Agenda
Scaled out AI Platform with Search (Bing and
Ads)
AI Platform is now used across Microsoft
Microsoft needs and drives Open and
Interoperable AI
Azure AI Services
Several Billion dollar AI businesses at
Microsoft
Learned a ton as we built them
Bing Ads Story
My AI Story
Step 3: Profit
Eric joins
MSFT
How
Bing Ads Execution
 Shipped once every 6 months
 3 experiments / month
 Big bets that didn’t work.
 Created dev teams with core metrics
 Pushed teams to move their metric
 Developed AI Infrastructure to do
rapid experimentation.
 Over a few years made >300x
improvement (1000/mo expr)
Experiment – 90% of ideas fail
Iterate – faster you try things, the more successful ideas you have
Build Infrastructure to enable fast iteration of experiments
RPM Gains
Search Scenario Evolution
Keyword Based Search
Natural Language Search
Voice, Vision, Context-Based Search
How to ensure best results?
 Vector-based search:
 Word is represented as vector.
 Vector captures word meaning – its semantics.
 Words with similar meanings get similar vectors.
 In Bing, we trained a GloVe model over
Bing’s query logs and generated 300-
dimensional vector, enabled by deep
neural network.
Deep Learning: Semantic Understanding
Vector-based Retrieval
 DL-generated vectors semantically represent queries, documents and passages
 Doc retrieval based on query-doc-passage semantic similarity (vector distance)
AND
long
does
…
canned
how
Query: {how long
does a canned
soda last}
canned
does
how
long
…
Posting 1
Posting 2
Posting 3
Posting 4
…
Matching
BM25
Perfect Match
Sem. Similarity
Vector
Recall
…
(0.78, 0.8, 0.4, 0.3, …)
(0.75, 0.6, 0.1, 0.8, …)
…
Approximate Nearest Neighbor
Search (ANN)
RANKING STACK
Bing AI Platform – Bing QnA
DATA PREP BUILD TRAIN DEPLOY INTELLIGENT APPS
ACTIONINTELLIGENCEDATA
Stored on
Cosmos
Æther with
TensorFlow
Philly
Ultra-fast Inferencing
using FPGAs
Youngji Kim
Principal Program Manager
Bing Relevance and AI
AI Expands Beyond Bing
AI Transforms all businesses across Microsoft
 AI is now done in almost every team
 Products that don’t seem to obviously use AI are powered by it
 Using the same tools and infrastructure to accelerate new teams
Took some time to warm up to the ideas…
To: Eric Boyd
From: xxxxx
Subject: ABC Experiment
…we have decided to ship the
XXX feature without running
the experiment first. We’re
quite confident in the
feature.
Our teams have a rich history
of shipping features without
any experimentation, something
that may be alien for folks
from Bing.
To: xxxxx
From: Eric Boyd
Subject: Re: ABC Experiment
I have a rich history of
driving my car with my eyes
closed. I’m quite confident I
know the way to work and have
only crashed a few times.
0
5k
10k
15k
Dailykeptcount
Rule-based Machine learning
Models trained on
exploration data
Filter using user
interactions
Offline model
Machine Learning for SmartArt Suggestions
0
5
10
15
20
25
30
35
40
45
1-Apr 15-Apr 29-Apr 13-May 27-May 10-Jun 24-Jun 8-Jul 22-Jul 5-Aug 19-Aug 2-Sep 16-Sep 30-Sep 14-Oct 28-Oct 11-Nov 25-Nov
Slidekeptrate
Machine Learning for SmartArt Suggestions
0
50k
100k
150k
Dailykeptcount
Rule-based Machine learning
0
5
10
15
20
25
30
35
40
45
1-Apr 15-Apr 29-Apr 13-May 27-May 10-Jun 24-Jun 8-Jul 22-Jul 5-Aug 19-Aug 2-Sep 16-Sep 30-Sep 14-Oct 28-Oct 11-Nov 25-Nov
Slidekeptrate
More SmartArts,
trained on all data
Offline model
Filter using user
interactions
Models trained on
exploration data
Office – PowerPoint Designer
DATA PREP BUILD & TRAIN DEPLOY INTELLIGENT APPS
ACTIONINTELLIGENCEDATA
Stored on
Cosmos
Æther with
TLC (ML.NET)
Power Point
User Behavior
Data prep using
Azure Databricks
osi
Anand Balachandran
Principal Program Manager
Office Intelligent Services
Every business at Microsoft investing in AI
+ more!
