019 Cisco and/or its affiliates. All rights reserved.
Unsupervised LearningAnd you repeat.You are done when no pointjumps to another group anymore 2019 Cisco and/or its affiliates. All rights reserved.
Unsupervised LearningAnd you repeat.You are done when no pointjumps to another group anymore 2019 Cisco and/or its affiliates. All rights reserved.
Unsupervised LearningAnd you repeat.You are done when no pointjumps to another group anymore 2019 Cisco and/or its affiliates. All rights reserved.
Unsupervised LearningAnd you repeat.You are done when no pointjumps to another group anymore 2019 Cisco and/or its affiliates. All rights reserved.
Neural Networks 2019 Cisco and/or its affiliates. All rights reserved.
Neural Networks: From Neurons to Perceptrons1. Inputs arrive from other neurons /nodes.Out putsignalsAct ivat ion(t hreshold)2. Receiving neuron / node processesthe inputs weighted sum.Signal (impulse) propagat ionInput signals(form ot her axon t erminas)x1x2w1w2s(z)Xz wi x in.xn3. Receiving neuron / node applies anactivation function (sigmoid).Weight ed sumwn 2019 Cisco and/or its affiliates. All rights reserved.Act ivat ionfunct ionOut put4. Output of activation function passedto subsequent neurons / nodes.
Neural Networks Then you put tons of units,possibly in multiple layers (inthis case, it is called deeplearning)Also called Artificial NeuralNetworks (ANNs)Convolutional Neural Networks(CNNs) are a common subclass of ANNs 2019 Cisco and/or its affiliates. All rights reserved.
Neural Networks: Image RecognitionTraining ImagesBackpropagation (just 10 years ago by Geoffry Hinton)Adjust weightsEventually, you trainmodel to recognize “car” 2019 Cisco and/or its affiliates. All rights reserved.
Types of Neural Networks The way you connect the unitscan vary immenselyAnd this is what makes thisfamily very richTons of possible applicationsdepending on what data you arelooking at, and what you try tofind 2019 Cisco and/or its affiliates. All rights reserved.
Part 2:The AI/ML Landscape at CiscoCollabIoTAISecUCS 2019 Cisco and/or its affiliates. All rights reserved.DNA
How Cisco Approaches AI/MLIntersightConsumptionProducts use AI/ML to do things better 2019 Cisco and/or its affiliates. All rights reserved.EnablementInfrastructure Supporting AI/ML workloads
AI/ML By Product CategoryReinvent theNetworkSecurity isFoundationalPower aMulticloud WorldUnlock thePower of DataCreate ompanyETACloudlockIntersightKineticCAMNetwork EarlyWarningCognitive dStealthwatchTalent TrendsTalosWebexUmbrella 2019 Cisco and/or its affiliates. All rights reserved.
AI/ML At Cisco:The Data Center 2019 Cisco and/or its affiliates. All rights reserved.
The Power of GPUs for Deep Learning Graphical Processing Units are specializedtypes of electronic circuitry designed to rapidlymanipulate memory for graphics GPUs support parallel processing, acceleratingtheir ability to execute algorithms that requireparallel processes GPUs are at the heart of deep learning andneural networks 2019 Cisco and/or its affiliates. All rights reserved.
Comparing CPUs and GPUs CPUs are capable of almost anytask – but at a priceCPUs are like a swiss army knife 2019 Cisco and/or its affiliates. All rights reserved. GPUs are highly-specializedprocessors used to solvecomplex math problemsGPUs are like specialized surgicalinstruments
Cisco AI/ML – Compute PortfolioTesting and development,and model trainingCisco UCS C240Cisco HyperFlex 2402 x P100/ V1002 x P100/V100 per nodeDeep learning/trainingCisco UCS C480Cisco UCS C480 ML6 x PCIe P100/ V1008x V100 with NVLinkInferencingC/HX 220C/HX 240Option of GPU-only nodesUnified managementCisco IMCXML APISimplified management, customer choice, Cisco Validated Design 2019 Cisco and/or its affiliates. All rights reserved.2 x P46 x P4
