Beruflich Dokumente
Kultur Dokumente
Summit 2017
20 – 22 March 2017 / London, UK
Classical Engineering
Seed
Weather and time
ML algorithms
Farming HPC ML solutions
Data
Data scientists
1997 ... 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016
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)
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)
1 b
y = a * x + b + error ∑
a
x
1 b
f(∑)
a
x
Sigmoid function
Intermediate presentation
Compute
Requirements
Extremely popular
Log Chart
Very popular
Occasional
Random Forest
SVMs Two-Layer Neural Nets
Decision Trees
Scorecards
Simple Regression Depth
15 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.
Deep learning attempts to improve machine learning, by
the creation of intermediate representations of the data.
Basic Complex
Edges shapes shapes
Output: 0 or 1
High-Performance
Deep
Compute Learning
Complexity
What Next? Algorithms
Data
Most data
The more
elements
data sources
have almost
the better
no meaning
Little or no
domain
knowledge
available
Microeconomics
Increasing abstraction via intermediate computational layers
Macroeconomics
Demand
Social
Transactions
Interactions
Traffic
... also at
the output layer
O
Audio
U
T
P
Text
U
T
S Tabular/ Classical Data Science
INPUTS
O
Audio
U
T
P
Text
U
T
S Tabular/ Classical Data Science
INPUTS
O
Audio
U
T
P
Text
U
T
S Tabular/ Classical Data Science
INPUTS
Input A Input B
INPUTS OUTPUT
Super Resolution
O
Audio
U
T
P
Text
U
T
S Tabular/ Classical Data Science
INPUTS
Super Resolution
O
Audio
U
T
P
Text Machine Translation
U
T
S Tabular/ Classical Data Science
INPUTS
Source: http://nlp.stanford.edu/projects/nmt/Luong-Cho-Manning-NMT-ACL2016-v4.pdf
38 © 2017 Gartner, Inc. and/or its affiliates. All rights reserved.
Wake-up call!
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
Structured Business Analytics
Medical Diagnosis Scene Classification
POS Tagging Video Analytics Compression
INPUTS
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
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
Insufficient:
Data
Domain insights
Compute infrastructure
ML task can be very complex
Skills
Luck Simply unknown if "good enough" solutions exist
Debugging is different to
classical software engineering
250
200
150
100
50
0
?
5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85
Out-of-Sample
In-Sample
Failure
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
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?