 1000s of Data Scientists & AI Developers
 Classical ML & Deep Learning
 Compliant & non-compliant data
 Internal & OSS frameworks
 Internal & OSS tools & languages
 Deployments to Azure, AP, Windows, Mobile, Edge
Building your own AI models for Transforming Data into Intelligence
Prepare Data Build & Train Deploy
8 Exabytes
30M events/sec
From 1B+ devices
Cosmos
Step 1: Prepare Data
GDPR compliant cloud data processing
system for near real-time ingestion &
processing
System for managing ML pipelines
Optimized for rapid prototyping, data reuse,
and collaboration
>20M Pipelines
>1.75M active Datasources (>2.7 total)
Æther
Step 2: Build and Train
DLIS
Run multiple models in parallel
ML, DL, Transforms & Featurizers
Abstracting platform details
(CPU/GPU/FPGA)
600K req/sec
<35ms latency
Æther
Step 3: Deploy
Hyperparameter Tuning
 As the models get more complex, the
space of hyperparameters to explore
increases exponentially!
 Grid or Random Search becomes too
costly
 Bayesian Optimization is used to optimize
Acquisition Functions
AI Platform
investments
Built an exabyte-scale data lake for everyone
to put their data of all types (structured and
unstructured)
Built AI tools approachable by any developer
Built machine learning tools for collaborating
across large experiment models
Summary
Project Kinect for Azure
How do we get from raw point clouds to cloud intelligence?
Wall
Floor
Ceiling
Table
Chair
Background
Project Kinect for Azure
Depth and Deep Learning
Using the power of our AI tools and infrastructure, we can take that raw output and train a
model capable of high fidelity environment perception.
Input: Project
Kinect for Azure
Raw Depth +
Active Brightness data Labelling Tool
Labelled
“Ground Truth”
Training Set Test Set
Philly
CNTK model
Analytics Client
ONNX model
Michelle Brook
Program Manager II
AI Perception and Mixed Reality
Labelled “Ground
Truth”
Raw Depth + Active
Brightness data
Labelling Tool Labelled “Ground
Truth”
Building the Training Data – How do we scale?
TABLE
Caffe2 | TensorFlow | PyTorch | CNTK Framework Backends
Hardware-Backed Libraries
e.g. CPU, GPU, FPGA
Open
ecosystem
for interoperable
AI models
Tools Should
Work Together.
ONNX enables models to be trained in one framework and transferred to another for inference.
CPUGPU
ML HW
DSPFPGA
High level API &
Framework Frontends
Hardware Vendor
Libraries & Devices
Any tools exporting ONNX models can benefit ONNX-compatible runtimes and libraries designed
to maximize performance on some of the best hardware in the industry.
Seamless Interoperability
ONNX.ai
Partners
Contribute
Get Involved
github.com/onnx
ONNX is a community project. We
encourage you to join the effort
and contribute feedback, ideas,
and code. Join us on Github.
Use ONNX
ONNX.ai
Start experimenting today. Check
out our Getting Started Guide,
Supported Tools, and Tutorials.
Follow Us
Stay up to date with the latest
ONNX news.
