Introduction to

Machine Learning
(with Azure Machine Learning service)
Contents
What is machine learning? 3 Democratizing AI with 21
Introduction to machine learning automated machine learning
Top objectives of automated
machine learning

Artificial intelligence versus 9

machine learning versus deep
learning Azure Machine Learning 24
A comparison of the 3 topics
service
How Azure Machine Learning service
meets customer objectives

Machine Learning: The Value it 12

Brings and case studies from
3 key industries How to get started 30
Next steps to get started
The value of deploying machine learning
in an organization

Who is this e-book for?

This e-book is designed to help business decision makers gain a

high-level understanding of machine learning and assist IT
decision makers in evangelizing the subject to stakeholders
within their organization.

2
What is
machine learning?
Machine learning (ML) is a rapidly developing field with a huge value
proposition for organizations across industries.

It evolved from artificial intelligence (AI) decades ago and with new
technological innovations, we are beginning to realize machine learning’s full
potential.

As new technologies like machine learning become critical to organizations,

business decision makers and employees who interact with them need to
expand their knowledge bases to achieve what is required for digital transformation.

Let’s start with a basic definition of machine learning.

3
 What is machine learning?
Data science technique: Machine learning Behaviors: Machine learning can predict
is used to mine predictions from statistical models. future behaviors based on past behaviors, such
Deep data science practices can also be applied that as surfacing suggestions based on previously
include modeless techniques and automatic identified patterns.
programming.

Machine learning is a
data science technique that allows
computers to use existing data to
forecast future behaviors,
outcomes, and trends.

Outcomes: Complex models are constantly Trends: Market index prices, forecasting
being tested to predict anything from sports the spread of diseases, and other trends can
1

outcomes to poverty outcomes to health be identified by recognizing patterns and

outcomes, and everything in between. existing trends in inputs.

By using machine learning, computers learn without being explicitly programmed.

A machine learning system uses data from sources such as apps, sensors,
networks, and devices to build its own logic to solve a problem or extract insight
*
(observations). There are different methods for using algorithms that find patterns
and insights.

1. https://www.beckershospitalreview.com/healthcare-information-technology/machine-learning-and-ai-are-taking-healthcare-by-storm-are-you-ready-to-rumble.html
2. https://docs.microsoft.com/en-us/azure/machine-learning/service/overview-what-is-azure-ml
*https://www.gartner.com/binaries/content/assets/events/keywords/catalyst/catus8/preparing_and_architecting_for_machine_learning.pdf
4
 Machine learning subdisciplines
Supervised learning is about having access to labeled data that is used as a guide to
teach the algorithm what conclusions to arrive to.
3

Supervised learning requires that the machine learning algorithm’s possible outputs are already known,
that target values are included in the data used for training and testing the algorithm.4 The algorithm
looks for patterns in the attributes of the data samples to try to predict target values correctly.5 After the
algorithm has discovered which features allow it to make accurate predictions for the training data, its
performance is confirmed by testing whether it can also make accurate predictions for the held-out
test data.6 The trained algorithm can then be used to make predictions on new data for which target
values are unknown.

Example:

Historical stock prices can be used to make predictions for

7

future prices.

There are several specific types of supervised learning that are represented within Azure Machine
Learning: classification, regression.8

Supervised learning
Supervisor

Training data set Desired output

Output
Input, labeled data

Algorithm Processing

3-4. https://www.datascience.com/blog/supervised-and-unsupervised-machine-learning-algorithms
5-8. https://docs.microsoft.com/en-us/azure/machine-learning/studio/algorithm-choice 5
 Machine learning subdisciplines
Unsupervised learning refers to the training of algorithms using information that is
neither classified nor labeled and allowing the algorithm to act on that information
without guidance.
9

Here the task of the algorithm is to group data samples according to patterns in similarities and
distances among them. Unlike supervised learning, no prior teaching is provided, that means no
training will be given to the machine. Therefore, the machine is restricted to find the hidden structure
in unlabeled data by itself.

Example:

Healthcare uses unsupervised learning to detect causality and to

identify correlations humans may miss by inputting health data
10
like blood pressure, heart rate, weight, prescriptions, etc.

