Transcription
MATLAB과 함께하는 딥러닝 4주 완성 부트캠프
세션2. 신호처리를 위한 머신러닝과 딥러닝MATLAB과 함께하는 딥러닝 4주 완성 부트캠프김종남 부장Application Engineer @ MathWorkscalebkim@mathworks.com
Understanding AI megatrendARTIFICIAL INTELLIGENCEAny technique that enables machines to mimichuman intelligenceMACHINE LEARNINGStatistical methods that enable machines to“learn” tasks from data without explicitlyDEEP LEARNINGprogrammingNeural networks with many layers that learnrepresentations and tasks “directly” from data1950s1980s2010s3
MATLAB AI used in Industries and ResearchConverting brain waves tospeech to help ALSpatients communicateUT AustinSeismic Event DetectionShellRestoration of arm andhand control byprocessing brain signalsBattelle4
Agenda AI Workflow Example 1: ECG Classification Example 2: Pavement Crack Identification Example 3: Human Activity Classification5
AI-driven system designData PreparationAI ModelingDeploymentData cleansing andpreparationModel design andtuningEmbedded devicesHuman insightHardwareaccelerated trainingEnterprise systemsSimulationgenerated dataInteroperabilityEdge, cloud,desktop6
Transforming raw data for useful analysis is a critical stepData PreparationAI ModelingDeploymentData cleansing andpreparationModel design andtuningEmbedded devicesHuman insightHardwareaccelerated trainingEnterprise systemsSimulationgenerated dataInteroperabilityEdge, cloud,desktop7
Synthetic Data Generation and Augmentation to deal with less dataSimulate data using models anddeep learningGenerate Radar ReturnsGenerate wireless waveformsGenerate Audio DataData PreparationData cleansing andpreparationHuman insightSimulationgenerated datatext2speech88
Use Signal Processing Apps to speed up Labeling and PreprocessingData PreparationData cleansing andpreparationSignal Labeler appHuman insightSimulationgenerated dataSignal Analyzer App9
Easily Extract Features from SignalsTime-Domain Features Signal Patterns Changepoints Peaks Signal Envelope Data PreparationData cleansing andpreparationHuman insightSimulationgenerated dataFrequency-Domain Features BWmeasurements Spectral Statistics Octave Spectrum Time-Frequency features SpectrogramMFCCConstant-QTransform Scalogram Domain-Specific Features Speech andaudio Navigation andSensor Fusion Radar Communication 10
Selecting and training the right model is importantData PreparationAI ModelingDeploymentData cleansing andpreparationModel design andtuningEmbedded devicesHuman insightHardwareaccelerated trainingEnterprise systemsSimulationgenerated dataInteroperabilityEdge, cloud,desktop11
Understand the approaches to creating AI modelsMachineLearning ModelsClusteringDecision treesSVMsAI ModelingModel design andtuningDeep Learningfrom scratchCNNsHardwareaccelerated trainingLSTMsInteroperabilityDeep Learningusing TransferLearning12
AI ModelingModel design andtuningHardwareaccelerated trainingData VolumeUnderstanding tradeoffs to building modelsInteroperabilityTime Required13
AI ModelingModel design andtuningHardwareaccelerated trainingPredictive PowerUnderstanding tradeoffs to building modelsInteroperabilityInterpretability14
MATLAB interoperates with other frameworksAI ModelingModel design andtuningHardwareaccelerated trainingKeras importerTensorFlow importerCaffe importerInteroperability15
Hardware acceleration and scaling are critical for trainingAI ModelingModel design andtuningHardwareaccelerated trainingInteroperability16
Completing your AI workflow with deploymentData PreparationAI ModelingDeploymentData cleansing andpreparationModel design andtuningEmbedded devicesHuman insightHardwareaccelerated trainingEnterprise systemsSimulationgenerated dataInteroperabilityEdge, cloud,desktop17
