Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5...
Transcript of Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5...
![Page 1: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/1.jpg)
1© 2015 The MathWorks, Inc.
Unlocking the power of
Machine Learning
Mandar Gujrathi
![Page 2: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/2.jpg)
2
AI will change the way we do things
https://www.cnbc.com/2018/01/05/how-artificial-intelligence-will-affect-your-life-and-work-in-2018.html
![Page 3: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/3.jpg)
3
Machine Learning has driven Innovation
Electric Grid
Load
Forecasting
Sentiment Analysis in Finance
Restore Arm
Control for
Quadriplegic
Robots mimic
complex human
behaviors
![Page 4: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/4.jpg)
4
ChallengeRestore arm and hand control to a quadriplegic man by processing signals from an
electrode array implanted in his brain
Products Used• MATLAB + Wavelet Toolbox
Approach• Used MATLAB to analyze signal samples
• Generated compact feature vectors using wavelets – wavelets allowed the
researchers to extract the important information from the signals for
classification.
• Applied machine learning to classify patterns mapped to movements, and
• Generate actuation signals for a neuromuscular electrical stimulator
Results▪ Control over paralyzed hand and arm restored
▪ Real-time processing performance achieved
Battelle Neural Bypass Technology Restores Movement to a
Paralyzed Man’s Arm and Hand
Link to full user story
“The algorithms we developed using
MATLAB gave the participant back
basic control of his arm and hand. By
the end of the study, he could grip a
bottle, pour out its contents, and set it
down, as well as pick up a stir stick and
execute a stirring motion.”
David Friedenberg
Battelle
Patient using the Battelle NeuroLife system.
![Page 5: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/5.jpg)
5
Outline
Machine Learning and its challenges
Developing a Heart Sound Classifier
Going beyond Machine Learning
![Page 6: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/6.jpg)
6
Key takeaways
Empower engineers to be productive in data science!
▪ Cover complete workflow (exploration to deployment)
▪ Apps for Machine Learning
▪ Support for Deep Learning
![Page 7: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/7.jpg)
7
2. Explore and Pre-Process
3. Extract Features
4. Train Models
Developing Machine Learning Applications
5. Deploy
1. Access Data
Sensors
Various ProtocolsDiverse data
Clean messy data
Discover patterns
Domain Knowledge
Select Features
Many Algorithms
Tune Parameters
Different platforms
Size/Speed
![Page 8: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/8.jpg)
8
Outline
Machine Learning and its challenges
Developing a Heart Sound Classifier
Going beyond Machine Learning
![Page 9: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/9.jpg)
9
Case Study: Heart Sound Classifier
Motivation
– Heart sounds require trained clinicians for diagnosis
– Lowered FDA requirements renewed interest
Goal: build a classifier and deploy in portable device
Data: Heart sound recordings (phonocardiogram):
– From PhysioNet Challenge 2016
– 5 to 120 seconds long audio recordings
Feature Extraction
Classification Algorithm
Normal
Abnormal
Heart Sound Recording
![Page 10: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/10.jpg)
10
Different Types of Learning
Machine Learning
Supervised
Learning
Classification
Regression
Unsupervised
LearningClustering
No output - find natural groups
and patterns from input data only
Output is a real number
(temperature, stock prices)
Output is a choice between
classes (Normal, Abnormal)Classification
![Page 11: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/11.jpg)
11
Step 1: Access & Explore Data
Challenges
– Different sampling rates
– Signal Management
– Large datasets (“big data”)
Easy Exploration of Data
– Time domain
– Frequency domain
– Time-Frequency domain
Signal Analyzer: Visual Data Exploration
![Page 12: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/12.jpg)
12
Step 2: Pre-process Signals
Challenges
– Preserving sharp features
– Overlap of signal and noise spectra
Automatic Denoising
Generate MATLAB code
Signal Pre-processing without writing any code
![Page 13: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/13.jpg)
13
Step 3: Extract Features
Challenges
– Find features for non-stationary signals
– Features occurring at different scales
– Feature selection
Spectral features
– Mel-Frequency Cepstral
– Octave band decomposition with Wavelets
![Page 14: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/14.jpg)
14
Step 4: Train Models
Challenges
– Knowledge of machine
learning algorithms
– Scale to large data sets
Quickly train model in App
– Define cross-validation
– Try all popular algorithms
– Analyze performance
93% on test data
Model Training with Classification Learner
Scale to large data sets
without recoding, “Tall” arrays
![Page 15: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/15.jpg)
15
![Page 16: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/16.jpg)
16
Step 4 Cont’d: Optimize Model
Challenges:
– Manual parameter tuning tedious
– Identify additional improvements
– Imbalanced data
Class Distribution
Normal 75%
Abnormal 25%
Iterative Model Optimization
– Bayesian Optimization of parameters
– Visually analyze performance
– Adjust for imbalances (data or
severity of misclassifications)
![Page 17: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/17.jpg)
17
MATLAB
Runtime
MATLAB
Compiler SDK
MATLAB
Compiler
MATLAB
MATLAB Coder
Step 5: Deploy
Embedded Hardware Enterprise Systems
MATLAB Production
Server
![Page 18: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/18.jpg)
18
Outline
Machine Learning and its challenges
Developing a Heart Sound Classifier
Going beyond Machine Learning
![Page 19: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/19.jpg)
19
Beyond Machine Learning: Deep Learning
Supervised Classification using Neural Nets with many layers
1. Convolutional Neural Networks (CNN)
– A versatile and flexible approach for Deep Learning
– Apply to signals by converting to time-frequency representation:
2. Long short-term memory networks (LSTM)
![Page 20: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/20.jpg)
20
Apply Deep Learning to Heart Sound Classifier
Steps
– Signal Time-Frequency
– Continuous Wavelet Transform
– Transfer Learning with GoogLeNet
Results
– Achieves 90% accuracy
– Just 10 lines of code
Deep Learning Training Visualiser
![Page 21: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/21.jpg)
21
2. Explore and Pre-Process
Visual Exploration
Wavelets
Feature Selection
3. Extract Features
4. Train Models
Quickly compare models in App
Automatically tune parameters
Explore Deep Learning
Summary: Making Machine Learning Easier
5. Deploy
1. Access Data
Support for industrial
sensors, phones, etc.
Automatically
Generate C/CUDA
Code
![Page 22: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/22.jpg)
22
Learn More
Complete user story for Battelle’s “NeuroLife” system
Download Heart Sounds Classification application from File Exchange
Watch “Machine Learning Using Heart Sound Classification”
Read:
– Machine Learning with MATLAB
– What is Deep Learning?
![Page 23: Unlocking the power of Machine Learning - …...Patient using the Battelle NeuroLife system. 5 Outline Machine Learning and its challenges Developing a Heart Sound Classifier Going](https://reader033.fdocuments.in/reader033/viewer/2022060320/5f0cf3857e708231d437f04f/html5/thumbnails/23.jpg)
23
Key takeaways
Empower engineers to be productive in data science!
▪ Cover complete workflow (exploration to deployment)
▪ Make machine learning easy
▪ Support for Deep Learning