CompetitiveAdvantageThroughDeepLearning (white paper)

W H I T E P A P E R
Competitive Advantage
through Deep Learning
Math, Technology, Applications, and Data:
Which Value Levers Are Right for Your Business?

TABLEOF
CONTENTS
01
05
07
16
Obstacles to Adoption
Introduction
Four Value Levers for Businesses
Conclusion
C O P Y R I G H T © 2 0 1 6
02History of Deep Learning
18References
08
12
13
Applications
Math
Data
10Technology

Contributions
Tarun Mehra, Principal
Author
tmehra@fuld.com | 857.202.5263 | Boston, MA
Robert Flynn, Principal
Principal Editor
rflynn@fuld.com | 857.202.5308 | Boston, MA
Ola Jachtorowicz, Content Director
Design
Allison Hackel, Marketing Coordinator
Production and Editorial Assistance

C O P Y R I G H T © 2 0 1 6
INTRODUCTION
Whether you realize it or not, chances are that you
have interacted with one or more deep learning
algorithms within the last few hours.
Your interaction may have happened while tagging
friends in a photograph suggested by Facebook,
reviewing a movie list suggested by Netflix, or
giving voice commands to your iPhone. You may
have even benefited from deep learning when you
engaged the autopilot mode in your last test-drive
with Tesla.
All of these seemingly intuitive and simple
experiences are possible because, behind the
scenes, complex deep learning algorithms are
running on huge amounts of trained data sets,
using high computational power across hundreds
of distributed computers. Deep learning -
computers’ ability to learn patterns, formulate
predictions, or make decisions based on data
sets - is now so commonplace that it often occurs
without end user recognition.
01
We are all increasingly
using deep learning.

C O P Y R I G H T © 2 0 1 6
02
HISTORY OF
DEEP LEARNING
It took the efforts of thousands of accomplished
academics and technologists globally to create
the “right conditions” for deep learning to take
shape and show its value in daily life. Its benefits
are beginning to impact the world’s business,
academic, and government circles. Given the rate
at which deep learning is becoming mainstream, it
will benefit almost all industries in some way in the
next five years.
Such a change will cause significant disruption.
Those companies that properly adopt deep learning
in their core businesses will not only survive the
disruption but will realize significant competitive
advantage.
Academic interest in deep learning preceded the
explosion of its real life applications. Deep learning,
a highly specialized field of machine learning and
artificial intelligence, is not a new discipline. The
field of artificial intelligence research was founded
at a conference at Dartmouth College in 1956.
However, deep learning only gained real
global traction in 2012 [see page 4], driven by
significant improvements in the mathematical and
computational efficiencies of artificial neural network
algorithms
On right: Researchers at Argonne National Laboratory
inspect the lab's first digital computer, AVIDAC, which began
operations in 1953. Source: Wikimedia Commons
After a long
period of
development, the
right conditions
now exist.

C O P Y R I G H T © 2 0 1 6
03
HISTORY OF
DEEP LEARNING
Artiﬁcial
Intelligence
Machine
Learning
Deep
Learning
A discipline to study and
develop machines exhibiting
human intelligence
One of the approaches to
achieve artiﬁcial intelligence
One of the techniques
for implementing
machine learning
1980+
1950+
2010+
Consider the following statistics to get a sense of
the enormous growth in academic interest in deep
learning: about 19,000 research papers related to
deep learning have been published in 2016 so far,
compared to 14,200 in 2015, and 1,740 in 2005.
Only five hundred research papers were published
on the topic in 2000 (source: Google Scholar).
Improvements in computational efficiencies
of deep learning algorithms have fed the
development of more complex technologies.
Open source communities started freely sharing
deep learning “libraries” for other developers
who could use them to write new applications.
A variety of deep learning platforms with simple
user interfaces have come to the market, and can
be used by non-coders to develop applications.
With these advanced tools and high level of
interest, a race began to develop or acquire
deep learning capabilities. Technology innovation
companies with deep pockets such as Google,
Apple, Intel, and Twitter got into a rush to grab
startups that were focusing on different parts of
the deep learning value map. These companies
(and their investors) have already made huge
investments in the deep learning space. More
investment is expected.
Thus, after decades of computer science efforts,
the "right conditions" now exist: conceptual and
technological advances, initial successes in a
variety of real life applications, visions for future
applications, and investors lining up to put their
money toward innovation.

