Tendencias en Inteligencia Artifiial y Maquinas

Gartner Data & Analytics Summit
Summit 2017
20 – 22 March 2017 / London, UK
Key Trends in Artificial Intelligence and

Machine Learning
Alexander Linden
CONFIDENTIAL AND PROPRIETARY

This presentation, including any supporting materials, is owned by Gartner, Inc. and/or its affiliates and is for the sole use of the intended Gartner audience or other intended recipients. This presentation may contain
information that is confidential, proprietary or otherwise legally protected, and it may not be further copied, distributed or publicly displayed without the express written permission of Gartner, Inc. or its affiliates.
© 2017 Gartner, Inc. and/or its affiliates. All rights reserved.
Artificial
Intelligence Technical discipline:
Solves business problems
through the extraction of
A sociotechnical knowledge from data.
construct:
Machine capabilities which Deep Learning

solve complex tasks that were
recently only possible by Machine Learning
humans (equally well or better)
Data Science
Predictive Analytics
Big Data Data

Advanced Mining IDA
Analytics
1 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

We Are an AI Company, an ML Company or Even AI-First

MEGATREND: Machine Learning Implies a Paradigm Shift
in Problem Solving
Classical Engineering
Seed
Weather and time
Soil Farmer Food
ML algorithms
Farming HPC ML solutions
Data
Data scientists

Extreme Progress in Just Five Years
MSFT's ImageNet Solution

AlphaGo
Google's NMT
Self-driving cars Baidu’s
in the desert Speech
Deep learning Recognition
Computer Self-driving Train drivers Solutions
beats top cars in (ImageNet, replaced by robots
chess normal Kaggle)
player traffic
1997 ... 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016
IBM Watson beats Tesla has over a

Jeopardy experts billion miles of
autonomous
driving data

Dramatic Progress: ... and Computers Opened Their Eyes ...
Misclassification Rate
30%
25%
20%
15%
Krizhevsky and Others,
University of Toronto
10%
Late 2014 Through 2016
5%
0%
2010 2011 2012 2013 2014 Baidu Microsoft Google … Microsoft Human
ImageNet Challenge (1,000 Categories and 1.2 Million Image Subset of 15 Million Image Dataset)

AI Misconceptions ...
"Watson is a cognitive technology

that can think like a human." lead to unproductive and
Quote from https://www.ibm.com/watson unrealistic fears:
captured 4 March 2017
"AI will take

85% of jobs
or kill us all"
Fearing a rise of killer robots is like

worrying about overpopulation on Mars
— Andrew Ng

If the human brain were so simple,
that we could understand it, we would
be so simple, that we could not.
— Emerson Pugh
By BruceBlaus — Own work, CC-BY-3.0,

https://commons.wikimedia.org/w/index.php?curid=28761830
Amazing Innovation
Broad AI Remains a Fantasy for >20 Years …
Key Issues
1. What is machine learning and AI?

2. What are the benefits?
3. What are risks, limitations and how to go from here?

Key Issues


… at the Core, ML Is About Creating Mappings
Informed by Input/Output Pairs
Type of Problem Inputs Outputs
Loan Application Application data Will the applicant repay the loan? (0 or 1)
Demand Prediction Market situation How many products will be bought? (n)
Self-Driving Cars Car sensory data Break, accelerate, tilt the wheel? (x, y)
Propensity to Buy Profile and transactions Will the customer buy or not? (0 or 1)
Failure Prediction Sensor readings Will a failure happen with 4 weeks (0 or 1)
Customer Churn Profile and activities Will customer cancel the contract? (0 or 1)
Medical Diagnosis Pixel data from a retinal scan Will the disease break out? (0 or 1)
Advertisement Ad + context + user profile Will the user click on ad? (0 or 1)

Neural Nets
1 b
y = a * x + b + error ∑
a
x
1 b
f(∑)
a
x
Sigmoid function
Intermediate presentation

Deep Learning Visualized: http://playground.tensorflow.org

There Is a Zoo of Machine-Learning Approaches Out There ...
Compute
Requirements
Extremely popular
Log Chart
Very popular
k-NN 20 years Less popular
Occasional
Random Forest
SVMs Two-Layer Neural Nets
Decision Trees
Scorecards
Simple Regression Depth
Deep learning attempts to improve machine learning, by
the creation of intermediate representations of the data.

Example: Face Recognition
Basic Complex
Edges shapes shapes
TASK: Does this

image contain a face?
Increasing abstraction through intermediate layers
Output: 0 or 1
Input: Raw pixel data

High-Performance Computing Often Required
10-Node Small GPU Large GPU Largest GPU

Desktop Spark Cluster Board Cluster Cluster
1. Face Recognition 1 year 4 months 4 weeks 4-5 days 12 hours
2. Demand Prediction 3 weeks 4-5 days 12 hours 3 hours 30 minutes
3. Machine Translation 20 years 2-4 weeks

Why Now? Virtuous Cycle
High-Performance
Deep
Compute Learning
Complexity
What Next? Algorithms
Data
Systems Virtuous Cycle

and Sensors

Key Issues


Deep Learning Works Best in Application Domains ...
Most data
The more
elements
data sources
have almost
the better
no meaning
Little or no
domain
knowledge
available
BUT: Don't write off shallow machine learning yet!

