This document discusses AI in the enterprise from past, present, and future perspectives. It provides an overview of the history and recent developments in AI and deep learning, including improved performance on tasks like image recognition. Case studies are presented showing how various large companies have successfully applied deep learning techniques like convolutional neural networks to problems in different industries involving computer vision, predictive maintenance, fraud detection, and more. The importance of data quantity for deep learning performance is highlighted. The final sections discuss challenges in AI adoption and the importance of piloting models before full production deployment.

1. AI in the Enterprise: Past, Present & Future
Paul Huibers, Think Big Analytics

2. 2
“
By 2020 AI will be a top five
investment priority for more
than 30% of CIOs.
—Gartner BI Summit,
February, 2017
“The Resurgence of AI
By 2019, deep learning will provide best-
in-class performance for demand, fraud,
and failure prediction. - Gartner

3. 3
AI First

4. 4
• Introduction to AI/DL
• AI in Industry
• Case Study: Financial Fraud
• Pilot to Production
Agenda

5. 5
What is AI?
Artificial Intelligence is usually defined as the
science of making computers do things that
require intelligence when done by humans.

6. 6
AI: A Brief History… …and now, Deep Learning!
• 1940’s – early concepts developed
• 1980’s – more concepts
– copy the brain, neurons, perceptron
– backpropagation for training
• 1990’s – LeCun handwriting reader
• AI winter
• 2009 – Netflix Prize $1M
• 2010 – first ImageNet competition
• 2012 - AI/deep learning comes of age
• ImageNet classification error:
– 2011: 25% using traditional methods
– 2012: 16% achieved by a ConvNet
– 2013: 11%
– 2014: 6.7%
– 2015: 3.6%
– 2016: < 3%
• 3% ~ human error rate (expert group)
• 0.3% mislabeling
• (1000 categories of images)
• What changed since the 1990s?
– 10,000X computing power, GPUs
– massive labeled datasets

7. 7
Deep Learning Innovation in Computer Vision
Recent ImageNet Results

8. 8
ImageNet
1.2 million images 1000 categories lots of animals…
a jungle of viewpoints,
lighting conditions, and variations of all
imaginable types.
…a jungle of viewpoints, lighting conditions, and variations of all imaginable types. – Karpathy

9. 9
What is Deep Learning?
• A machine learning method that involves learning data representations
rather than task-specific algorithms
• Deep Neural Networks – an artificial neural network with multiple hidden
layers of “neurons” between the input and the output
• Artificial Neural Networks – computing systems inspired by biological
neural networks, involving a collection of connected units, with learned
weights and activation functions between the units
How is it achieved?

10. 10
Deep Neural Networks
How are they different?
• Multiple hidden layers in neural network with intermediate data representations
to facilitate dimensional reduction
• Interpret non-linear relationships in the data through activation functions
• Derive patterns from data with very high dimensionality
Why do we care?
• Ability to create value with little
or no domain knowledge
required
• Ability to incorporate data from
across multiple, seemingly
unrelated sources
• Ability to tolerate very noisy data

11. 11
Data Quantity Drives Deep Learning Performance
Andrew Ng
Amount of Labeled Data
ModelPerformance
1990’s
Small Training Sets
Traditional ML
Small NN
Medium NN
Large NN

12. 12
Deep Learning Architectures
Convolutional Neural Network (ConvNet or CNN)
• CNN = Convolution + Pooling + ReLu + Fully Connected
• Convolution Layers are composable so can be chained
• Primary use: any problem that has a high
dimensional input (ex.: Image Labeling)

13. 13
Specialized APIs General Purpose Frameworks
AI Framework Landscape
Vision
Language
Speech
Keras
• Pretrained (fast)
• Public
• Google/Microsoft/Amazon
• Need to be trained (expensive)
• Private
• Fully customizable

14. 14
Touched by AI…
• Cognitive successes
• Siri, Alexa, OK Google!
– Understanding words
– Understanding context
– Language translation
• Face detection in images
• Recommender systems
• How about some practical
examples from industry?