C O P Y R I G H T © 2 0 1 6
HISTORY OF
DEEP LEARNING
04
IBM’s Deep Blue supercomputer defeated world chess
champion Garry Kasparov in a six-game match in 1997. IBM
Watson won Jeopardy! in 2011.
While these earlier examples used artificial intelligence and
machine learning, 2012 proved to be a turning point for
deep learning, driven by substantial improvement in math,
technology, and demonstrable applications.
Using 16,000 Computers to Spot Cats in 10 Million YouTube Videos
JUNE 2012. Google scientists built a neural network across 16,000 computer processors with over one
billion connections. It randomly accessed 10 million YouTube videos with one simple objective: to identify
cats in those videos. The simulation ran for three days, and started recognizing pictures of cats after only 20,000 videos.
The team, led by Stanford computer science professor Andrew Ng (now with Baidu) and Google fellow Jeff Dean, fixated one
simulated neuron in the model on images of cats, while others focused on human faces, yellow flowers, and other objects. Deep
learning algorithms helped the system identify these objects even though no definitions, labels, or distinguishing features of
these objects were provided to the system (a process called ‘unsupervised learning').
When the system was asked to categorize objects into 22,000 granular categories, the accuracy was about 16%. While it may not
sound great, this outcome was 70% better than previous methods. When the system was asked to categorize the objects into 1,000
more generic categories, the accuracy rate drastically improved to 50%.
AUGUST 2012. Merck launched Merck Molecular Activity Challenge with a goal to help develop safe and
effective medicines by predicting molecular activity. The objective of the competition was to “identify
the best statistical techniques for predicting biological activities of different molecules, both on- and off-
target, given numerical descriptors generated from their chemical structures”. The challenge was based
on 15 molecular activity datasets, each for a biologically relevant target.
The competition ended two months later on October 16th 2012, with the leaderboard topped by a team called “gggg” from the
startup Kaggle. This team consisted of two professors at the University of Toronto (Geoff Hinton and Ruslan Salakhutdinov) and
three PhD students (George Dahl, Navdeep Jaitly, Christopher “Gomez” Jordan-Squire), all highly active researchers in the field
of deep learning.
OCTOBER 2012. Richard F. Rashid, one of Microsoft’s top scientists, delivered a lecture at a conference in
Tianjin, China. Deep learning was on a bold display at this event since while Dr. Rashid was delivering his
lecture in English, some complex algorithms were recognizing his words, and simultaneously displaying
them on a large screen on the stage.
While that was a significant achievement, what really got Dr. Rashid a huge applause at that event was what came after that. In his
next demonstration, he paused after each sentence, and his words were not only translated into Mandarin characters, but were
also accompanied by a simulation of his own voice in Mandarin. Dr. Rashid himself had never spoken those words in Mandarin.
2012 proved to be a
turning point for deep learning
Deep Learning Experts Win 'Merck Molecular Activity Challenge'
Dr. Rashid Uses Deep Learning to Speak Mandarin Without Learning It