onnxai
onnxai
Empowering Physicians with Medical ImagingAI
Data Scientists Developers
Intelligent Disease Prediction
Data Prep Build Train Deploy Intelligent Apps
Azure Machine
Learning
IoT Hub
WindowsML
IOT Edge
Model:
DenseNet-121
Code:
Keras +
TensorFlow
National Institute
of Health
Chest Xray Data
112,120 images
14 pathology labels
30,805 unique patients
Visual Studio
Tools for AI
Azure
Deep Learning GPU VM
VSTS +
CI/CD
CosmosDB +
Azure Functions
NuGet
Chris Lauren
Principal Program Manager
AI Platform
With , you don’t have to be the
size of Bing to solve the problem
Bringing the best of AI to Azure and the best of Azure to AI
Pre-Build AI
Azure Cognitive Services
Conversational AI
Azure Bot Services
Custom AI
Azure Machine Learning
Integrated with Azure Machine Learning
Create new deep learning projects easily
Scale Out with Azure Batch AI
Monitor model training progress & GPU utilization
Visualize your model processing with integrated open
tools like TensorBoard
Get started quickly with the Samples Gallery
Productive AI developer tools to train
models and infuse AI into your apps
The Machine Learning framework made by and for .NET developers
Proven & Extensible
Open Source
Supported on Windows, Linux, and macOS
Developer Focused
Join at github.com/dotnet/machinelearning
Customizable
ML.NET Preview
Cross-platform open Source Machine learning framework for .NET
Extensively used across Microsoft: Windows, Bing, Azure
High productivity for complete workflow
Extensible to other frameworks (TensorFlow, CNTK…)
Microsoft scale Tools/Services available/coming to you
 Cognitive Toolkit (CNTK)
 ML.Net
 ONNX
 Azure Batch AI
 Hyper Parameter Tuning in Azure ML
 Project BrainWave (FPGA model acceleration) in Azure ML
 Visual Studio Tools for AI
And more coming soon…
SeeingAI.com
seeingai@microsoft.com
 Cloud data processing system
 Near real time ingestion
 Near real time processing
 Fully GDPR compliant
 8 Exabytes
 30M events/sec
 From 1B+ devices
COSMOS Æther
 System for managing ML
pipelines
 Optimized for rapid
prototyping, data reuse, and
collaboration
• >20M Pipelines
• >1.75M active Datasources
(>2.7 total)
Data Prep -> Build/Train -> Deploy
DLIS
 Run multiple models in
parallel
 ML, DL, Transforms &
Featurizers
 Abstracting platform details
(CPU/GPU/FPGA)
 600K req/sec
 <35ms latency
AI @ Microsoft, How we do it and how you can too!
AI @ Microsoft, How we do it and how you can too!

AI @ Microsoft, How we do it and how you can too!

  • 3.
    Agenda Scaled out AIPlatform with Search (Bing and Ads) AI Platform is now used across Microsoft Microsoft needs and drives Open and Interoperable AI Azure AI Services
  • 4.
    Several Billion dollarAI businesses at Microsoft Learned a ton as we built them Bing Ads Story My AI Story
  • 5.
  • 6.
  • 7.
    Bing Ads Execution Shipped once every 6 months  3 experiments / month  Big bets that didn’t work.  Created dev teams with core metrics  Pushed teams to move their metric  Developed AI Infrastructure to do rapid experimentation.  Over a few years made >300x improvement (1000/mo expr)
  • 8.
    Experiment – 90%of ideas fail Iterate – faster you try things, the more successful ideas you have Build Infrastructure to enable fast iteration of experiments
  • 10.
  • 12.
    Search Scenario Evolution KeywordBased Search Natural Language Search Voice, Vision, Context-Based Search
  • 13.
    How to ensurebest results?
  • 14.
     Vector-based search: Word is represented as vector.  Vector captures word meaning – its semantics.  Words with similar meanings get similar vectors.  In Bing, we trained a GloVe model over Bing’s query logs and generated 300- dimensional vector, enabled by deep neural network. Deep Learning: Semantic Understanding
  • 15.
    Vector-based Retrieval  DL-generatedvectors semantically represent queries, documents and passages  Doc retrieval based on query-doc-passage semantic similarity (vector distance) AND long does … canned how Query: {how long does a canned soda last} canned does how long … Posting 1 Posting 2 Posting 3 Posting 4 … Matching BM25 Perfect Match Sem. Similarity Vector Recall … (0.78, 0.8, 0.4, 0.3, …) (0.75, 0.6, 0.1, 0.8, …) … Approximate Nearest Neighbor Search (ANN) RANKING STACK
  • 17.