15. 15
• Introduction to AI/DL
• AI in Industry
• Case Study: Financial Fraud
• Pilot to Production
Agenda

16. 16
Proven applications of Deep Learning
ANOMALY
DETECTION
Enables real-time
detection of
abnormal patterns of
data, usually time-
series events.
PREDICTIVE
MAINTENANCE
Improves preventative
measures &
performance with
greater accuracy at
the asset &
component level
RECOMMENDER
SYSTEMS
Enable more effective
search rankings based
on context, in
accordance with a
particular objective
such as purchase or
click-through
SPEECH
RECOGNITION
Enable capture of
voice to text with
higher fidelity of
speech transcription
and improved
precision of speaker
identification
COMPUTER
VISION
Enables dramatically
more accurate visual
recognition tasks
that include image
classification,
detection and
localization
DOCUMENT
AUTOMATION
Enables automation
of manual, paper-
based processes
that are human-
intensive with higher
speed, accuracy
and fidelity

17. 17
Industry Specific Use Cases
High-Dimensional Data
Image
Video
Audio
Time Series
Text
• Many already have working solutions using non-DL Machine Learning Techniques
• Deep Learning is delivering improvement in performance on complex problems
Automotive Retail
• Navigation, Guidance, Assistance
• Predictive Maintenance
• Visual Search
• Recommendation
• Text Analytics
• Assistants
• Brand Analytics
Manufacturing & High-Tech Health Care
• Image/Audio/Video
• Reinforcement Learning – Systems
Optimization
• Plant Operations Optimization
• Image-based Analysis
• Drug Discovery
Financial Services & Insurance Cross-Industry
• Anti-Fraud
• Portfolio Optimization
• Damage Assessment
• Cyber Security
• Call Center Audio

18. 18
Large European Logistics Provider
• Increasing use of plastic bags in
shipping
• Challenges with existing package
sorting and identification system
• Use Deep Learning Image
Analytics to improve identification
and sorting
• Tools: TensorFlow, Hadoop
• Techniques:
– Deep Learning: Convolutional
Neural Network

19. 19
• Road objects, traffic and accident
events are manually reported or not
at all
• Automated object detection and
scene labeling system from car
camera feed to improve navigation
and traffic
• Tools: Darknet, Caffe, TensorFlow
• Techniques:
– Object Detection: Single Shot MultiBox
Detector (SSD), You Only Look Once
(YOLO)
– Scene Labeling: Convolutional Neural
Network
Large Auto Parts Manufacturer Use Case
Real-Time
Streaming
Streaming
Results
Traffic Data Service
Navigation Update
Darknet/Darkflow –
Object Detection
TensorFlow – Scene
Labeling
Cloud GPU
Based Training
TF Serving
Cloud GPU
Based Inference
Model
Updates

20. 20
• Handwritten check volume is
decreasing however processing
checks has many fixed costs
• Handwriting recognition to reduce
manual processing and fraud
examination resulting in cost savings
• Tools: Spark, Hadoop, TensorFlow
• Techniques:
– Convolutional Neural Network
– Image Processing
Large US Multinational Bank
Check Images
To Hadoop
ImageMagick
Processing
Handwriting
Recognition
Fraud
Detection

21. 21
• Predict failure of pistons on large
container ships to reduce unplanned
and costly maintenance
• Utilized sensor data to predict piston
wear between 70-80%
• Failures are extremely infrequent so there
is a risk of overfitting
• Tools: R, Hadoop, Spark, AWS
• Techniques:
– ROC curve
– Internet of Things data
– Methods to prevent overfitting
Large Container Shipping Company Container Ship Sensors
Predict
Failures
1 month (December)
High
Low
Abnormalcylinderbehaviour*
Lead time
Port stays
PROB1
0.0
0.2
0.4
0.6
0.8
2015-12-06 2015-12-13 2015-12-20 2015-12-27
Piston ring change
Cylinders
Piston ring(s) changed Threshold
Abnormal behavior:
Everything above
the threshold
triggers an alarm
Other cylinders
below threshold
Worn piston ring was
changed
Each point is a
combination of
selected sensor data
for a specific cylinder