05
C O P Y R I G H T © 2 0 1 6
OBSTACLES TO
ADOPTION
Despite all the attention in recent years, deep
learning is a topic understood in its entirety by
only a few. These people include deep learning
researchers, highly specialized developers,
technology entrepreneurs and innovators, or
a select breed of corporate executives and
investors. A vast majority of other executives
in corporate and investment firms still have a
relatively high-level, conceptual view of what
deep learning is.
While this high level view is helpful for executives,
it’s not enough to use it to create value for their
businesses. For that, a good understanding of
the value levers of deep learning is required.
Value levers in this context mean those specific
components of deep learning that could be
capitalized on by businesses, organically or
inorganically, irrespective of the industry, to
create significant additional value. Value levers
of deep learning are outlined in the next section.
Executives who do understand the value levers of
deep learning are already prioritizing it on their
strategic agenda and, as a result, they will be much
better placed than others to make better business
and investment decisions to win in the future.
Why aren't a majority of executives
actively and urgently pursuing deep learning
to create value for their businesses?

06
C O P Y R I G H T © 2 0 1 6
OBSTACLES TO
ADOPTIONThere are three reasons why deep learning isn’t being
pursued actively and urgently by many executives
despite the impending nature of its impact on virtually
all industries within the next five years.
For those executives who want to capitalize on the
deep learning opportunity, not understanding its
right value levers can result in wrong strategic
decisions.
A deep learning start-up may focus on the wrong value
levers in an investor-pitch and may get rejected, costing
them their future. A large corporation may make a wrong
acquisition, partnership or investment decision without
realizing that the value levers of involved entities are
not aligned. An investor may misunderstand the value
lever being pitched by a startup or a corporation and
may make wrong investment decisions.
1
2
3
VALUE LEVERS NOT UNDERSTOOD
VALUE LEVERS MISUNDERSTOOD
VALUE LEVERS UNDERSTOOD,
BUT FACE RELUCTANCE TO CHANGE
Some executives might find the initial exposure to the concepts
around deep learning difficult to understand and assume that
none of the value levers are relevant for their businesses.
Some executives may make the mistake of choosing wrong
value levers for their business growth or assume that all value
levers are relevant for them.
Some executives may be too protective of their traditional
models or find it challenging to change the status quo.

07
C O P Y R I G H T © 2 0 1 6
FOUR VALUE LEVERS
FOR BUSINESSES
Value Levers - specific components of deep learning that can be
capitalized on by businesses, organically or inorganically, irrespective
of the industry, to create significant additional value.
Deep learning has four key value levers.
Relevance of these value levers to a particular
business is a function of its industry, competitive
position, and maturity level of its technological
and innovation efforts.
MATH
TECHNOLOGY
APPLICATIONS
DATA
Definitions and algorithms that
constitute artificial neural networks
Platforms and libraries for building deep
learning applications
Use cases for deep learning such as natural
language processing and object recognition
Behavioral, visual, auditory, and textual information
that can be used to train artificial neural networks

When we look at an image of a living room, we
can quickly figure out the particulars - a sofa, a
table, a television, a rug, a window, etc.
When an image is provided as an input to a
computer, it identifies data packets (pixels,
color frequencies, etc.). Unlike a human being,
a computer doesn’t intuitively identify what that
image consists of, because it’s not "trained" to
make deductions about the real world based on
the data from the image.
The neural networks in our brains help us
parse through our memory and make real time
computations to decide what an image is. This
dynamic is made possible through years of
“training” we undertake about the world around
ENVIRONMENTALLY
FRIENDLY FURNITURE
08
C O P Y R I G H T © 2 0 1 6
FOUR VALUE LEVERS
FOR BUSINESSES
us, and through the immense amount of data
(and relationships) our brains hold.
It’s easy to extend this example to other
normal human functions - speech recognition
(translating sound waves from what we
hear to their real world meaning), and facial
recognition (remembering and recalling who
a person is in the real world by looking at the
person’s face), etc.
The Math part of deep learning essentially
encapsulates all the definitions and efficient
execution of such neural network algorithms
in the context of computers instead of
humans. Neural network algorithms and
approaches in context of mathematics
MATH
The neural networks in
our brains help us parse
through our memory
and make real time
computations to decide
what an image is.
The human brain is able to categorize similar objects (such as dinosaurs) thanks to years of data collection.