By 2019, deep
learning will provide
best-in-class
performance for
demand, fraud and
failure prediction

"Using deep learning ...
PayPal has cut its false-
alarm rate in half."
Hui Wang, Sr. Director of Global Risk Sciences, PayPal
in American Banker, 1 September 2016

Deep Learning Addresses One of the Biggest "Big Data"
Challenges: Data Fusion
Data
Weather
Microeconomics
Increasing abstraction via intermediate computational layers
Macroeconomics
Demand
Social
Transactions
Interactions
Traffic

Deep learning
moves the burden
from data
preparation to
network
architecture
selection
and (currently) to extreme computing requirements

Deep learning enables extremely rich content analytics and
motor control
Previously
Now and soon
... also at
the output layer

Image
O
Audio
U
T
P
Text
U
T
S Tabular/ Classical Data Science
Structured Business Analytics
Tabular Text Audio Image Video Diverse
INPUTS

Image Visual Search
O
Audio
U
T
P
Text
U
T
INPUTS

Give me more like this
DCGAN (Facebook, 2015)

Visual Search
Image Image Filtering
O
Audio
U
T
P
Text
U
T
INPUTS

Content Synthesis
and also Manuel Ruder and Thomas Brox (Uni Freiburg)

Works for Videos and also Manuel Ruder and Thomas Brox (Uni Freiburg)
Credit to https://deepart.io (Germany)

Content Synthesis
Input A Input B
INPUTS OUTPUT

Visual Search
Super Resolution
O
Audio
U
T
P
Text
U
T
INPUTS

Super Resolution
Original Low-Resolution Image First Iteration Second Iteration
Try yourself at http://waifu2x.udp.jp

Visual Search
Super Resolution
O
Audio
U
T
P
Text Machine Translation
U
T
INPUTS

Google Neural Machine Translation
Source: http://nlp.stanford.edu/projects/nmt/Luong-Cho-Manning-NMT-ACL2016-v4.pdf
Wake-up call!
Text analytics + Google NMT reported

error reduction
chatbots rates of 30% to 80%

Image Search Visual Search
Image Synthetic animation Image Filtering
Image Generation Super Resolution
Visual Q&A
O
Voice Synthesis Lip Reading
Audio Real-Time Translation
U
Speech Imitation
T
Text Creation Image Captioning
P NLG
Text Machine Translation Speech Recognition OCR, ICR
U
Chatbots Speak Proofreading
T Image Recognition Surveillance
Classical Data Science Chatbots Listen Robotics
S Tabular/ Speech UI Visual Search Surveillance
Information Extraction Self-Driving Cars
Medical Diagnosis Scene Classification
POS Tagging Video Analytics Compression
INPUTS

Key Issues


Still Early ...
... too expensive

... mostly conducted in isolation
Projects are yet ... not agile and collaborative enough
... just utilizing maximum 3-4 data sources
Tools ...
... have often insufficient UIs (too old — or too juvenile)
... lack in collaboration feature
... lack in cross-platform model management
... are yet either pure cloud or no elastic scale or pure on-premises

Current Data Science Platforms Are Not (Yet) Satisfactory
From "Magic Quadrant for Data Science Platforms,"

14 February 2017 (G00301536)
Landscape of ML Solutions MAKE
Skymind's Deeplearning4j
Salesforce Einstein
Caffe
SAP Clea Google's TensorFlow Theano
Microsoft Cognitive Toolkit
Business H2O.ai's Deep Water
Users Application Baidu’s Pebble Intel BigDL
Engineers Amazon Web Services’
Apache MXNet
Embedded Machine-Learning
Machine Learning APIs ML
From "Magic Quadrant for Data Science Platforms," 14 February 2017 (G00301536)
Engineers
Data R, Python,
Scientists Data Science Scala, Matlab
Data Smart Data Discovery Platforms
Analysts Deep-Learning
Frameworks
Data Analysis
Software Deep-Learning
Intel's Nervana Cloud Platforms
Azure Deep-Learning Hardware
Rescale
AWS
Nvidia, AMD, IBM, Intel
BUY
Google Cloud Platform
Deep-Learning Adoption Patterns
 20-30 firms
 Own tooling
Skilled Inexperienced
End Users End Users
Tools and APIs Only APIs
 3-5K firms  Most Gartner clients
 Have data scientists  Have no data scientists on staff
 Let others come up with solutions!

APIs, SaaS, APIs, SaaS,  Evaluate solutions — understand costs
packaged apps and Tools and APIs packaged apps and
custom-made solutions custom-made solutions
 800-1000 organizations
Advanced  End users use tools for strategic advantage
 Service providers create custom-made solutions
 IT vendors create APIs, SaaS or packaged apps

ML Projects Fail ...
Complexity of the pipeline/stack
Insufficient:
 Data
 Domain insights
 Compute infrastructure
ML task can be very complex
 Skills
 Luck Simply unknown if "good enough" solutions exist

Insufficient Data
Debugging is different to
classical software engineering
Too dirty Too little
Inconsistent Too biased
Incomplete Too dynamic

Too Little Data
No. of Data No.