22. 22
Large European Railway
• Detecting rail switch failures
• Allows for switches to be fixed
ahead of time thus not delaying
trains
• Tools: R (Shiny and Studio),
Hadoop, Spark
• Techniques:
– Survival Analysis
– Machine Learning
– Internet of Things data
Railway Switch Sensor Data
Visualize
Failures
and Act

23. 23
• Fraud detection across products
• Trends
– Mobile payments exploding
– Fraud evolving rapidly, increased
sophistication
• Significant improvements over
traditional rules-based techniques
• Tools: Spark, Hadoop, TensorFlow
• Techniques:
– Boosted Decision Trees
– Convolutional Neural Network
Large European Bank

24. 24
State of AI in Industry
Successes
• Computer vision (e.g., ImageNet)
• Speech & NLP
• Simplification of general-purpose
ML (e.g., recommendation)
• Rapid advance of state of art,
growth of expertise & applications
• Major investment programs in
industry
Challenges
• Research-driven, fundamentals
change
• Mostly empirical, little theory
• Complexity in solution design
• Limited access to talent
• AI/DL still requires governed data,
and Analytics Ops integration
• Gaps in enterprise deployment
beyond lock-in clouds

25. 25
• Introduction to AI/DL
• AI in Industry
• Case Study: Financial Fraud
• Pilot to Production
Agenda

26. 26
Fighting Financial Fraud
with Artificial Intelligence
at Danske Bank

27. 27
Data Driven Approach to Fight Fraud
Fast evolving fraud sophistication
Ambitions for Fraud Project
Danske Bank advanced
analytics blueprint
Data driven approach to
real time scoring of
transactions
Reduce false-positives &
Enhance fraud detection rate
ONLY ~40%
of fraud cases are detected using rules
Low Detection Rate
99.5%
of cases are not
fraud related
Many false positives
Challenges for Fraud Detection
Tens of Millions
€ lost each month
High Fraud Loss
© 2017 Teradata

28. 28
Modeling
Challenges
© 2017 Teradata
• Class imbalance
(100,000:1 non-fraud vs. fraud)
• Assigning fraud labels from
historic data
• Fraud is ambiguous
• Not all features available in
real-time (balance, etc.)
• Most machine learning sees
transactions atomically

29. 29
Current models can
only catch ~70% of
all fraud cases
Deep Learning Opportunity
Traditional ML models
view transactions
atomically
Often missed fraud
transactions are part
of a series
Capturing
correlation across
many features
© 2017 Teradata

30. 30
Machine Learning Results
(Live System: 60 transactions/sec.)
Ensemble of boosted decision trees
and logistic regression.
From online validation of the model:
● 25-30% false positive reduction, with
over 35% increase in detection rate
● Opportunity to expand model with
additional features, retrain on recent
data and add additional models to
the ensemble.
● Models can be expanded to
additional channels
Rule Engine on
validation set
© 2017 Teradata

31. 31
Three Deep Learning Architectures to Deliver Value
• Designed for spatial
correlated features,
but by transforming
transactions into a
2D image, we can
learn temporal
correlated features.
• Deeper ConvNet
allows learning more
complex & general
features.
Goal: Learn kernels from
temporal & static
features to gain insight
into the characteristics of
fraud.
• Learn temporal
information and
classify if the
sequence of
transactions
contains fraud.
• Shares knowledge
across learning time.
Goal: Learn transaction
patterns within a window.
Two solutions can be
tested: flag fraud or
predict next transaction
and define an error.
• Learn how to
generate normal
transactions,
potentially large
volumes of non-fraud
data.
• AE provide a low level
representation of the
data.
Goal: Build a model that
learns how to generate
non-fraud data. To
detect fraud, define a
reconstruction error rate
for the fraud cases
Auto-Encoders
LSTM
ConvNet
© 2017 Teradata