have been around for a few decades. But
it’s only been in the last few years that these
algorithms have been improved significantly
to lend themselves to real life applications.
Like any algorithms, deep learning algorithms
need to be provided inputs (an image, a voice,
a video, a data-set) with a goal to ascertain
the outcome (what’s in the image, does the
image have a certain person, what is the voice
saying, translate the voice).
Mathematically, this is done through a network
of nodes, virtual neurons analagous to neurons
in the human brain, which are the fundamental
building blocks of deep learning algorithms. The
algorithms may have many connected nodes
spreadacrossmanylayers,includinginput-layers,
output-layers, and intermediate hidden-layers.
These nodes and their relationships typically
have computational properties called
“biases" and “weights”. These weights and
biases determine how each simulated node
responds – with a mathematical output
between 0 and 1 – to a digitized feature such
as the edge of or a shade of red in an image.
The goal of the algorithm typically is to
minimize the error function between the
algorithm output and the actual truth. This
is possible over a period of time if the right
amount and type of training data is fed into
the algorithm, and as the algorithm adjusts
(learns) its nodes and their properties over
many iterations and tests.
Many of today’s artificial neural networks
can train themselves to recognize complex
patterns. Some of these deep learning
algorithms are DBN (Deep Belief Nets),
RBM (Restricted Boltzmann Machine), CNN
(Convolutional Neural Nets), RNN (Recursive
Neural Nets), AutoEncoders, and RNTN
(Recursive Neural Tensor Net).
09
C O P Y R I G H T © 2 0 1 6
FOUR VALUE LEVERS
FOR BUSINESSES
MATH
A variety of efficient artificial neural network algorithms
are available to application developers, who typically
choose the right algorithm depending on the nature
and the goal of their end applications
Restrictive Boltzmann Machine
Deep Belief Network
Edge detection enables object recognition.

10
C O P Y R I G H T © 2 0 1 6
There are various technological options to
access and deploy relevant deep learning
algorithms discussed in the previous
section. Broadly, they fall in two categories
- a) Deep Net Platforms, and b) Deep Net
Libraries.
A deep net platform can be understood
as a set of tools that other people can use
to build their own application. This is not
very different from applications that can
be built off of the tools provided by iOS
and Android, Windows and MacOS, etc.
So, in a nutshell, a deep learning platform
provides a set of tools and interface for
building custom deep nets.
Typically, these platforms provide users
with a selection of deep nets from which to
choose, along with the ability to integrate
data from various sources, adjust data,
and manage those models through an
easy user interface (UI). Some technology
platforms can also help with performance
if a net has to be trained with a large
data set. Some popular deep learning
platforms are Ersatz Labs, H2O.ai, and
Dato GraphLab.
A deep net library, on the other hand, is a
premade set of deep learning functions and
modules that can be called through your
own software programs.
These libraries are not designed to provide
easy user interface (UI) to directly build
models, but rather are typically created
and regularly maintained by highly
qualified developers and technology
teams. Many libraries are open sourced
and are surrounded by big communities
that provide support and contribute to the
codebase. Some of the most popular open
source libraries are Theano (Python based),
TensorFlow (Google), Torch, Caffe and
DeepLearning4J. Each of these libraries
has its own pros and cons, but they are
all fairly popular amongst the developer
communities.
TensorFlow is particularly worth highlighting
because it was originally developed by
FOUR VALUE LEVERS
FOR BUSINESSES
TECHNOLOGY
Deep Net Platform - out-of-box application
thatprovidesthetoolstoapplydeeplearning
algorithms using an intuitive user interface
Deep Net Library - premade set of deep
learning functions (often open-source) that
can be called through custom software