Tying it Together: Big Data with Machine LearningCisco Validated Hadoop Design with Cloudera on GPU-Powered AI/ML Workloads2 x UCS 6332Fabric Interconnect16 x UCS C240 819202122232425262728293031Popular ML/DL frameworks4232123BCN4STSENV41LSL1L2CISCO 2526272829303132123BCN404STSENVLSL1L2CISCO D ReWritable800 GB800 GB2 TB800 GB800 GB800 GB800 GB2 TB800 GB800 00NVMEHWH800NVME SSDNVME SSDSATA HDDSATA HDDSATA HDDSATA HDDNVME SSDNVME SSDNVME SSDNVME SSDSATA HDDSATA HDDSATA HDDSATA HDDNVME SSDNVME SSD2 TB2 TB2 TB2 TB2 TB2 TBMU L T IRECO RDERReW ritabl eUCSC480 0707071211X10DVD ReWritable09C240 M5datanodes800 GB800 GB2 TB800 GB800 GB800 GB800 GB2 TB800 GB800 00NVMEHWH800NVME SSDNVME SSDSATA HDDSATA HDDSATA HDDSATA HDDNVME SSDNVME SSDNVME SSDNVME SSDSATA HDDSATA HDDSATA HDDSATA HDDNVME SSDNVME SSD2 TB2 TB2 TB2 TB2 TB2 TBMU L T IRECO RDERReW ritabl eUCSC480 M509Cisco UCS C480 ML M5 each withCompute:8x V100 Nvidia Tesla GPU2x6132 CPUS060605050504040407XX0608XXXXXXXXXXXXXXXX06DVD ReWritable800 GB800 GB2 TB800 GB800 GB800 GB800 GB2 TB800 GB800 00NVMEHWH800NVME SSDNVME SSDSATA HDDSATA HDDSATA HDDSATA HDDNVME SSDNVME SSDNVME SSDNVME SSDSATA HDDSATA HDDSATA HDDSATA HDDNVME SSDNVME SSD2 TB2 TB2 TB2 TB2 TB2 TBMU L T IRECO RDERReW ritabl eUCSC480 M5S0504X03030303XX020202010101XXXXXXXXXXXXXXX02DVD ReWritableCisco Cloudera CVD:https://www.cisco.com/c/en/us/td/docs/unified computing/ucs/UCS CVDs/Cisco UCS Integrated Infrastructure for Big Data with Cloudera 28node.html 2019 Cisco and/or its affiliates. All rights reserved.800 GB800 GB2 TB800 GB800 GB800 GB800 GB2 TB800 GB800 00NVMEHWH800NVME SSDNVME SSDSATA HDDSATA HDDSATA HDDSATA HDDNVME SSDNVME SSDNVME SSDNVME SSDSATA HDDSATA HDDSATA HDDSATA HDDNVME SSDNVME SSD2 TB2 TB2 TB2 TB2 TB2 TBMU L T IRECO RDERReW ritabl eUCSC480 M501Nvidia CUDA containersorchestrated with YARN for DeepLearning
iCAMIntelligent Comprehensive Analytics and Machine Learning on Nexus switchesiCAM Overview: Analytics & Telemetry, natively on theswitch/router Security access control analytics Internal hardware tables usage analytics Top/bottom heavy hitters Anomaly visualization Build apps on top of iCAM Historical Analytics Predictive Analysis Streaming telemetryBenefits: Order of magnitude OPEX savings : reduction inconfiguration, and ease of deployment. Order of magnitude CAPEX savings : Natively onthe switch/router: Wiring, Power, Rackspace andCost savings Scalability : Multi-Terabits/s Compute & Storage for Analytics, Historical Data 2019 Cisco and/or its affiliates. All rights reserved.
AI/ML At Cisco:Cognitive Collaboration 2019 Cisco and/or its affiliates. All rights reserved.
Webex Endpoints Built on Powerful AIWebex Board 70Webex Board 55Room 70DRoom 70SRoom 55Room Kit PlusRoom KitNVIDIA Jetson Platform - The same electronics engine powering self-driving cars 2019 Cisco and/or its affiliates. All rights reserved.
Webex Assistant“OK Webex” 2019 Cisco and/or its affiliates. All rights reserved.
Audio distance: 1.90 Quality: 0.58 Far end: 0AI that detects you 2019 Cisco and/or its affiliates. All rights reserved.F: 0.0% T: 86.4% U: 0.0% N: 0.0% S: 235People count: 6
AI that Recognizes youAddyJeff BlakeRuiAndreJohnAlicia 2019 Cisco and/or its affiliates. All rights reserved.Dyan
Noise Detection and Attenuation Noise detection on desktop clientsgenerally available since FY18-Q1 2019 Cisco and/or its affiliates. All rights reserved.
Classification: From Signals to ImagesDogSirenTappingSpectrogramWAVVoiceVoices and “noise” have a distinct “image” that can be detected and filtered.Deep Learning at Work in Cisco Collaboration Systems 2019 Cisco and/or its affiliates. All rights reserved.
AI/ML At Cisco:Enterprise Networking 2019 Cisco and/or its affiliates. All rights reserved.
The Power of AI/ML in the Network Anomaly detection Root cause analysis Dynamic network performance at different times and on different network conditionsDifferent expected performance on different SSIDs and/or locations for the samecustomerDifferent expected performance for different customersStatic thresholds (even if configurable) would likely raise many false positives or missrelevant eventsAutomatic selection of relevant KPIs explaining an issueCross-correlation across multiple devicesLong-term trending Automatically identifying trends and behavior changes on network entities/locations 2019 Cisco and/or its affiliates. All rights reserved.
