You are on page 1of 24

​ Bringing Artificial Intelligence to the Enterprise

​ Chad Meley, VP of Artificial Intelligence Product Marketing at Teradata


AI Is Happening in the Enterprise

PayPal John Deere Lowe’s JP Morgan Wells Fargo Walmart


2
Waves of Industry Disruption Powered by AI
Financial Services

Incumbents

Challengers

3
Waves of Industry Disruption Powered by AI
Auto Manufacturing

Incumbents

Challengers

4
AI makes complex AI improves
decisions productivity

Incorporates
traditional data sets

Improved financial fraud Chatbots to improve call


detection and reduced center experience
costly false positives

Leverages previously
intractable data sets Predict costly unplanned Computer vision
downtimes by analyzing improves operational
maintenance records, efficiencies in logistics
sensors AND asset imagery

5 © 2017 Teradata
Business Challenges Faced

Only 40% of fraud caught

99.9% of cases investigated


Were not fraud

Getting worse because of new payment


methods and fraudster sophistication

6 Copyright 2017 Teradata


Dramatic Improvement in Fraud Detection
Deep •  50% increase in
Learning fraud detection
rate

•  60% decrease in
True Positive Rates

false positive rate

Classic Rules •  From design to


Machine Engine models in
Learning production in 8
sprints

7 © 2017 Teradata
False Positive Rates
Convolutional Neural Networks

In machine learning, a convolutional neural network (CNN) is a class of deep, feed-


forward artificial neural networks that has successfully been applied to analyzing visual
imagery.

CNNs use relatively little pre-processing compared to other image classification


algorithms. This means that the network learns the filters that in traditional algorithms were
hand-engineered. This independence from prior knowledge and human effort in feature
design is a major advantage.

They have applications in image and video recognition, recommender


systems and natural language processing.

8 © 2017 Teradata
How Can We Create an Image From Bank Transactions?
Add correlated features Image size is:
Input Output in a clock-wise manner [10 x 3, 50 x 3, 1]

Raw Features N

t0 X_0, X_1, ... X_n X_3 X_41 X_5 X_31 X_42 X_40 X_49 X_3 X_15 X_1 X_26 X_9
...
t1 X_0, X_1, ... X_n t0 X_37 X_0 X_30 X_16 X_1 X_32 X_31 X_2 X_4 X_17 X_n X_40

X_10 X_31 X_29 X_2 X_35 X_15 X_40 X_28 X_1 X_2 X_28 X_35
t2 X_0, X_1, ... X_n dt
X_3 X_41 X_5 X_31 X_42 X_40 X_49 X_3 X_15 X_1 X_26 X_9
...

...
t1 X_37 X_0 X_30 X_16 X_1 X_32 X_31 X_2 X_4 X_17 X_n X_40

ts X_0, X_1, ... X_n X_10 X_31 X_29 X_2 X_35 X_15 X_40 X_28 X_1 X_2 X_28 X_35
dt

...

...
...

...
Top k Features Correlation X_3 X_41 X_5 X_31 X_42 X_40 X_49 X_3 X_15 X_1 X_26 X_9
...
ts X_37 X_0 X_30 X_16 X_1 X_32 X_31 X_2 X_4 X_17 X_n X_40

X_10 X_31 X_29 X_2 X_35 X_15 X_40 X_28 X_1 X_2 X_28 X_35

(examples of k: client location, merchant location, size of transaction,


9 © 2017 Teradata relative size of transaction, frequency of transactions, others….)
Inside the model

Non-fraud Fraud

64 Filters
Activations
After the CNN
Residual Blocks

10 © 2017 Teradata
Improving Operational Excellence with Artificial Intelligence (AI)
Major Logistics Company

Business challenge Solution Benefits


•  Small percentage of packages wind •  Designing and deploying AI assisted •  Reduction of warehouse search window
up as “orphaned” solution, image to image approach. from weeks to minutes.

•  Reducing resolution time on •  Improving business and technical •  Orphaned matching on target to
orphaned packages currently processes through research and achieve 90% successful match rate by
costing the company tens of millions assessment to accelerate business results. year two.
per year
•  Increasing progress with active learning •  Improvement in customer satisfaction due
on AI. to a quick resolution of lost and found
cases.