the researchers on the Google Brain team
within Google’s machine intelligence research
organization. However, the library has since
been open sourced and made available to the
general public. A primary benefit of TensorFlow
is said to be distributed computing, particularly
across multiple-GPUs.
A commonly asked question is whether one
should use a platform or a library to build an
application. The answer to that question goes
back to understanding the advantages and
disadvantages of each.
As discussed earlier, a platform is an out-of-
box application that lets you configure a deep
net’s hyper-parameters through an intuitive UI.
So, with a platform, you don’t need to know
anything about coding in order to use the
tools. However, the downside is that you are
ENVIRONMENTALLY
FRIENDLY FURNITURE
11
C O P Y R I G H T © 2 0 1 6
FOUR VALUE LEVERS
FOR BUSINESSES
constrained by the platform’s selection of
the deep nets and configuration options.
For anyone looking to quickly deploy a
deep net, a platform may be the best way
to go.
On the other hand, a deep learning library
is, as discussed above, a set of functions
and modules that you can call through
your own code in order to perform
certain tasks. So, deep learning libraries
provide a lot of extra flexibility with deep
net selections and hyper-parameter
configurations. Although, the downside is
the coding experience required to utilize
their functions and modules. For example,
there aren’t many platforms that let you
build a recursive neural tensor net (RNTN),
but you can code your own with the right
deep net library.
TECHNOLOGY
Multiple deep learning platforms are available online.

Deep learning is just getting started. New
applications emerge all the time. Here is a quick
look at prominent applications today.
12
C O P Y R I G H T © 2 0 1 6
Y
E
FOUR VALUE LEVERS
FOR BUSINESSES
APPLICATIONS
Allows images to be accessible through a standard
search through machine generated captions.
Many companies, including Facebook, use such
applications to scan pictures for faces at many
different angles, and then suggest the names of the
people in those pictures.
Used to recognize objects within images; allows
for images to be searchable based on the objects
within them.
Makes use cases such as driverless cars, remote
robots, and theft identification possible.
Used to extract relationships and facts from
text (fact extraction), automatically translate
text to other languages (machine translation),
and run analysis to understand people’s
opiniona (sentiment analysis) related to
global events, new products, politics, etc.
Use cases include vocal search, real-time
translation, and machine-generated subtitles.
Automatic Tagging
Object Recognition
Video Parsing
NATURAL LANGUAGE
PROCESSING (NLP)
SPEECH RECOGNITION
MACHINE VISION
weather in massachusettswhat is the
In the medical field, deep learning has a variety
of applications, including cancer detection,
drug discovery, and radiology. A team from
Stanford uses deep nets to identify over 6,000
factors that help predict the chances of a
cancer patient surviving. Another team of deep
learning experts from Toronto University and
Washington University won a drug discovery
challenge by Merck in 2012 [see page 4]. Deep
nets can also help detect tumors through the
use of MRI, fMRI, EKG, and CT scan data.
MEDICINE
Deep learning algorithms can be used to make
buy and sell predictions based on data inputs
from market movements, and design elements
such as risk profiling, portfolio allocations, short-
term vs. long-term trading.
FINANCE
Marketing and advertising have numerous
uses of deep learning, including sentiment
analysis, segmentation, cross selling, promotion
decisions, and personalized ads in real time.
MARKETING
AND ADVERTISING

13
C O P Y R I G H T © 2 0 1 6
FOUR VALUE LEVERS
FOR BUSINESSES
At this point there are so many sophisticated
deep learning algorithms out there that the
defining success factor for a business really
becomes the data – raw or
labeled.
Companies like Facebook and
Google are sitting on a wealth
of high quality data at scale
(behavioral, search, social, etc.),
and they are clearly way ahead
of others with computing,
intellectual, and financial
resources to make use of that
data for real life applications.
There are many other companies
that have large amounts of data,
but not all the resources to
benefit from it.
They may or may not understand
what to do with that data in the context of
deep learning. They might initially try to
label the data internally, and soon realize
the size and scope of that task and that they
don’t have the optimal resources to do it.
They then try to figure how to get their data
labeled externally, and to do so they go to
companies like Spare5 (a “Training Data as
a Service” company) and present them with
some examples of how they want their data
to be labeled (some annotations and specs).
From there, companies like Spare5 will
do the heavy lifting, use their network of
communities, efficient algorithm, and access
to other bigger datasets to do the data
labeling with quality at speed and scale.
Data can be the sustainable competitive
advantage for businesses.
Math and associated deep learning
algorithms will continue to improve, but
they will be more or less accessible to
everyone. Technology and its deep learning
libraries and platforms are easily available
to everyone too.
The defining success
factor for a business
really becomes the data
- raw or labelled.
DATA
Consumer data, such as online shopping behavior, can be used to train
deep learning networks.