    Bing AI Platform– Bing QnA DATA PREP BUILD TRAIN DEPLOY INTELLIGENT APPS ACTIONINTELLIGENCEDATA Stored on Cosmos Æther with TensorFlow Philly Ultra-fast Inferencing using FPGAs
  • 18.
    Youngji Kim Principal ProgramManager Bing Relevance and AI
  • 20.
  • 21.
    AI Transforms allbusinesses across Microsoft  AI is now done in almost every team  Products that don’t seem to obviously use AI are powered by it  Using the same tools and infrastructure to accelerate new teams
  • 22.
    Took some timeto warm up to the ideas… To: Eric Boyd From: xxxxx Subject: ABC Experiment …we have decided to ship the XXX feature without running the experiment first. We’re quite confident in the feature. Our teams have a rich history of shipping features without any experimentation, something that may be alien for folks from Bing. To: xxxxx From: Eric Boyd Subject: Re: ABC Experiment I have a rich history of driving my car with my eyes closed. I’m quite confident I know the way to work and have only crashed a few times.
  • 25.
    0 5k 10k 15k Dailykeptcount Rule-based Machine learning Modelstrained on exploration data Filter using user interactions Offline model Machine Learning for SmartArt Suggestions 0 5 10 15 20 25 30 35 40 45 1-Apr 15-Apr 29-Apr 13-May 27-May 10-Jun 24-Jun 8-Jul 22-Jul 5-Aug 19-Aug 2-Sep 16-Sep 30-Sep 14-Oct 28-Oct 11-Nov 25-Nov Slidekeptrate
  • 26.
    Machine Learning forSmartArt Suggestions 0 50k 100k 150k Dailykeptcount Rule-based Machine learning 0 5 10 15 20 25 30 35 40 45 1-Apr 15-Apr 29-Apr 13-May 27-May 10-Jun 24-Jun 8-Jul 22-Jul 5-Aug 19-Aug 2-Sep 16-Sep 30-Sep 14-Oct 28-Oct 11-Nov 25-Nov Slidekeptrate More SmartArts, trained on all data Offline model Filter using user interactions Models trained on exploration data
  • 27.
    Office – PowerPointDesigner DATA PREP BUILD & TRAIN DEPLOY INTELLIGENT APPS ACTIONINTELLIGENCEDATA Stored on Cosmos Æther with TLC (ML.NET) Power Point User Behavior Data prep using Azure Databricks osi
  • 28.
    Anand Balachandran Principal ProgramManager Office Intelligent Services
  • 31.
    Every business atMicrosoft investing in AI + more!  1000s of Data Scientists & AI Developers  Classical ML & Deep Learning  Compliant & non-compliant data  Internal & OSS frameworks  Internal & OSS tools & languages  Deployments to Azure, AP, Windows, Mobile, Edge
  • 32.
    Building your ownAI models for Transforming Data into Intelligence Prepare Data Build & Train Deploy
  • 33.
    8 Exabytes 30M events/sec From1B+ devices Cosmos Step 1: Prepare Data GDPR compliant cloud data processing system for near real-time ingestion & processing
  • 34.
    System for managingML pipelines Optimized for rapid prototyping, data reuse, and collaboration >20M Pipelines >1.75M active Datasources (>2.7 total) Æther Step 2: Build and Train
  • 35.
    DLIS Run multiple modelsin parallel ML, DL, Transforms & Featurizers Abstracting platform details (CPU/GPU/FPGA) 600K req/sec <35ms latency Æther Step 3: Deploy
  • 36.
    Hyperparameter Tuning  Asthe models get more complex, the space of hyperparameters to explore increases exponentially!  Grid or Random Search becomes too costly  Bayesian Optimization is used to optimize Acquisition Functions
  • 37.
    AI Platform investments Built anexabyte-scale data lake for everyone to put their data of all types (structured and unstructured) Built AI tools approachable by any developer Built machine learning tools for collaborating across large experiment models Summary
  • 39.
  • 40.
    How do weget from raw point clouds to cloud intelligence? Wall Floor Ceiling Table Chair Background Project Kinect for Azure
  • 41.