Unsupervised learning

Input, non-labeled Output

data
Algorithm

Interpretation Processing

9. https://www.geeksforgeeks.org/supervised-unsupervised-learning/
10. http://people.csail.mit.edu/dsontag/courses/mlhc_summer18/day2/causal_inference.pdf 6
 Machine learning subdisciplines
Reinforcement learning is when a machine learning algorithm interacts with its
environment, either simulated or real, exploring different strategies to determine the
optimal sequence of choices to maximize reward.
Reinforcement learning methods resemble how humans and animals learn: The algorithm tries various
behavioral strategies, trying to maximize the reward it receives in response. Each choice the algorithm
makes when it encounters a new data point determines which data point it will encounter next.11 A
sequence of good decisions is reinforced with rewards. With experience the algorithm learns to produce
the optimal sequence of decisions within a given context.12

Example:
*
Enabling computers to learn to play games or drive vehicles.
Reinforcement learning is common in robotics, where the set of
sensor readings at one point in time is a data point, and the 13
algorithm must choose the robot's next action.

Reinforcement learning

Environment

Output
Input, labeled data

Best Action
Reward

Selection of
State Algorithm

Agent

11-12. https://docs.microsoft.com/en-us/azure/machine-learning/studio/algorithm-choice
*https://www.gartner.com/binaries/content/assets/events/keywords/catalyst/catus8/preparing_and_architecting_for_machine_learning.pdf
13. https://docs.microsoft.com/en-us/azure/machine-learning/studio/algorithm-choice
7
Machine learning subdisciplines
Deep learning uses what’s called a neural network architecture to do its learning. The neural
network architecture mimics how a brain works rather than using traditional statistical frameworks.
In this way, reinforcement learning could be placed under the deep learning umbrella.

The deep part of deep learning refers to the number of hidden layers in the neural network architecture,
which are usually several and going up to one-hundred fifty layers or more. There are many different
types of deep learning neural networks: convolutional, recurrent, multilayer perception and beyond.

Example:

Speech recognition, image recognition, and Natural Language

Processing (NLP) are all typically achieved through the use of
neural network architectures.

8
Artificial intelligence

VS

Machine learning

VS

Deep learning
9
Artificial intelligence
VS machine learning

VS deep learning

As a business decision maker, it can sometimes be difficult to understand the contrasts

between artificial intelligence, machine learning, and deep learning. Here’s a breakdown
of what differentiates the three topics so that you can gain a better understanding of how
they affect your business.

Artificial intelligence is the ability of a system to do tasks normally performed by

humans. It is an umbrella term in which both machine learning and deep learning fall
under.

Think of machine learning as a type of artificial intelligence, where its human task
is to learn from data. It is a statistical method that uses features and attributes in order
to describe an instance in a data set.

Deep learning draws inspiration from how the brain processes information. It is called
deep learning, because these models are architectures with several stacked layers on top of
each other. Just as in the human brain, deeper layers form more abstract interpretations of
the input data. This approach is especially useful when working with data without structured
attributes or features, leaving it up to the algorithm to come up with its own interpretation of
what the input represents. In comparison to most conventional machine learning techniques,
deep learning requires massive amounts of compute power, more training time, and very
large datasets. Deep learning techniques have been around for many years, but only because
of recent breakthroughs in the size of available datasets and computational resources has it
become possible to apply them to hard real-world problems.

10
 Machine Learning:
The Value it Brings
and case studies
from 3 key industries

11
 Machine learning:
The value it brings
One of the most intriguing aspects of machine learning’s value proposition is that
it doesn’t require as much advanced programming as other methods to gain
useful insights. This can help organizations cut down on time and costs, while
gaining a better understanding.

The two key pain points for organizations and developers interested in deploying
machine learning services are that they require both large amounts of raw data,
*
as well as high computing power to perform the computations needed to learn.

(and case studies from 3 key industries)

“ It's not enough just to sort of have AI capability that we

can exercise—you also need the ability to democratize it so
that every business can truly benefit from it...
That to me is our identity around AI. ”
14

– Satya Nadella

With powerful storage and data offerings, and world-class compute offerings,
Azure is a great place to start your machine learning journey to meet your
business objectives while alleviating some of the pain points that have been
associated with machine learning in the past.