14
C O P Y R I G H T © 2 0 1 6
However, having data at scale, and being able
to train, label, and augment that data (in-house,
through service-providers, or through strategic
partnerships) are becoming the most important
differentiators for businesses. And this is true
across industries.
Despite this strong point of differentiation,
executives sometimes need to be convinced
to pool their (big) data to get faster and
better results.
ConsiderMobileye,anIsraelicompanythatwants
competing carmakers to contribute on-the-road
data to help teach automated cars how to drive
safely. Mobileye supplies advanced computer
hardware and software to many automotive
manufacturers to enable cars to spot objects
on the road. The company is now developing
ways to train cars to drive themselves. This will
essentially be done by feeding computers huge
quantities of driving behavior data (sourced from
multiple carmakers) as training data into their
deep learning simulations.
Currently, many experimental self-driving cars
follow rules that are programmed manually,
which obviously makes it difficult to account for
every possible eventuality. Demonstrating the
continuing evolution, Mobileye has recently
been at the center of a controversy over the
limits of car automation, as its vision technology
is used in Tesla’s autopilot system, which was
involved in a car crash in Florida.
Executives sometimes
need to be convinced
to pool their (big)
data to get faster and
better results.
FOUR VALUE LEVERS
FOR BUSINESSES
DATA
Autonomous vehicles require the ability to recognize driving conditions and react accordingly.

15
C O P Y R I G H T © 2 0 1 6
FOUR VALUE LEVERS
FOR BUSINESSES
analysis methods are either insufficient or
sub-optimal in drawing insights from such
datasets. Deep learning is changing all of
that, given its application in a number of
healthcare areas.
Massachusetts General Hospital (MGH)
is one of the pioneers in this area, using
deep learning supercomputer to help
improve everything from detection to
diagnosis to treatment and disease
management, by training a deep neural
network on its repository imaging, genetic
and other data. Similar initiatives are seen
across other innovative hospitals in the
areas of pediatrics, Alzheimer’s, facility
management, etc.
Some of the interesting healthcare start-
ups using AI include:
- Medical Imaging & Diagnostics
- Baylabs, Arterys, Visexcell, Deep
Genomics, Entopsis, Zebra, Imagia,
Deep 6, SemanticMD, Behold.ai
- Wearables – Cyrcadia, Magnea,
physIQ, Sentrian, TinyKicks, QMedic
- Health & Lifestyle Management –
AiCure, Healint, Wellframe, Lucina,
Ovuline, PeerWell
- Mental Health – LifeGraph, Ginger.io,
TAO
- Drug Discovery – Numerate, Globavir,
Atomwise
- Virtual Assistants – Sensely, Your.MD,
Babylon, medwhat
- Insights and Risk Management –
Apixio, Pathway Genomics, lumiata,
Ensodata, Oncora
Healthcare most
attractive for Venture
Capitalists and
Corporate Investors
for AI investments
Equity financing in AI increased from
$282Min2011to$2.4Bnin2015,including
about $1Bn invested by corporate
investors across 90 deals. Most active
corporate investors in AI include Intel
Capital, Google Ventures, GE Ventures,
Samsung Ventures, and Bloomberg Beta.
Most active venture capital (VC) firms
in the space include Khosla Ventures,
Intel Capital, Data Collective, Google
Ventures, New Enterprise Associates,
Andreesen Horowitz, Formation 8,
Horizon Ventures, and Accel Partners.
Healthcare stands out compared to
other industries with about 60 AI deals
in 2015, accounting for about $260M.
Some of the most well-funded AI
companies in Healthcare are Welltok,
iCarbonX, Stratified Medical, Butterfly
Network, and Apixio.
Thereisnowexabytes(1millionterabytes)
of data in healthcare and it continues
to grow every day with the addition
of data from wearables, EHR systems
and other patient-generated systems.
Needless to say that the traditional data