Deploy to any processor with best-in-class performanceCPUDeploymentEmbedded devicesCodeGenerationGPUEnterprise systemsEdge, cloud,desktopFPGA18
ECG Waveform Segmentation ECG signals contain P,QRSand T waves. Identifying thesecan help with diagnosis andclassification. Dataset contains 210 ECGsignals , 15 minutes long,labeled by cardiologist19
ECG Segmentation using a LSTM NetworkPreprocess Dataand FeatureExtractionBuild Deep Networkfrom ScratchDeploy & Integrate20
Building a Deep Learning Model from Scratch21
Script -Video22
Getting data from Bitalino card23
Deployment to a Raspberry Pi Board Raspberry Pi 3 B with ARM Cortex Aprocessor Compute Library from ARM Repository of low-level optimizedfunctions For ARM Cortex A processors andARM Mali family of arning-on-raspberry-pi.html24
Using Wavelets for AIWhy are wavelets and multiscale techniques useful? Efficient representations for datacharacterized by transient components andlong duration trends or oscillations Provides state-of-the- art results for : Anomaly detection and HealthMonitoring Biomedical Signal Analysis Seismic Analysis Radar and Comms applications Financial Analysis25
Which wavelet techniques can be used for AI?Continuous Wavelet TransformWavelet ScatteringWavelet Statistics.and many more26
Crack Identification in Pavements with Wavelets Image analysis of pavements is not alwaysfeasible Data is collected using accelerometersensor installed in cars Dataset: Mendeley Data open datarepository – 327 samples Data is collected at different car speeds forvarying sizes of pavement cracks27
Crack Identification with Pretrained NetworkPreprocess Dataand FeatureExtractionUse Deep Learningwith TransferLearningWaveletTransform28
Using Transfer LearningLoad pretrained networkEarly layers that learnedlow-level features(edges, blobs, colors)Last layers thatlearned taskspecific features1 million images1000s classesReplace final layersNew layers to learnfeatures specificto your dataFewer classesLearn fasterTrain networkTraining imagesTraining optionsPredict and assessnetwork accuracyTest images100s images10s classesTRAINED NETWORK29
Script Video30
Human Activity Classification Classify from 5 activities:––––– WalkingClimbing upstairsGoing downstairsLying downstandingData captured from sensors on mobilephones31
Human Activity Classification with AutoMLPreprocess Dataand AutomaticFeature ExtractionWaveletScatteringTrain and TestMachine LearningModelsFeature and ModelSelection32
Wavelet Scattering33
Script Video34
MATLAB supports the entire AI-driven system designData PreparationAI ModelingDeploymentData cleansingand preparationModel design andtuningEmbeddeddevicesHuman insightHardwareacceleratedtrainingEnterprise systemsInteroperabilityEdge, cloud,desktopSimulationgenerated dataSignal Processing appsGenerate DataQuickly build modelsDeploy to targets withcode generationFeature Extraction TechniquesAccelerate training35
Wavelet ToolboxSignal Processing ToolboxDeep Learning ToolboxGPU CoderStatistics & Machine LearningToolboxMATLAB CoderMATLAB36
Many resources to get started arning/deep-learning-signal-processing.html37
MATLAB과 함께하는 딥러닝 4주 완성 부트캠프 일정7월 7일7월 14일7월 21일영상 분석을위한 딥러닝신호처리를위한 머신러닝과 딥러닝딥러닝프로젝트를위한 데이터준비 기법7월 22일7월 28일MATLABMATLAB으로Deep Learning시작하는Day 2021강화학습딥러닝 부트캠프 응용편라이다 및딥러닝 기반무선통신을레이다를 위한이상탐지 기법위한 딥러닝딥러닝바이오,의료분야를위한 딥러닝https://bit.ly/3hfSm24오늘 등록하세요!
감사합니다
Machine Learning Models Wavelet Scattering Feature and Model Selection. 33 Wavelet Scattering. 34 Script Video. 35 . Signal Processing Toolbox Wavelet Toolbox Statistics & Machine Learning Toolbox Deep Learning Toolbox GPU Coder MATLAB Coder. 37 Many resources to get started with