*https://www.gartner.com/binaries/content/assets/events/keywords/catalyst/catus8/preparing_and_architecting_for_machine_learning.pdf
14. https://www.forbes.com/sites/bobevans1/2018/06/04/microsoft-ceo-satya-nadella-on-the-extraordinary-potential-of-ai/#986818e162ff 12
Use cases and case studies
from some key industries

To comprehend how these methods can be evangelized within your own

company, it is worthwhile to see how they are used in real-world settings. This
section will provide a few key industries that are already implementing artificial
intelligence, machine learning, deep learning, and automated machine learning.
We’ll look at a few use cases that these industries benefit from, and then dive
deeper into some companies within each industry that are utilizing Azure
Machine Learning services to meet their own business objectives.

13
 Healthcare and insurance
There are many use cases being tested in the healthcare and insurance industries.
Below, we’ll take a closer look at how leading life insurance company TAL (leading
life insurance company in Australia) used AI to improve quality assurance and the
customer experience.

Genomics and
FAST-Q
precision medicine Faster innovation
DNA BAM
SAM Single cell sequencing for drug
sequences VCF
Expression
Biomarker, genetic, variant, and population analytics development
ADAM and HAIL on Databricks

HL7/CCD
Clinical and claims data Improved
Real-world
837 Claims data warehouse outcomes and
Pharmacy Readmission predictions
analytics Registry Efficacy and comparative analytics increased
EMR Prescription adherence revenue
Market access analysis

GPU image processing

MRI Diagnostics
Image X-Ray Graphic intensive workloads leveraging
processing CT Deep learning using Tensor Flow
Ultrasound Pattern recognition machine learning

IoT device analytics Predictive

Sensor Readings
Aggregation of streaming events
analytics
Time series
data Event data Predictive maintenance transforms
Anomaly detection quality of care

Social analytics
Improved patient
Social data Social media
Real-time patient feedback via topic modeling communications
Adverse events
listening Analytics across publication data
Unstructured and feedback

14
 Healthcare and insurance

Traditionally, TAL’s quality assurance team could only review a randomly

15
selected 2-3 percent of cases. In an effort to increase the number of cases
they can review, TAL is embracing AI to improve quality assurance and the
customer experience. Using Azure Machine Learning service, the team is now 16

able to review 100 percent of cases.

“deploy
Azure Machine Learning regularly lets TAL’s data scientists
models within hours rather than weeks or months –
delivering faster outcomes and the opportunity to roll out
many more models than was previously possible. There is
nothing on the market that matches
Azure Machine Learning in this regard.
17

”
- Gregor Pacnik, Innovation Delivery Manager, TAL

15-17. https://azure.microsoft.com/en-us/blog/azure-machine-learning-service-a-look-under-the-hood/ 15
 Oil, gas, and energy
Machine learning and AI are revolutionizing the oil, gas, and energy industry.
Below, you’ll see how the global oil and gas company BP is utilizing AI and
machine learning to analyze their raw data quickly and with increased accuracy.

Upstream Digital oil field/ Faster

Field data
optimization, oil production
Asset data innovation
Demographics Production optimization
maximize Production data Integrate exploration and seismic data for revenue
well life Minimize lease operating expenses growth
Decline curve analysis

Sensor stream
Industrial IoT Improved
Grid operations,
asset inventory
data
Pipeline monitoring
outcomes and
UAV images
optimization Inventory data Preventive maintenance increased
Smart grids and microgrids
Production data
Grid operations, field service
revenue
Asset performance as a service

Supply-chain
Optimizing
Transaction data optimization
Supply-chain Demographics supply-chain
Trade monitoring, optimization
optimization Purchasing history
Retail mobile applications with machine
Trends
Vendor management - construction, transportation, learning
truck, and delivery optimization

Sensor stream
data Safety and security Predictive
Transport analytics
Risk Retail data
Grid production
Real-time anomaly detection transforms
optimization Predictive analytics
data
Industrial safety safety and
Refinery tuning
parameters
Environment health and safety security

Clickstream data Sales and marketing Improved

Products analytics customer
Recommendation Services
engine Market data
Fast marketing and multi-channel engagement engagement
Develop new products and monitor
Competitive data
acceptance of rates with machine
Demographics
Predictive energy trading learning
Deep customer insights

16
Oil, gas, and energy

BP is a global oil and gas company that operates in 70 countries. They needed
to overcome uncertainty in the exploration and production of hydrocarbons.
There are close to 200 qualitative and quantitative features that BP records to
summarize any reservoir. Knowing which of these is most important when
predicting recovery factor is a challenge. They often relied on human bias, or
other reservoirs with similar geological or geographical settings on which to
base comparison studies, but what if other features not considered here could
bring more insights to the problem and improve their predictability?

