Alain Goossens & Jean-Pierre Van Loo Data scientists – SII Belgium

Slides:



Advertisements
Similar presentations
Observation Pattern Theory Hypothesis What will happen? How can we make it happen? Predictive Analytics Prescriptive Analytics What happened? Why.
Advertisements

MACHINE LEARNING AND ARTIFICIAL NEURAL NETWORKS FOR FACE VERIFICATION
Why consider the cloud? Cloud innovation presents challenges for IT.
Data Mining. 2 Models Created by Data Mining Linear Equations Rules Clusters Graphs Tree Structures Recurrent Patterns.
Andy Roberts Data Architect
Azure Machine Learning My first Data Science experiment Using Azure Machine Learning.
Introduction to Azure Machine Learning and Data Mining algorithms Oleksandr Krakovetskyi CEO, DevRain Solutions PhD, Microsoft Regional
Agility Dev TestDeploy Learn Agility.
AZ PASS User Group Azure Data Factory Overview Josh Sivey, Solution Partner October
A Suite of Products that allow you to Predict Outcomes, Prescribe Actions and Automate Decisions.
Artificial Intelligence and Machine Learning in Big Data and IoT: The Market for Data Capture, Analytics, and Decision Making 2016 – 2021 Phone No.: +1.
Advancements in Analytics with Azure Machine Learning James Wang Technical Evangelist Microsoft Taiwan Slide modified from
Cloud Analytics Platforms Christian Frey. About AIDA Our mission is to advance knowledge in data analytics through research, education and outreach Our.
9/24/2017 7:27 AM © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN.
Business Insights Play briefing deck.
11/19/2017 9:41 PM © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN.
Bhakthi Liyanage SQL Saturday Atlanta 15 July 2017
Energy Demand Forecasting
2/13/2018 4:38 AM © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN.
4/19/ :02 AM © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN.
4/18/2018 3:49 PM © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN.
Azure Machine Learning
Data Platform and Analytics Foundational Training
Predicting Azure Consumption using Ensemble Learning
ANOMALY DETECTION FRAMEWORK FOR BIG DATA
Cortana Intelligence Overview
Parcel Tracking Solution Parcel Tracking What to look for Architecture
Working With Azure Batch AI
Orchestrating Data and Services with Azure Data Factory
Siemens Enables Digitalization: Data Analytics & Artificial Intelligence Dr. Mike Roshchin, CT RDA BAM.
School of Computer Science & Engineering
Prepared by: Mahmoud Rafeek Al-Farra
AI development using Data Science Virtual Machines (DSVM) in Azure
Machine Learning in practice
Enable the Hybrid Data Platform
Optimizing Edge-Cloud IoT Applications for Performance and Cost
Energy Demand Forecasting
Microsoft Build /22/ :52 PM © 2016 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY,
Remote Monitoring solution
Accelerate your advanced analytics practice using solution templates
Azure Infrastructure as a Service
9/21/2018 3:41 AM BRK3180 Architect your big data solutions with SQL Data Warehouse & Azure Analysis Services Josh Caplan & Matt Usher Program Managers.
Migrating Your BI Platform To Azure
Intro to Machine Learning
Dive into Predictive Maintenance using Cortana Intelligence Suite
Object Recognition & Detection
Artificial Intelligence Changes the Security Landscape
CSSE463: Image Recognition Day 11
Microsoft Azure Enables Big-Data-as-a-Service Applications for Industry and Government Use “Microsoft Azure is the most innovative and robust suite of.
Overview of Machine Learning
Analytics in the Cloud using Microsoft Azure
Context about the Data Warehouse
Intro to Machine Learning
What’s New and What’s Coming…
Azure Machine Learning on Databricks
Presentation By: Eryk Helenowski PURE Mentor: Vincent Bindschaedler
Agenda Need of Cloud Computing What is Cloud Computing
Machine Learning for Space Systems: Are We Ready?
What's New in eCognition 9
Get High Score By Just Preparing These Dumps Microsoft DP-201 Dumps Pdf
Introduction to Azure Data Lake
Data Wrangling for ETL enthusiasts
Customer 360.
Michael French Principal Consultant 5/18/2019
The Modern Data Warehouse and Azure
Microsoft Business Analytics and AI
SQL Server 2019 Bringing Apache Spark to SQL Server
Analytics on Azure What to Use When Christina E. Leo
COGNITIVE SERVICES MACHINE LEARNING FOR DEVELOPERS
Architecture of modern data warehouse
Presentation transcript:

Alain Goossens & Jean-Pierre Van Loo Data scientists – SII Belgium Applying neural networks to anomaly detection Alain Goossens & Jean-Pierre Van Loo Data scientists – SII Belgium Data Innovation Summit 27th June 2018 #DISUMMIT

Underground cabling, infrastructure and pipelines Applying neural networks to anomaly detection, Alain Goossens & Jean-Pierre Van Loo, @alaingoossens4 @jpvl1 #DISUMMIT

Current anomaly detection process Desk top analysis Visual analysis of pictures Risk analysis Verification of permits When necessary send an intervention team Intervention team Applying neural networks to anomaly detection, Alain Goossens & Jean-Pierre Van Loo, @alaingoossens4 @jpvl1 #DISUMMIT

Artificial intelligence Concept of the POC YES there is an anomaly NO there is no anomaly Artificial intelligence INPUT pictures Applying neural networks to anomaly detection, Alain Goossens & Jean-Pierre Van Loo, @alaingoossens4 @jpvl1 #DISUMMIT