32. 32
How Can We Create an Image From Bank Transactions?
t0 X_0, X_1, ... X_n
dt
t1 X_0, X_1, ... X_n
t2 X_0, X_1, ... X_n
ts X_0, X_1, ... X_n
...
Top k Features Correlation
...
X_0
X_41 X_5
X_30
X_29X_31X_10
X_37
X_3
X_1
X_42 X_40
X_32
X_15X_35X_2
X_16
X_31
X_2
X_3 X_15
X_4
X_1X_28X_40
X_31
X_49
X_n
X_26 X_9
X_40
X_35X_28X_2
X_17
X_1
...
X_0
X_41 X_5
X_30
X_29X_31X_10
X_37
X_3
X_1
X_42 X_40
X_32
X_15X_35X_2
X_16
X_31
X_2
X_3 X_15
X_4
X_1X_28X_40
X_31
X_49
X_n
X_26 X_9
X_40
X_35X_28X_2
X_17
X_1
...
...
...
...
...
X_0
X_41 X_5
X_30
X_29X_31X_10
X_37
X_3
X_1
X_42 X_40
X_32
X_15X_35X_2
X_16
X_31
X_2
X_3 X_15
X_4
X_1X_28X_40
X_31
X_49
X_n
X_26 X_9
X_40
X_35X_28X_2
X_17
X_1
N
dt
t0
t1
ts
Input Output
Raw Features
Add correlated features
in a clock-wise manner
© 2017 Teradata
Image size is:
[10 x 3, 50 x 3, 1]
Time
Strides of 3

33. 33
2D Transaction Image Example
© 2017 Teradata
Non-fraud Transaction Image
Non-fraud
Fraud Transaction Image
X-axis: features, Y-axis: time
Fraud
Non-fraud

34. 34
Network Architecture for CNNs
Fraud
Normal
50
30
25
15
13
8CNN
Fraud
Normal50
30
25
15
25
15
25
15
25
15
© 2017 Teradata

35. 35
Inside the ResNet model
64 Filters
Activations
After the CNN
Residual Blocks
FraudNon-fraud
© 2017 Teradata

36. 36
Dramatically Improve Fraud Detection over Traditional Rule Engine
© 2017 Teradata
Our Deep
Learning
Rules
Engine
False Positive Rates
Our Deep
Learning
vs
Rule
Engine at
9% FP
40% Gain
TruePositiveRates
Classic
Machine
Learning
Our Deep
Learning vs
Machine
Learning at
1% FP
44% Gain

37. 37
Lessons Learned: Take-Aways From Danske Bank
Deep
learning
adoption
from
pictures to
financial
transactions
Enhancement
of data quality
& cluster
capabilities
with data
ingestion
Building
Analytics
Ops
capabilities
to support
business
units
Leveraging
experience
from Fraud
advanced
analytics to
deliver extra
use cases
© 2017 Teradata

38. 38
• Introduction to AI/DL
• AI in Industry
• Case Study: Financial Fraud
• Pilot to Production
Agenda

39. 39
Operationalization is Hard
“We evaluated some of the new methods offline
but the additional accuracy gains that we
measured did not seem to justify the engineering
effort needed to bring them into a production
environment.”
- Netflix, 2012

40. 40
Focus First on Pilot into Production
Sets up Phase Two: Scale COE, Standardize Capabilities
Investigate
Test
Engineer
SimulateIntegration
Analyze
Data
Go Live
Handover
Validate
Activities: Define business
opportunity, understand data
available, test model
approaches, potentially
generate data
Outcome: Proposed solution
approach
Discovery/Insights
Activities: Architecture
selection, software engineering
of model and simulation
Outcome: Predicted impact of
model
Live Test
Activities: Integration into
live business process
(Champion/Challenger),
analysis, iteration
Outcome: Benefit
measurement, live learnings,
improvement
Production
Activities: Go Live, Analytics
Ops integration, Hand Over
Outcome: System scaled,
application teams and ops
trained and operating
Assessment
Insights
Production
Live Test
Cross-Functional
Teams
Cross-Functional Teams

41. 4141
For more information, please contact:
Paul.Huibers@ThinkBigAnalytics.com
603-395-6567
Thank You StampedeCon!
stampedecon.com/ai-summit-2017-st-louis/

42. 42
The Future… and more…
Architectural innovations: RNN, LSTM, GAN and more
Better training through new optimization, new activation functions and more
Transfer learning, pre-training and more
Theory catching up with practice (Tishby) – relevant information, bottlenecks
10,000X speed improvement would make many things possible – Moore’s Law
Unsupervised learning
Learning with few samples
AGI – artificial general intelligence
Singularity