WIXLIB

Responding to the AI ChallengeLearning from Physical IndustriesDavid Richdrich at mathworks dot com 2019 The MathWorks, Inc.1

How can other industries inform AI in finance? Four learnings from outside of financeThree areas of explorationTwo quick MATLAB PSAs (public service announcements)AI in this talk includes;machine learning, deep learning, reinforcement learning 2

Our Customers / Key IndustriesAerospace and DefenseAutomotiveBiological SciencesBiotech and PharmaceuticalCommunicationsElectronicsEnergy ProductionFinancial ServicesIndustrial MachineryMedical DevicesProcess IndustriesNeuroscienceRailway SystemsSemiconductorsSoftware and Internet3

Four Learnings from Other Industries1.2.3.4.Plenty of value away from the “obvious” applicationsThere’s no reason not to look for your keys under the street light:If you have data use itRegulations can be tough – but perhaps not for advice.If you don’t have data, can you create it?4

Four Learnings from Other Industries1.2.3.4.Plenty of value away from the “obvious” applicationsThere’s no reason not to look for your keys under the street light:If you have data use itRegulations can be tough – but perhaps not for advice.If you don’t have data, can you create it?5

Society of Automotive EngineersLevels of Autonomous VehiclesAdvanced DriverAssisted Systems (ADAS)Automated Driving Systems6

Subaru (a customer)Advanced Driver-Assistance SystemsCritical safety features for everyoneDetects obstacles, applies brakes, adjustscruise control, and stays in lane7

BMW - Machine Learning to Detect Oversteering“With little previous experience with machinelearning, we completed a working ECUprototype capable of detecting oversteeringin just three weeks.” Tobias Freudling, BMWGroupA BMW M4 oversteering on a test track. Engineers gathering and cleaned dataExplored many machine learning approacheswith Classification Learner AppGenerated code for vehicles on test trackLink to technical article8

Four Learnings from Other Industries1.2.3.4.Plenty of value away from the “obvious” applicationsThere’s no reason not to look for your keys under the street light:If you have data use itRegulations can be tough – but perhaps not for advice.If you don’t have data, can you create it?10

Four Learnings from Other Industries1.2.3.4.Plenty of value away from the “obvious” applicationsThere’s no reason not to look for your keys under the street light:If you have data use itRegulations can be tough – but perhaps not for advice.If you don’t have data, can you create it?11

Musashi Seimitsu Industry Co.,Ltd.Detect AbnormalitiesAutomated visualinspection of 1.3 millionbevel gear per monthManufacturersoften have a troveof labelled data12

Four Learnings from Other Industries1.2.3.4.Plenty of value away from the “obvious” applicationsThere’s no reason not to look for your keys under the street light:If you have data use itRegulations can be tough – but perhaps not for advice.If you don’t have data, can you create it?13

GenentechDeep Convolutional Neural Networks for Digital Pathology AnalysisGenerate training data iteratively Model is iteratively improved by addingmore data Removes need to annotate tumor by handGreen tumorRed necrosisBlue otherSegment tumor tissue from necrosis Segmentation of massive 25k x 25k images Trained and deployed U-Net semanticsegmentation algorithmNot a diagnosis!Assists pathologistPresented at American Conference on Paramacometrics ( 7th October 2018)Deep Convolutional Neural Networks for Digital Pathology AnalysisJeffrey Eastham-Anderson, Kathryn Mesh, Jeff Hung, Andrea Dranberg (MathWorks)14

MATLAB PSA #115

Use the Live Editor to create scripts that combine code,output, and formatted text in an executable notebook.16

Now integrated withSymbolic Math ToolboxSymbolic and Numericin one Live Editor Notebook17

Four Learnings from Other Industries1.2.3.4.Plenty of value away from the “obvious” applicationsThere’s no reason not to look for your keys under the street light:If you have data use itRegulations can be tough – but perhaps not for advice.If you don’t have data, can you create it?18

Predictive Maintenance: Reciprocating PumpPredict pump failures in real-time using sensor data“I keep my machineshealthy and running sohow do I get failuredata to train a model?”19

Generate DataBuild modelObtain sensor dataSimulink ModelRun simulationsSensor DataGeneratedFailure DataFine tune modelIncorporate failure modesInject Failures20

Preprocess DataData PreprocessingMethodsFailure Data (Sensors/Simulation) Time Domain Frequency Domain Time-FrequencyDomainPreprocessed Data21

Feature Extraction & Condition MonitoringFeature ExtractionMethodsPreprocessed Data Order/Modal Analysis Time-FrequencyAnalysis Input-Output Models Model Coefficients &States Residual Generation .Frequency PeaksHealth Indicators22

Predictive Model TrainingPredictive MethodsFrequency PeaksHealth Indicators Anomaly Detection Fault Classification Remaining Useful Life Trending Hazard Distributions Time seriesForecasting .Remaining UsefulLife 43.4 days100 Days toFailure0 Days toFailureFault Classifier23

Three Topics to Watch24

Three Topics to Watch#3 Fragmentation in Hardware Architecture? (for Deep Learning)25

Fragmentation in Hardware Architecture?Multiple vendors doing ARM NVIDIA is the standard fordata-center deep learningBut there are challengers;–––– FPGA from Xilinx, Intel, othersAMD RadeonGoogle’s TPUEmbedded processors from ST,TI, Renasas, InfineonOver 1B of ventureinvestment in AI chip startupsCloud is an acceleratorTraining vs. Inference26

GPU falling out of favor as hardware for embedded deployment?Edge computing hardware zoo:Intel Neural Compute Stick 2 (left, top)Movidus Neural Compute Stick (left, bottom)NVIDIA Jetson Nano (middle, top)Raspberry Pi 3, Model B (middle, bottom)Coral USB Accelerator (right, top) Google TPUCoral Dev Board (right, bottom) Google TPUImage courtesy of Dr. Allasdair Allanhttp://bit.ly/great-big-roundup27