Overview Advanced analytics pattern in Azure MODEL TRAINING SERVING STORAGE AZURE ML AZURE ML STUDIO ML SERVER AZURE DATABRICKS (Spark ML) SQL Server (In-database ML) DATA SCIENCE VM BATCH AI SENSORS AND IOT (UNSTRUCTURED) COSMOS DB APPLICATIONS LONG TERM STORAGE DATA PROCESSING SQL DB r LOGS, FILES AND MEDIA (UNSTRUCTURED) SQL DB DATA LAKE STORE AZURE STORAGE COSMOS DB DATA LAKE ANALYTICS AZURE DATABRICKS HDINSIGHT SQL DW TRAINED MODEL HOSTING ORCHESTRATION AZURE ANALYSIS SERVICES BUSINESS / CUSTOM APPS (STRUCTURED) DASHBOARDS SQL Server (In-database ML) AZURE CONTAINER SERVICE DATA FACTORY Applying neural networks to anomaly detection, Alain Goossens & Jean-Pierre Van Loo, @alaingoossens4 @jpvl1 #DISUMMIT

Data preparation Categorization of the images Rural - Industrial 523 with bulldozers 567 with no anomalies Enrichment to have more images Cropping, flipping or translating images Applying neural networks to anomaly detection, Alain Goossens & Jean-Pierre Van Loo, @alaingoossens4 @jpvl1 #DISUMMIT

The models used A simple neural network with Microsoft Azure Machine Learning Studio  Convolutional neural network (CNN) and R-CNN on a deep learning Virtual machine Custom Vision Service Vision API + Microsoft Azure Machine Learning Studio based on output of the Vision API Applying neural networks to anomaly detection, Alain Goossens & Jean-Pierre Van Loo, @alaingoossens4 @jpvl1 #DISUMMIT

The models used A simple neural network with Microsoft Azure Machine Learning Studio  Convolutional neural network (CNN) and R-CNN on a deep learning Virtual machine Custom Vision Service Vision API + Microsoft Azure Machine Learning Studio based on output of the Vision API Applying neural networks to anomaly detection, Alain Goossens & Jean-Pierre Van Loo, @alaingoossens4 @jpvl1 #DISUMMIT

Two-class Neural network – AUC optimized

Principle of convolutional neural network (CNN) anomaly no anomaly Applying neural networks to anomaly detection, Alain Goossens & Jean-Pierre Van Loo, @alaingoossens4 @jpvl1 #DISUMMIT

Deep learning virtual machine Labelled data set, anomaly Labelled data set, normal condition Two class classification/prediction: YES anomaly No anomaly Applying neural networks to anomaly detection, Alain Goossens & Jean-Pierre Van Loo, @alaingoossens4 @jpvl1 #DISUMMIT

Convolutional neural network (CNN) Results 50 epochs Accuracy: Training = 98,5 % Validation = 92,9 % Test = 93 % True positives : 94,4 % False negatives : 5,6 % False positives : 8,4 % True negatives : 91,6 % Bulldozer detected when there is one No bulldozer detected when there is no No bulldozer detected but there is one Bulldozer detected but there is no Applying neural networks to anomaly detection, Alain Goossens & Jean-Pierre Van Loo, @alaingoossens4 @jpvl1 #DISUMMIT

Regions of interest on big images Applying neural networks to anomaly detection, Alain Goossens & Jean-Pierre Van Loo, @alaingoossens4 @jpvl1 #DISUMMIT

Custom Vision Service Applying neural networks to anomaly detection, Alain Goossens & Jean-Pierre Van Loo, @alaingoossens4 @jpvl1 #DISUMMIT

Computer Vision API 2000 known tags Applying neural networks to anomaly detection, Alain Goossens & Jean-Pierre Van Loo, @alaingoossens4 @jpvl1 #DISUMMIT

Machine learning studio on vision API tags Find a correlation between tags identified and the presence of a bulldozer Tested models: Two-class boosted decision tree Two-class logistic regression Two-class support vector machine Two-class decision jungle Two-class neural network Applying neural networks to anomaly detection, Alain Goossens & Jean-Pierre Van Loo, @alaingoossens4 @jpvl1 #DISUMMIT

Business conclusions Fairly good results with the main models used but with a dataset: trivial limited in variability Customer learned about AI, ML and NN in a real world context Further steps: Extend dataset variability Apply other models Applying neural networks to anomaly detection, Alain Goossens & Jean-Pierre Van Loo, @alaingoossens4 @jpvl1 #DISUMMIT

Business conclusions Fairly good results with the main models used but with a dataset: trivial limited in variability Customer learned about AI, ML and NN in a real world context Further steps: Extend dataset variability Apply other models Applying neural networks to anomaly detection, Alain Goossens & Jean-Pierre Van Loo, @alaingoossens4 @jpvl1 #DISUMMIT

Scientific conclusions Cropping images is absolutely required Learning speed of the neural network is quite satisfactory (for a relatively simple dataset) Ad hoc trained models perform better but pre- trained models are not to be excluded Using classifiers on metatags from a generic Vision API gives a fairly good result Applying neural networks to anomaly detection, Alain Goossens & Jean-Pierre Van Loo, @alaingoossens4 @jpvl1 #DISUMMIT

Scientific conclusions Cropping images is absolutely required Learning speed of the neural network is quite satisfactory (for a relatively simple dataset) Ad hoc trained models perform better but pre- trained models are not to be excluded Using classifiers on metatags from a generic Vision API gives a fairly good result Applying neural networks to anomaly detection, Alain Goossens & Jean-Pierre Van Loo, @alaingoossens4 @jpvl1 #DISUMMIT