“runsWiththrough
one line of code, [automated machine learning]
different algorithms... that previously were
manually tested by the scientists. The power of the cloud
comes in here. [Azure Machine Learning service] makes data
scientists more productive, which means faster time
to market for machine learning (ML) projects. ”
-Manish Naik, Principal for digital innovation, BP

To address this problem, BP has worked with Microsoft to develop a machine

learning model that can rapidly predict recovery factor. The model is now
deployed and available globally for BP’s internal community of reservoir
engineers and geoscientists. They can use the tools to test and screen multiple
assumptions and scenarios, comparing with parallel studies to help them make
better-informed decisions faster.

17
 Retail
The retail industry is finding some interesting ways to make use of machine
learning. Find out how global retailer, ASOS, is improving their customer experience
and innovating faster.

Next best and

Customer profiles personalized offers Faster innovation
Recommendation Shopping history for a better
Online activity Customer 360/consumer personalization
engine Social network Right product, promotion, at right time customer
analysis Multichannel promotion experience

Store design and

Effective Shopping history ergonomics Improved
customer
Online activity
Path to purchase
consumer
Floor plans
engagement App data In-store experience outcomes and
Workforce optimization
increased revenue

Demand plans Data-driven stock, Omni-channel

Forecasts inventory, ordering shopping
Inventory Sales history
optimization Trends Predict inventory positions and distribution experience
Fraud detection
Local events
Market basket analysis
with machine
Weather patterns
learning

Demographics
Assortment
Buyer perception optimization Predictive
Inventory Consumer research analytics
Economic modeling
allocation Market/
Optimization for foot traffic, online interactions transforms
competitive
Flat and declining categories
analysis growth

Real-time pricing Improved

Historical sales data optimization
Price scheduling customer
Consumer
engagement
Segment level Demand elasticity engagement
price changes Personal pricing schemes
Promotion events with machine
Multi-channel engagement learning

18
 Retail

ASOS, a global fashion retailer, needs the ability to scale up and out,
innovate quickly, and incorporate new customer engagement models to
improve their customers’ experiences. Azure and Azure Machine Learning
service gave ASOS the ability to do just that.

“andIn a5,000
world where we have 85,000 products on the site
products going live each week, we need to
make sure that the right subset of those products is in front
of our consumers. Now, the products and content will
be more relevant to [shoppers].
18

”
-Bob Strudwick, Chief Technology Officer, ASOS

18. https://customers.microsoft.com/en-us/story/asos 19
Democratizing AI
with automated
machine learning

7
20
 Democratizing AI with
Automated Machine Learning
Automated machine learning makes machine learning more accessible across an
organization.19 It gives data scientists an automatic algorithm selection process and the
ability to run automated, systematic processes on raw data. There are many repetitive, tedious
and manual tasks associated with traditional machine learning. Automated machine learning
automates many of these tasks so that the training aspect is not the only automated component.
Automated machine learning can also help remove much of the risk of bias that manual model
selection poses.

19. https://medium.com/microsoftazure/democratize-artificial-intelligence-with-automated-machine-learning-169b348a9509 21
 Automated machine learning –
top objectives
Simplify and accelerate building, training, and deployment
Slow tuning times, identifying suitable algorithms, and deploying to the cloud or edge are all
challenges many developers face when deploying machine learning services. Automation allows
developers to tune hyperparameters fast and identify the right algorithm to use and deploy to
the cloud or edge with one click.20

Improve productivity and reduce costs

The typical machine learning life cycle consists of four parts: business understanding, data
acquisition, modeling, and operationalization. Automated model selection and hyperparameter
tuning reduces building time from weeks or months to days.

Automate analytical queries

Analytic queries are increasingly generated via search, natural language, or even generated
automatically. Automated analytical queries allow data scientists to stop spending time on
searches in order to focus more time on their business objectives.21

20. https://azure.microsoft.com/en-us/blog/new-automated-machine-learning-capabilities-in-azure-machine-learning-service/
21. https://searchenterpriseai.techtarget.com/feature/Automating-machine-learning-puts-analytical-models-on-autopilot 22
 Automated machine learning –
top objectives (continued)

of data science tasks will be automated by 2020,

resulting in increased productivity and broader

40%+ usage of data and analytics by data scientists,

*
according to Gartner, Inc. There are many ways that
data scientists can benefit from automated machine
learning.