CiscoDNAKairosAnalyticsProjectForForWireless,Wired Networksand andIoTWireless,WiredIOTAnomaly detection across hundredsof thousands of devices andthousands of networksAPsCognitive AnalyticsAnomaly detectionRoot-cause AnalysisLong-term trending 2019 Cisco and/or its affiliates. All rights reserved.
2019 Cisco and/or its affiliates. All rights reserved.ThroughputUse regression topredict upper and lowerband.Client Device Packet Retriesper secmost impactedApp Throughput – High Packet RetriesCorrelate with otherpotential issues that areexperiencing peaks /valleys in performance.Determine if issue isspecific to any specificdevices.
AI/ML At Cisco:Security 2019 Cisco and/or its affiliates. All rights reserved.
Cisco Cognitive IntelligenceEarly Detection & Response with Artificial Intelligence10Brequestsper day20Kincidentsper icEventsIncidents 2019 Cisco and/or its affiliates. All rights reserved.
Cisco Cognitive IntelligenceLayer 1Unsupervised LearningLayer 2Classification (Supervised)10Brequestsper icEventsIncidents 2019 Cisco and/or its affiliates. All rights reserved.20Kincidentsper day
Stealthwatch for SecurityDetecting Malware Embedded in Encrypted TrafficInitial Data PacketMake the most of theunencrypted fieldsSequence of PacketLengths and TimesIdentify the content typethrough the size and timingof packetsC2 messageData exfiltrationSelf-Signed certificate 2019 Cisco and/or its affiliates. All rights reserved.Global Risk MapWho’s who of the Internet’sdark sideBroad behavioralinformation about theservers on the Internet.
Limitations, Society, Ethics, Future ofAI/ML 2019 Cisco and/or its affiliates. All rights reserved.
Limitation of AIPerformance limitation Not easy to explain how AI make the decision visualization for peopleBiased AI through biased data - people doesn't want discrimination, but often they do such and if AI gets biased data,they could make a biased decision AI unhealthy stereotypes - if man is king or father, then women is queen or mother, but when man isprogrammer, women is homemaker facial recognition better works for light-skinned than dark-skinned - this is unfair and needs to be biasedCombating bias: technical solution: "zero out" the bias in words some words manually put to zero use less biased and/or more inclusive data (e.g. for facial recognition system include data for multi-ethnics) transparency and/or auditing proc
Artificial Intelligence, Machine Learning, and Deep Learning (AI/ML/DL) F(x) Deep Learning Artificial Intelligence Machine Learning Artificial Intelligence Technique where computer can mimic human behavior Machine Learning Subset of AI techniques which use algorithms to enable machines to learn from data Deep Learning
This Artificial Intelligence Master's Program provides training for the most sought after skills required for a successful career in Artificial Intelligence. As you undertake your Artificial Intelligence Engineer training, you will master the concepts of Deep Learning, Machine Learning, Natural Language Processing plus
Artificial Intelligence -a brief introduction Project Management and Artificial Intelligence -Beyond human imagination! November 2018 7 Artificial Intelligence Applications Artificial Intelligence is the ability of a system to perform tasks through intelligent deduction, when provided with an abstract set of information.
This Artificial Intelligence Master's Program covers the crucial skills you need for a successful career in artificial intelligence (AI). As you undertake your Artificial Intelligence course, you'll master the concepts of the machine and deep learning—plus the programming languages needed to excel in an AI career with exclusive
Artificial Intelligence enabled by Machine and Deep Learning Machine Learning . Real Truth of Artificial Intelligence by Whit Andrews Presented at Gartner Data & Analytics Summit 2018, March 2018. 5 Big Data Compute Power Machine Learning . Data Analytics Business System MATLAB Produc
Deep learning Machine learning methods based on neural networks. The adjective 'deep' refers to the use of many hidden layers in the network, two hidden layers as a minimum but usually many more than that. Deep learning is a subset of machine learning, and hence of artificial intelligence more broadly. Artificial neural networks
CS8691-Artificial Intelligence. 23 of 32 Learning to Imitate a Human Action CS8691-Artificial Intelligence. 24 of 32 Result after Learning HumanAction Imitation CS8691-Artificial Intelligence. 25 of 32 The Future of AI Massiv
Learn the history of deep learning and the business stakeholders in a deep learning project Learn the tools used by deep learning engineers INTRODUCTION TO . Course 4: Operationalizing Machine Learning Projects on SageMaker This course covers advanced topics related to deploying professional machine learning projects on
His catalogue includes a ballet, orchestral, chamber, piano, vocal and choral compositions. His unrecorded Symphonies are: Nos. 2 for Pipa and Large Orchestra (1981), 7 for Chinese Orchestra "The Great Wall" (2004) and 8 for Organ, hoir and Orchestra "This oundless Land’ (2007). Symphony No. 1 (1979) Wing-Sie Yip/Russian Philharmonic Orchestra ( Symphony No. 3 and Morning Sun) HUGO HRP 795 .
