Azure AI ML

Udit Goel
Solution Area Specialist
 Microsoft AI portfolio
Applications

Power Platform
Business Users
& Citizen Developers
Power BI Power Apps Power Automate Power Virtual Agents

Azure AI

Applied AI Services

Bot Service Cognitive Search Form Recognizer Video Indexer Metrics Advisor Immersive Reader
Cognitive Services

Developers &
Data Scientists
Vision Speech Language Decision OpenAI Service

ML Platform

Azure Machine Learning

 A copilot for every Microsoft Cloud experience

Microsoft 365 Copilot Dynamics 365 Copilot Copilot in Power Platform

Works alongside you in the apps Turbocharge your workforce with Imagine it, describe it, and Power
you use every day a copilot for every job role Platform builds it

Microsoft Security Copilot Windows Copilot GitHub Copilot

Defend at machine speed with The first centralized AI assistance Increase developer productivity
Microsoft Security Copilot on a platform to accelerate innovation
 Azure AI

Applied AI Services

Bot Service Cognitive Search Form Recognizer Video Indexer Metrics Advisor Immersive Reader

Cognitive Services

Vision Speech Language Decision OpenAI Service

ML Platform

Azure Machine Learning

 Azure OpenAI Service
Large pretrained foundation AI models
custom-tunable with your data

GPT-4 GPT-3.5

ChatGPT DALL·E 2
 Azure OpenAI Service on your data

Data Sources
(search, files, databases, storage etc.)

API & SDK

App or Azure OpenAI

Copilot agent Service on your data
Additional 3P Data Sources
(files, databases, storage data etc.)
 Your Azure OpenAI Service instance
is isolated from every other customer

Azure AI Your data is not used to train

the foundation AI models
AI you can trust without permission
Your data is protected by
the most comprehensive enterprise
compliance and security controls
 Introducing Azure AI Studio
Build, customize, train, evaluate, and deploy AI responsibly

Build and Ground Azure Built-in Retrieval Create AI safety

train your OpenAI Service vector augmented prompt built-in
own models and OSS models indexing generation workflows
using your data made easy
Data is the fuel that powers AI
 Microsoft Fabric
Data analytics for the era of AI
 Microsoft Fabric
Azure Synapse
Analytics
Azure Data
Factory
Microsoft
Power BI

Unified storage & compute

Unified experience
Unified governance
Unified business model
 Microsoft Fabric
Data analytics for the era of AI

Complete analytics Lake centric Empower every Copilot

platform and open business user powered
Azure ML
Customer momentum
 Azure Machine Learning
Customer promise
Azure Machine Learning is Microsoft’s ML platform that lets data scientists work how they work
best while getting models into enterprise applications quickly, reliably, safely, and cost effectively.

Accelerating the end-to-end machine learning lifecycle

Accelerate time-to-value Infuse responsible AI

Collaborate and streamline Build and enterprise-grade
with rapid model throughout the
the management of models models with confidence on
development to gain a development lifecycle to
across teams to reduce secure, hybrid, and
competitive business build fair, explainable, and
operational costs compliant platform
advantage responsible models
 Turning data into business insights
Microsoft Intelligent Data Platform
Ingest & Prepare Data Utilize & Govern

Structured Data
Govern

Purview
Open Datasets

Unstructured Data
Analytics
Azure Machine Learning

Data Bricks Synapse Power BI

All other ML scenarios;
NLP, Vision, IoT, etc.