22

Reduce time-consuming tasks

Many of the time-consuming machine learning tasks, such as hyperparameter tuning and
model training, testing and selection, which typically take up as much as 80 percent of a data
scientist’s time, can increasingly be automated.

23

Explore more options

It’s virtually impossible for a data scientist to test every possible solution to a machine learning
problem. Automated machine learning allows data scientists to rapidly test algorithms against
a chosen data set and find the best one to use for that problem.21

24

Measure model performance

With automated machine learning, it’s possible to evaluate the importance of model features
and measure the relationships between those features and a model’s outputs automatically. This
allows a data scientist to more quickly identify potential areas of improvement.

25
Achieve quicker time to value
Automated machine learning tools can streamline the process of choosing algorithms, running
them, and comparing the results, ultimately allowing a data scientist to put the right model into
production much sooner.

*https://www.gartner.com/en/newsroom/press-releases/2017-01-16-gartner-says-more-than-40-percent-of-data-science-tasks-will-be-automated-by-2020
22-25. https://cdn2.hubspot.net/hubfs/532045/How%20to%20Use%20Automated%20Machine%20Learning%20.pdf?t=1500943259879 23
 Azure
Machine
Learning
service

24
Azure Machine Learning service
Scale easily on the cloud and to the edge
Brings the Azure approach to data science with enterprise-ready security, compliance and
virtual network support. Users can use any data and deploy models anywhere with Azure
Kubernetes Service (AKS) and lower costs with autoscaling cloud compute.
2

Power productivity and ease of use

Users can access Azure Machine Learning service from any Python environment and use
their favorite frameworks and tools. They can leverage multi-cloud interoperability with
ONNX, rapidly build and deploy models across skill levels with automated machine learning,
and quickly build Interactive dashboards with built-in Power BI integration.

Accelerate machine learning with MLOps

MLOps is DevOps for machine learning. With machine learning pipelines, Azure Machine
Learning service brings collaboration and simplification across the end-to-end lifecycle,
from data preparation to deployment and monitoring.

25
 Azure Automated Machine Learning meets all the customer objectives listed
above along with a few other considerations:

Automated Machine Learning CPUs FPGAs

Data Professionals with all skill levels can easily use
Automated ML to identify an end-to-end machine
learning pipeline for any problem, achieving higher GPUs
accuracy while spending less time. It enables a
significantly larger number of experiments to be run,
resulting in faster iteration towards production-ready
intelligent experiences.

Bring AI to Deploy & manage

the edge models everywhere DevOps for machine learning
Increase productivity with experiment tracking,
Increase your rate model management and monitoring, integrated
of experimentation CI/CD, and machine learning pipelines.26 And with
advanced machine learning pipelines, you can
collaborate on all the steps of data preparation,
model training and evaluation, and deployment.
27

Powerful infrastructure CPUs FPGAs

Optimized for flexibility and performance. Train
models with ease and reduce costs by autoscaling
powerful FPGAs, CPU, and GPU.28 By using Azure GPUs
Machine Learning service, you can start training
on your local machine and then scale out to the
cloud. You can train models faster and scale them
up and out.

26. https://docs.microsoft.com/azure/machine-learning/service/overview-what-is-azure-ml
27-28. https://azure.microsoft.com/en-us/services/machine-learning-service/ 26
Simple deployment
Deploy models on-premises, to the cloud, and at the edge
with a few lines of code. When you have the right model, Docker Containers
you can easily deploy it in a container such as Docker. So,
it's simple to deploy to Azure Container Instances or
Azure Kubernetes Service. Alternatively, you can use the
container in your own deployments, either on-premises
or in the cloud.