Collect more Points Needed
2
Variables
1
4 2
8 3
16 4
256 8
Syndicate data 65.536 16
~ 4 billion 32
~ 16 quadrillion 64
Generate data Utilize partial solutions

from simulations (deep features, transfer learning)

Biased Data
250
200
150
100
50
0
?
5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85
Analyze Data science audits Document

ML Projects Hidden Costs: Dynamic Data
Out-of-Sample
In-Sample
Failure
Past NOW Future

Training
Projects can become

VERY expensive

Boundary Between Impossible,
Doable and Must-Have

Scratching the AI Boundary
RISK
Tactical Decisions
High
 >5 seconds ... min.
 Nonroutine analysis
 Not simple I/O
FAQs, new product
Medium description, new features, Strategic Decisions
Operational Decisions price changes, chats  Unstructured
 Routine  Many steps
 Less than 2 seconds  Nonroutine activities
 Simple I/O New product lines, new
Low web design, acquisitions,
Cross-selling, image business plans
recognition, marketing,
retention, management, task
assignment, failure prediction
Low Medium High

Individual Business Impact
How Do Companies Gear Up?
Get funding Project work

Peers
Screen
Own projects Get more funding
Ideation
Feasibility/ROI estimation
Get more resources
Involve Different stakeholders
Grass roots Collect Data —
Pilots internal,
Limited budget external ...
Catalog
Data Agile and efficient
Get resources Scientist Secure processing pipelines
ML
Hire Engineer Clarify Deployment
With funding Data  Advisors
Engineers
 Student interns
Upskill  Hackathons
(Ongoing) Scoping of AI/Machine-Learning Projects
Events and Decisions Data Assessment Business Exploration Deployment Monitoring Text
1 3 5 7 8 Images
Audio
Videos
6 Testing and
Measurements (Geo., ...)
2 Storytelling
Ideation Transactions
9
4 Machine-Learning Potential Transactions
Prototyping Recalibration
Events
What is the "old" solution? What is the desired latency during "inference"?
What can we do better? Can we buy, outsource or make?
How can we define "better"?
How often do we have to recalibrate?
Do we collect the best data?
Can we buy/acquire more? What is the desired transparency?

Recommendations and Next Steps
 Make machine learning part of your digital strategy.

 Build a data science initiative (Data Lab, COE, CC, team).
 Revisit old problems that resisted a good solution so far.
 Shallow machine learning will remain superior for many problems!
 Adventure into deep learning only if your organization has the
necessary skills:
– Others will adopt deep-learning-based solutions via APIs, applications or
external service providers.
– But they must still be able to evaluate.

Recommended Gartner Research
 Innovation Insight for Deep Learning

Alexander Linden, Tom Austin and Svetlana Sicular (G00319191)
 Magic Quadrant for Data Science Platforms
Alexander Linden, Peter Krensky, Jim Hare and Others (G00301536)
 Machine-Learning and Data Science Solutions: Build, Buy
or Outsource?
Peter Krensky and Alexander Linden (G00315415)
For more information, stop by Gartner Research Zone.


Tendencias en Inteligencia Artifiial y Maquinas

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Tendencias en Inteligencia Artifiial y Maquinas

Hochgeladen von

Copyright:

Verfügbare Formate

Gartner Data & Analytics Summit

Key Trends in Artificial Intelligence and

CONFIDENTIAL AND PROPRIETARY

Machine capabilities which Deep Learning

Big Data Data

1 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

2 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

Soil Farmer Food

3 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

MSFT's ImageNet Solution

IBM Watson beats Tesla has over a

4 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

5 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

"Watson is a cognitive technology

"AI will take

Fearing a rise of killer robots is like

6 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

By BruceBlaus — Own work, CC-BY-3.0,

1. What is machine learning and AI?

10 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

1. What is machine learning and AI?

11 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

12 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

13 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

14 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

k-NN 20 years Less popular

16 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

TASK: Does this

Input: Raw pixel data

17 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

10-Node Small GPU Large GPU Largest GPU

2. Demand Prediction 3 weeks 4-5 days 12 hours 3 hours 30 minutes

3. Machine Translation 20 years 2-4 weeks

18 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

Systems Virtuous Cycle

19 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

1. What is machine learning and AI?

20 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

BUT: Don't write off shallow machine learning yet!

21 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

22 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

23 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

24 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

25 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

Now and soon

27 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

Structured Business Analytics

Tabular Text Audio Image Video Diverse

28 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

Structured Business Analytics

Tabular Text Audio Image Video Diverse

29 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

DCGAN (Facebook, 2015)

Structured Business Analytics

Tabular Text Audio Image Video Diverse

31 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

and also Manuel Ruder and Thomas Brox (Uni Freiburg)

32 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

Credit to https://deepart.io (Germany)

34 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.

Image Image Filtering

Structured Business Analytics

Tabular Text Audio Image Video Diverse