If multiple architectures become viable, then what? Evaluate HW for purpose – choose targetDevelop in high level languageTransform to target executableA small number of finance customers doing this today(GPU, FPGA). Will this grow?28

MATLAB PSA #229

Run MATLAB code faster with redesigned execution engine. All MATLAB code is now JIT compiled Incremental improvements each release– Faster assignment into large table, datetime,duration, and calendarDuration arrays– Construct objects and set properties faster – Render plots with large numbers ofmarkers faster using less memory– Increased speed of MATLAB startup30

Three Topics to Watch#3 Fragmentation in Hardware Architecture? (for Deep Learning)#2 Reinforcement Learning31

Reinforcement Learning in the News Focuses Mainly On Board Gameso Chesso Go Video Gameso Atario DoTA, Starcraft RecommendationSystems32

But Increasingly Being Seen In Context of Autonomous Systems Learn Complex Taskso Manipulationo Planningo Navigationo ControlSource: OpenAISource: Google33

Traditional “Controls” Customers Have Proactively Engaged100 customers have spoken to us about Reinforcement Learning since 2018 Reinforcement learningneeds a lot of data, usuallygenerated from models Models can incorporateconditions hard to emulate inthe real world Many of them haveMATLAB and Simulinkmodels that can be reusedAutonomous Systems(e.g. Robots)Calibration Problems(e.g. Engine Maps)Controller Design forNonlinear SystemsAutomated Driving34

Using Reinforcement Learning to Improve Driving ControlModels like this are used by our customers to develop controllers and other algorithmsRL VehicleModels35

RL for Autonomous Driving – Co-simulating with Unreal EngineProject With A Major Automotive Company Step 1: Trained deep neuralnetwork (DNN) based driver Step 2: Use RL to improveperformance of DNN-based driver Step 3: Use improved DNN toaugment traditional controller Result: 2 sec (7%) faster than theoriginal driver36

RL Interest Growing in Finance Notably JP Morgan LOXM (Limit OrderExecution Management)Positive results for our first experiments– Stock trading– Hedging Academic activity – mostly automated tradingPetter Kolm from NYU speaking later;– Dynamic Replication and Hedging:A Reinforcement Learning Approach37

Three Topics to Watch#3 Fragmentation in Hardware Architecture? (for Deep Learning)#2 Reinforcement Learning#1 Explainability and V&V for AI38

Explainability in Finance: Principles of fairness requirebeing able to explain why the model is making decisions2. Use of personal attributes as input factors for AIDA-drivendecisions is justified.4. AIDA-driven decisions are regularly reviewed so that modelsbehave as designed and intended.8. Firms using AIDA are accountable for both internallydeveloped and externally sourced AIDA models.13. Data subjects are provided, upon request, clearexplanations on what data is used to make AIDA-drivendecisions about the data subject and how the data affects thedecision.“AIDA” refers to artificial intelligence or data analytics, which are defined as technologies that assist orreplace human decision-making.https://www.mas.gov.sg/ al.pdf39

Trade-off between predictive power and explainabilityEnsemble methodsExplainable estsBoostedTreesDeepNeuralNetworksLarge quantities of dataAutomated feature extractionClassical Machine Learning40

Attribution Reveals the Why Behind Deep Learning DecisionsClassified as “keyboard” due in partto the presence of the mouseIncorrectly classified “coffee mug”as “buckle” due to the /help/deeplearning/ug/gradcam-explains-why.html41

Across Industries there are different meanings for Verification & Validation for AIEXPLAINABILITYCan you explainthe working of AImodel in SDATA PRIVACYRIGOR & TRUSTCan you observeand trace causeand effect in an AImodel and explainthe rationale ofthe decision?Is AI systemimmune fromspoofing and othercommon attacks ?Can an attackerdeduce sensitivetraining data fromoutput of AI modelor system?Has AI systembeen developedwith defined,traceable andrigorous process?SAFETYCERTIFICATIONHas AI systembeen developedwith safetylifecycle as keycomponent.42

Common safety practicesRedundancyRedundantimplementationwith votingFMEAMonitoringFailure modeanalysis to ensuresafe behavior inanticipatedscenariosMonitoring andlogging to recorddecisions for postmortem analysisAuditingAudit trails ofdevelopmentactivitiesDataSeparationFirewalls betweentraining andtest/validationdata to ensureappropriateaccuracy metricsare computedBestPracticesDocumentation ofbest practicesbeing followed toimproverepeatability43

Safety Standards Updates:Very Early Phase TUV SUD– Open GENESIS– Started in May 2019 SAE and EUROCAE– Joint Working Group WG-114– Kick-off in August 2019 RTCA (Aerospace, US)– Still evaluating member’s interests ISO JTC 1/SC 42– Standardization program on Artificial Intelligence– In “Preparatory” phase – work not yet started44

When designing physical productsTest &VerificationModeling & SimulationCode Generation45

Much of the process may be regulated46

Connected Systems Means st & VerifyGenerateCode, DeployOperate47

Positive improvements from field SimulateTest & VerifyGenerateCode, DeployOperate48

Model Risk Management – The Model is the Product Risk managementBoard and stakeholdersRegulatorModel OwnersEmphasis onMonitoring,Reporting andPerformanceAssessment– Process– People– Inventory of Assets– Execution Phase– Reporting IT, front-office Model developers,quants, analysts.Business linesImplementationand DeploymentDefinition andDevelopmentModelInventoryReview andApprovalIndependent modelreview and auditRegulatorModel UsersLearn more– A Platform for Risk ModelsPaul Peeling, idationSLODModel Validation, IT49