Users can deploy models on-premises, to the cloud, and at

Azure Kubernetes
the edge, with a few lines of code.29 When you have the
service
right model, you can easily deploy it in a Docker container.
Models can be deployed to compute targets including:
Azure Kubernetes Service (AKS), Azure Machine Learning
Compute, Azure Container Instances, Azure IoT Edge and
Field-programmable gate array (FPGA).
30

Tool-agnostic Python SDK

Jupyter PyCharm
Azure Machine Learning service can be used in any
popular Python development environment, including
Visual Studio Code, PyCharm, and Jupyter and Azure PC
Databricks notebooks.31 Various rich tools are Visual Studio
available for each of these environments, including Code
a web portal for Azure, the Visual Studio code
extension for Azure Machine Learning services, and
widgets for Jupyter Notebooks. These tools make
it easy to explore and transform data, to then
develop and deploy models.
32

PyTorch Scikit-learn Support for open source

frameworks
Use your favorite machine learning frameworks
and tools, such as PyTorch, TensorFlow, and
TensorFlow
scikit-learn. By using the Azure Machine Learning
33

SDK for Python, along with open source Python

packages, you can build and train highly accurate
machine learning and deep-learning models
yourself in an Azure Machine Learning service
Workspace.34

29. https://azure.microsoft.com/en-us/services/machine-learning-service/
30. https://docs.microsoft.com/en-us/azure/machine-learning/service/overviewwhat-is-azure-ml
31. https://azure.microsoft.com/en-us/services/machine-learning-service/
32. https://docs.microsoft.com/en-us/azure/machine-learning/service/overview-more-machine-learning 27
33. https://azure.microsoft.com/en-us/services/machine-learning-service/
34. https://docs.microsoft.com/en-us/azure/machine-learning/service/overview-what-is-azure-ml
Azure supports experimentation by allowing users to leverage service-side
capture of run metrics, output logs, and models. Users can manage training jobs
Azure supports experimentation by allowing users to leverage service-side
locally then scale up or out. They can also use leaderboards, side-by-side run
capture of run metrics, output logs, and models. Users can manage training jobs
comparison, model selection, and conduct a hyperparameter search on ML
locally then scale up or out. They can also use leaderboards, side-by-side run
or DNN.
comparison, model selection, and conduct a hyperparameter search on ML
or DNN.
To learn more about how to collect and prepare data at scale, check out Azure
Data Factory, Azure Blob Storage, and Azure Databricks.
To learn more about how to collect and prepare data at scale, check out Azure
Data Factory, Azure Blob Storage, and Azure Databricks.

Azure supports experimentation by allowing users to leverage service-side

capture of run metrics, output logs, and models. Users can manage training jobs
locally then scale up or out. They can also use leaderboards, side-by-side run
comparison, model selection, and conduct a hyperparameter search on ML
or DNN.

To learn more about how to collect and prepare data at scale, check out Azure
Data Factory, Azure Blob Storage, and Azure Databricks.

28
 Machine learning with Azure
Machine Learning service
Data Scientists can use Azure Machine Learning service to build custom machine
earning models - train, deploy, and manage ML models at cloud scale.

Prepare Experiment Deploy

...
CODE

Prepare Build model Train & Register & Build Deploy service
your data (your favorite IDE) test model manage models image monitor model

Developers can now use Azure Machine Learning service:

• Identify suitable algorithms and tune hyperparameters faster.
35

• Improve productivity and reduce costs with autoscaling compute and DevOps for
36
machine learning.

• Seamlessly deploy to the cloud and the edge with one click.
37

• Access all these capabilities from your favorite Python environment using the
latest open source frameworks, such as PyTorch, TensorFlow, scikit-learn,
38

and MXNet.

35-38. https://azure.microsoft.com/en-us/services/machine-learning-service/ 29
 How to
get started

30
How to get started
You probably downloaded this e-book because you’re curious about machine
learning and how to get started. Now, you understand what machine learning
is; some of the customer objectives it meets; what Azure Machine Learning
service delivers; some of the top, industry, machine learning use cases; and
you’ve seen it in action through industry case studies. Here are a few options to
start utilizing machine learning on your own, or for your business:

Get started with Azure Machine Learning service today

Build, train, and deploy machine learning models with an Azure free account

Sample Notebooks
Explore Azure Machine Learning service with sample notebooks

Quickstarts
Use a cloud-based notebook server to get started with Azure Machine Learning service
Use your own notebook server to get started with Azure Machine Learning service

Tutorials
Train models with Azure Machine Learning service
Deploy models in Azure Container Instances
Prepare data for regression modeling
Use automated machine learning to build your regression model

31