16
C O P Y R I G H T © 2 0 1 6
Whether you are a deep learning start-up trying
to figure out how to build a company that
will be attractive to investors, a large corporation
strategizing future organic or inorganic growth
options building upon big data sitting in-house,
or an Investor evaluating hundreds of deep
learning deals a year, it is critical to understand the
components and the value levers of deep learning.
Ifyouarealargeglobalcorporationintheconsumer
goods industry, your strong deep learning value
lever is likely to be data (years of global data from
sales, customers, inventory, promotions, etc.).
It’s important to keep your data in mind while
strategizing for future growth options using deep
learning. Any acquisitions or partnerships should
ensure that the value levers of the two entities
are aligned.
Use deep learning value
levers to separate the signal
from the noise.
CONCLUSION

17
C O P Y R I G H T © 2 0 1 6
You may partner with Google or Apple for their
math, technology, or even data to augment
your own consumer behavior data.
Or you may partner with or acquire a predictive
consumer behavior analytics start-up, to help
create monetization opportunities with their
data. Both those options and potentially
many others are valid, but it’s important
to understand and evaluate the alignment
of the value levers to avoid surprises and
disappointment in the future.
If you are a predictive healthcare data
analytics start-up, your relevant and strong
deep learning value levers in an investment
pitch should mostly be about healthcare
applications or data, but only if you also
have data. In that case, trying to impress
investors by overemphasizing buzzwords
such as Convolutional Neural Nets (math)
or TensorFlows (technology) is not going to
make you look strong.
Investors who understand the right value
levers will understand very quickly that you
are highlighting something that’s not a
differentiator for your venture (anyone can use
CNN and TensorFlows). On the other hand, if
you are a “training data as a service” startup
like Spare5, your value lever is technology, not
math, or data. So, in a nutshell, stick to the
value lever that makes sense to you.
For obvious reasons, some of which are
mentioned in the examples above, it makes
sense for investors to consider detailed
review and evaluation of the value levers
as part of their investment due diligence
process. The fact that investors are focused
on the future, it becomes all the more
important for them to make investment
decisions across industries.
Needless to say, not putting deep learning
on your strategic agenda can be dangerous,
because chances are that your competitors
are already aggressively working on the issue.
Not putting deep learning on
your strategic agenda can be
dangerous, because chances
are your competitors are
already aggressively working
on the issue.
Stick to the value lever that makes sense for you.
CONCLUSION