Tools &
Technologies

VELOCITY ™ – Artificial Intelligence


Strategy
4
11 © 2017 Teradata
Greater Operational Efficiencies with Deep Learning
State run postal service

Business challenge Solution Benefits


•  115M parcels shipped per year; 7.5m •  Several techniques used to process and •  Processed around 10K images in an
parcels are plastic bags but only a small % transform 10K images automated pipeline
are sorted correctly into an automatic roll
•  Pre-trained models vs. new Models were •  Achieved 80% accuracy on testing dataset
cage container
compared and results benchmarked; for identification of plastic bags
•  Key challenge: •  Keras with Theano backend was used for
the final results due to more user-friendly
•  Identify plastic sacks from images & interface and limited time for the project
measure size

•  Automate sorting to automatic roll cage


container

•  Optimize load/unload of parcels & number


of containers

Tools and Technologies

VELOCITY ™ – Artificial Intelligence Strategy

12
Customer Intimacy Enabled by AI
Global Mobile Communication Provider

Business challenge Solution Benefits


▪  Reducing cost of customer-driven ▪  Building an AI solution to replicate 332 ▪  Reducing the customer interaction costs
communication types of queries in 2 languages (English & with agents.
Roman Urdu and their mixture).
▪  Developing solutions with AI – machine ▪  Decreasing resource driven challenges as
intelligence and natural language ▪  Delivering user interface for interacting virtual agent able to work 24/7.
processing. with a virtual agent and host the solution
▪  Scalable system enabling future use case
on Teradata Azure server for testing.
projects.
▪  Developing a robust solution easy to
extend to other response types.

Tools & Technologies

VELOCITY ™ – Artificial Intelligence Strategy

13
Data For Predictive Maintenance
Structured Data Unstructured Data
•  Time Series •  Text
•  Events •  Image
•  Graph •  Sound

14 © 2017 Teradata
What is AI?
•  ”Artificial intelligence" is •  Artificial intelligence is technology •  AI generally refers to the
applied when a machine that appears to emulate human ability of machines to
mimics "cognitive" performance typically by exhibit human-like
functions such as learning, coming to its own intelligence—for example,
"learning" and "problem conclusions, appearing to solving a problem without
solving.” understand complex content, the use of hand-coded
engaging in natural dialogs with software containing
people, enhancing human detailed instructions.
cognitive performance or
replacing people on execution of
non routine tasks.

15 © 2015 Teradata
AI Evolution

ARTIFICIAL
INTELLIGENCE MACHINE
LEARNING DEEP
Early theory and
accomplishments LEARNING
Predictive
stirs interest and
accuracy
excitement Allows more-complex problems to
improves with
be tackled, and others to be solved
more data
with higher accuracy, with less
K-means clustering, cumbersome manual fine-tuning
Bayesian Networks,
Recurrent Neural Networks,
Support Vector Convolutional Deep Neural
Machines Networks, Generative Adversarial
Network, Deep Feed Forward

1950s mid 2000s mid 2010s


16 © 2017 Teradata
What makes Deep Learning so intriguing?

•  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

17 Copyright 2017 Teradata


AI Technology Landscape
SaaS Analytics AI APIs Frameworks for AI Model
Development and Execution
Visual Vision

Assistive Speech

Keras

Operations Language

Proprietary cloud APIs. Software frameworks.


Pre-packaged Applications
Leaders are open source.
Vendor-owned Models Models built on Enterprise Data

18 Copyright 2017 Teradata


AI Project Execution Options
SaaS Analytics AI APIs Custom AI
Development
Models Commodity, limited One size fits all Tailored & Unique
configuration
Data Least common Data...not informed Leverage all data,
denominator by enterprise data including enterprise
Business Process Disconnected Limited Fully Integrated
Integration
Deployment Isolated SaaS Dependent PaaS Any (Hybrid)

AI Expertise None Minimal Significant

Implementation Low cost & low risk Moderate cost & High cost & high risk
moderate risk
Competitive Parity Tactical Strategic & Game
Advantage “me too” Changing

19 Copyright 2017 Teradata


Challenges Delivering Enterprise AI
AI and Deep Learning initiatives introduce Unique and Challenging Requirements

Strategic Technology Operational Data Talent

•  Significant business •  Academic code •  Transparency of •  Governance and •  Shortage of AI talent


processes change •  Specialized models quality •  Immature best
•  Broad hype and hardware •  Integration of AI with •  Requires curated, practices in all
interest •  Support gaps for broader ecosystem high value data at required capabilities
•  Limited range of open source •  Deployment of scale •  AI change
applications to date •  Poor interoperability large-scale GPU •  Need to blend with management
clusters high velocity data emerging
•  Performance and •  Training data
scale curation

20
21 © 2017 Teradata
​ Tips for a Successful Enterprise AI Initiative

•  Get aligned on AI use cases and roadmap


•  Seek early wins to get momentum and solidify stakeholder interest
•  Acknowledge and address skill and capability gaps
•  Rethink processes that were originally designed pre-AI

© 2015 Teradata
22 ​ © 2017 Teradata

Over the next decade, AI won’t replace
managers, but managers who use AI
will replace those who don’t.

’’
HBR July 2017

23
www.teradata.com/Insights/Artificial-Intelligence

24
24 Copyright 2017 Teradata

You might also like