Prepare Build Manage &

Deploy
Data & Train Monitor
 Azure
Machine Learning
Seamless studio experience
Structured Data
Business Apps
Notebooks Automated ML Designer

Responsible ML tools
Models: AKS
Comprehensive MLOps Batch predictions: SQL DB
Open Datasets

Reproducibility Automation Deployment Re-training

Unstructured Data Security and Governance Analytics & Govern

Unified management across clouds and on-premises

Azure Edge & Hybrid

All other ML scenarios; Managed Azure Arc-enabled
NLP, Vision, IoT, etc. Powerful Compute Edge/IoT Devices
(CPU, GPU, FPGA) Kubernetes Kubernetes

AI-first toolchain
Prepare Data Build & Train Deploy Manage & Monitor
 Responsible AI
State-of-the-art technology to put responsible AI principles into
practice in machine learning and build trust with end users

Development Usage Governance

Assess your machine learning Improve your confidence and Produce the responsible AI
models for responsible AI quality of data driven decision scorecard encapsulating all the
concerns during development. making with “what-if” analysis. responsible AI metrics for your
machine learning models.
Bringing together 7 tools to Apply real-life interventions and
evaluate the fairness, policies with causal analysis in Export to PDF and share your
explainability, and error analysis the responsible AI dashboard. scorecard with both technical
for debugging. and non-technical audiences like
Generate diverse counterfactual compliance officers.
explanations to inform end users.
 MLOps
The union of people, process, and platform to
enable to continuous delivery of value with Machine Learning

People Process Product

Blend together the work of individual Provide templates to bootstrap your Safely deliver features to your customers as
engineers in a repository. Work together to infrastructure and model development soon as they're ready
accelerate prototyping environment, expressed as code
Monitor your pipelines, infrastructure and
Each time you commit, your work is Model, code, environment and data are products in production and know when they
automatically built and tested, and bugs are tracked to ensure models can be reproduced aren't behaving as expected
detected faster
Automate the entire process from code
Code, data, models and training pipelines commit to production
are shared to accelerate innovation

Automation Collaboration Reproducibility

 MLOps benefits

Automation Collaboration Reproducibility

Automate repeatable Facilitate collaboration across Ensure full visibility and
actions and trigger business team through easy exchange reproducibility into the end-to-
process that relate to an and handoff of assets end ML lifecycle in your team
AI-infused application

== VELOCITY and SECURITY (For ML)

 Secure, hybrid, and compliant platform

Identity and access control Network isolation Data protection

User and system managed All traffic on virtual network Data encryption in transit
identity for workspace with with no public IP using TLS1.2 and at rest
minimum required permissions
Control inbound and outbound Encryption with system managed or
AAD conditional authentication access of your environment customer managed keys
with RBAC control for custom policies

Policies and monitoring Planning and deployment Platform trust

Azure Monitor and Guidance for security and scale Customer Lockbox to approve or reject
Security Center integration customer data requests
Infrastructure-as-code templates and
Azure Policy to enforce landing zone reference implementations Transparency on OS and container updates
configuration settings and vulnerability management
Train and deploy models in
Usage and cost reporting cloud or on the edge
 Machine Learning end-to-end lifecycle

Data Model Model Model Model

preparation building training & deployment management &
assessment monitoring

Responsible AI
MLOps

ML Platform (Secure, Compliant, Hybrid)

AI will reinvent every industry
 More effective patient treatment plans
Personal Accurate drug development and discovery
Healthcare Precise medical imaging and diagnoses

Protect loan/credit privacy with sharable data

Safe Detect suspicious and fraudulent transactions
Finance Turn data trends into risk assessments

AI will reinvent
every industry Sustainable Tailor inventory and pricing based on demand

Retail Authentic virtual try-ons and less returns

Generative product design and style transfers

Efficient Optimize inventory tracing in the supply chain

Manufacturing Amplify humans with collaborative cobots

Navigate the inner complexity of machines
 Microsoft Cloud
Industry partner ecosystem
Get started today
Azure Machine Learning:

https://azure.com/ml

Start building with a free trial

https://azure.microsoft.com/en-
us/free/sql-data-warehouse/
Thank You!
 Improve time-to-value with
rapid innovation and integrated tools
Experiencing <1% error rates in predictions
for package pickup cycles Read story here • Accelerate model creation
with our AutoML
Challenge • Increase team efficiency
Customers expect timely and accurate information on their packages. Delivery services
giant FedEx wants to gain better information on its packages, provide real-time • Access flexible
visibility to its customers, and reduce its late-delivery fees. infrastructure
Solution • Start local and
FedEx used Azure Machine Learning across cloud and on-premises to build scale to cloud
containerized solution with Azure ML to develop and deploy models for packages
prediction using MLOps. • Build your way with any
framework or IDE
Impact
• The only platform with
The Company reduced customer service calls, improved real time visibility, improved Pytorch enterprise
visibility for millions of packages FedEx achieved an error rate of <1% at the height of
their pickup cycle.
 Operationalize at
scale with MLOps
Realized time savings in training image models and
operationalizing with just a few clicks Read story here
• Reuse reproducible
pipelines
Challenge
To keep pace with increasing customer expectations, BMW wanted to optimize
• End to End tracking
production lines to increase flexibility and drive quality control for different of models
components on the same production line.
• Code re-use through
Solution Automated ML and
BMW implemented an innovative solution combining industrial robotics, computer reusable components
vision and Azure ML based learning models. The movements of robotic arm are tailored
for quality control based on ML models. MLOps is used for their RCV workflows, image • Deploy models anywhere
annotation, model training, deployment and inferencing.
• Monitor end-to-end cycles
Impact
Achieved shorter turnaround times and increased rate of components traversing
through production lines. BMW realized time saving in training image models and
operationalizing them with just a few clicks.
 Develop on a secure, hybrid,
and compliant platform
Improved predictions by more than 40%
for product recommendations Read story here
• Ease of secure setup using templates
(Terraform, ARM, Bicep)
Challenge
• Virtual networks and end to end
PepsiCo wanted to generate insights on knowledge of the industry/the market, in Private IP addresses
relation to consumers shopping habits and their preferences. Dealing with large
datasets across all stores and locations was a challenge to overcome. • Role-based access control and
managed identity support
Solution
• Integration with Azure Policy, Azure
PepsiCo developed an application, StoreDNA, moving voluminous data to Azure -
Monitor, and Azure Security Center
breaking data silos and built models to provide prioritized list of top 3 product actions
per locations. PepsiCo met regulatory compliance while deploying models to cloud with • Data encryption in transit and at rest
MLOps and CI/CD pipeline.
• Control Costs
Impact
• Over 60 compliance certifications
PepsiCo acted on more than 85 percent of the application’s recommendations, and
the company determined that it was improving predictions by more than 40 percent.
Automation with MLOps removed weeks of work from ML projects and cut time
spent on routine tasks, improved accuracy and repeatability of models.
 Develop & Deploy
responsible solutions
Reduced disparity from 7% to 0.5% between
men and women in loan decisions Read story here

Challenge • Responsible AI dashboard

EY wanted to understand their customers better, identify fraud and security breaches sooner,
and make fairer loan decisions faster. Driving model transparency was also a challenge.
• Explaining model outcomes
• Mitigate model fairness
Solution
• Analyze root cause of errors
EY used Fairlearn to assess a models’ fairness by looking at its performance across
different demographics and retraining as appropriate. EY improved model • Reproduceable and
interpretability using InterpretML toolkit to drive model transparency.
automated workflows
to support
Impact
EY mitigated model disparity of 7% between men and women in the accuracy of
approving/denying loans. Post mitigation, the disparity was less than 0.5%.
 Machine Learning the REAL end-to-end lifecycle

Managing
V A N C E P R O J E CT PRE
P ARE MOD EL
AD

lect online metrics

Col
e Clo Va
rat ud
BUSINESS
abo tr

lid
NEED
l

ate
l

ai
co

nin
Hypo

te
d
the

on
an

ida
e si
at

te

line
Deploy cand
Itera
alu

sG
eriment and ev

ath

performance
er
and prep da
Exp

ta

lt s
Va
lid

esu
Bu
ild o at
r a c q u ir e

er
e

at
M

of
a

lu
Ev o

flin
e

ni
to
r
e
Pro m ot

REV RETR
ERT PR OJECT AIN MO D E L

Prototyping Training Operationalizing

MLOps
Responsible AI