    Depth and DeepLearning Using the power of our AI tools and infrastructure, we can take that raw output and train a model capable of high fidelity environment perception. Input: Project Kinect for Azure Raw Depth + Active Brightness data Labelling Tool Labelled “Ground Truth” Training Set Test Set Philly CNTK model Analytics Client ONNX model
  • 42.
    Michelle Brook Program ManagerII AI Perception and Mixed Reality
  • 43.
    Labelled “Ground Truth” Raw Depth+ Active Brightness data Labelling Tool Labelled “Ground Truth” Building the Training Data – How do we scale? TABLE
  • 46.
    Caffe2 | TensorFlow| PyTorch | CNTK Framework Backends Hardware-Backed Libraries e.g. CPU, GPU, FPGA
  • 49.
  • 50.
    ONNX enables modelsto be trained in one framework and transferred to another for inference. CPUGPU ML HW DSPFPGA High level API & Framework Frontends Hardware Vendor Libraries & Devices Any tools exporting ONNX models can benefit ONNX-compatible runtimes and libraries designed to maximize performance on some of the best hardware in the industry. Seamless Interoperability ONNX.ai
  • 51.
  • 52.
    Contribute Get Involved github.com/onnx ONNX isa community project. We encourage you to join the effort and contribute feedback, ideas, and code. Join us on Github. Use ONNX ONNX.ai Start experimenting today. Check out our Getting Started Guide, Supported Tools, and Tutorials. Follow Us Stay up to date with the latest ONNX news. onnxai onnxai
  • 54.
    Empowering Physicians withMedical ImagingAI
  • 55.
  • 57.
    Intelligent Disease Prediction DataPrep Build Train Deploy Intelligent Apps Azure Machine Learning IoT Hub WindowsML IOT Edge Model: DenseNet-121 Code: Keras + TensorFlow National Institute of Health Chest Xray Data 112,120 images 14 pathology labels 30,805 unique patients Visual Studio Tools for AI Azure Deep Learning GPU VM VSTS + CI/CD CosmosDB + Azure Functions NuGet
  • 58.
    Chris Lauren Principal ProgramManager AI Platform
  • 60.
    With , youdon’t have to be the size of Bing to solve the problem
  • 61.
    Bringing the bestof AI to Azure and the best of Azure to AI Pre-Build AI Azure Cognitive Services Conversational AI Azure Bot Services Custom AI Azure Machine Learning
  • 62.
    Integrated with AzureMachine Learning Create new deep learning projects easily Scale Out with Azure Batch AI Monitor model training progress & GPU utilization Visualize your model processing with integrated open tools like TensorBoard Get started quickly with the Samples Gallery Productive AI developer tools to train models and infuse AI into your apps
  • 63.
    The Machine Learningframework made by and for .NET developers Proven & Extensible Open Source Supported on Windows, Linux, and macOS Developer Focused Join at github.com/dotnet/machinelearning Customizable ML.NET Preview Cross-platform open Source Machine learning framework for .NET Extensively used across Microsoft: Windows, Bing, Azure High productivity for complete workflow Extensible to other frameworks (TensorFlow, CNTK…)
  • 64.
    Microsoft scale Tools/Servicesavailable/coming to you  Cognitive Toolkit (CNTK)  ML.Net  ONNX  Azure Batch AI  Hyper Parameter Tuning in Azure ML  Project BrainWave (FPGA model acceleration) in Azure ML  Visual Studio Tools for AI And more coming soon…
  • 67.
  • 71.
     Cloud dataprocessing system  Near real time ingestion  Near real time processing  Fully GDPR compliant  8 Exabytes  30M events/sec  From 1B+ devices COSMOS Æther  System for managing ML pipelines  Optimized for rapid prototyping, data reuse, and collaboration • >20M Pipelines • >1.75M active Datasources (>2.7 total) Data Prep -> Build/Train -> Deploy DLIS  Run multiple models in parallel  ML, DL, Transforms & Featurizers  Abstracting platform details (CPU/GPU/FPGA)  600K req/sec  <35ms latency