18
C O P Y R I G H T © 2 0 1 6
1. MIT Technology Review. Sep 27, 2016. “Google’s New Service Translates
Languages Almost as Well as Humans Can.” https://www.technologyreview.
com/s/602480/googles-new-service-translates-languages-almost-as-well-as-
humans-can/
2. MIT Technology Review. “Deep Learning.” https://www.technologyreview.
com/s/513696/deep-learning/
3. MIT Technology Review. Oct 10, 2016. “An Ambitious Plan to Build a Self-
Driving Borg.” https://www.technologyreview.com/s/602531/an-ambitious-
plan-to-build-a-self-driving-borg/
4. Harvard Business Review. Oct 11, 2016. “Technology Will Replace Many
Doctors, Lawyers, and Other Professionals.” https://hbr.org/2016/10/robots-
will-replace-doctors-lawyers-and-other-professionals
5. MIT Technology Review. Sep 9, 2016. “The Extraordinary Link Between
Deep Neural Networks and the Nature of the Universe.” https://www.
technologyreview.com/s/602344/the-extraordinary-link-between-deep-
neural-networks-and-the-nature-of-the-universe/
6. MIT Technology Review. May 19, 2016. “Inside Vicarious, the Secretive AI
Startup Bringing Imagination to Computers.” https://www.technologyreview.
com/s/601496/inside-vicarious-the-secretive-ai-startup-bringing-
imagination-to-computers/
7. MIT Technology Review. Aug 30, 2016. “What Robots Can Learn from
Babies.” https://www.technologyreview.com/s/602246/what-robots-can-
learn-from-babies/
8. McKinsey. June 2015. “An executive’s guide to machine learning.” http://
www.mckinsey.com/industries/high-tech/our-insights/an-executives-guide-
to-machine-learning
9. McKinsey. Sep 2014. “Artificial intelligence meets the C-suite.”http://
www.mckinsey.com/business-functions/strategy-and-corporate-finance/our-
insights/artificial-intelligence-meets-the-c-suite
10. Forbes. Oct 7, 2016. “3 Steps To Jumpstart A Machine Learning Strategy.”
http://www.forbes.com/sites/maribellopez/2016/10/07/3-steps-to-jumpstart-
a-machine-learning-strategy/#4f8b90096d95
11. Fortune.com. Sep 2016. “Why deep learning is suddenly changing your
life.” http://fortune.com/ai-artificial-intelligence-deep-machine-learning/
12. CBinsights. Aug 31, 2016. “From Virtual Nurses To Drug Discovery: 90+
Artificial Intelligence Startups In Healthcare.” https://www.cbinsights.com/
blog/artificial-intelligence-startups-healthcare/
REFERENCES

19
C O P Y R I G H T © 2 0 1 6
13. Samsung Insights. July 2016. “Deep Learning: The Next Step in Applied
Healthcare Data.” https://insights.samsung.com/2016/07/12/deep-learning-
the-next-step-in-applied-healthcare-data/
14. New York Times. Nov 2012. “Scientists See Promise in Deep-Learning
Programs.” http://www.nytimes.com/2012/11/24/science/scientists-see-
advances-in-deep-learning-a-part-of-artificial-intelligence.html?_r=1
15. Nvidia. July 29, 2016. "What’s the Difference Between Artificial
Intelligence, Machine Learning, and Deep Learning?" https://blogs.nvidia.
com/blog/2016/07/29/whats-difference-artificial-intelligence-machine-
learning-deep-learning-ai/
16. Numenta. Jan 11, 2016. "What is Machine Intelligence vs. Machine
Learning vs. Deep Learning vs. Artificial Intelligence (AI)?" http://numenta.
com/blog/2016/01/11/machine-intelligence-machine-learning-deep-
learning-artificial-intelligence/
17. CBinsights. Oct 7, 2016. "The Race For AI: Google, Twitter, Intel, Apple In
A Rush To Grab Artificial Intelligence Startups." https://www.cbinsights.com/
blog/top-acquirers-ai-startups-ma-timeline/
Audio and Video Sources
1. This Week in Machine Learning. Sep 29, 2016. "Talk #6 - Angie Hugeback
(Principal Data Scientist at Spare5) – Generating Training Data for Your ML
Models." https://twimlai.com/twiml-talk-6-angie-hugeback-generating-
labeled-training-data-mlai-models/
2. This Week in Machine Learning. Oct 9, 2016. "Talk #7 – Carlos Guestrin -
(the Amazon professor of Machine Learning at the University of Washington) –
Explaining the Predictions of Machine Learning Models." https://twimlai.com/
twiml-talk-7-carlos-guestrin-explaining-predictions-machine-learning-models/
3. This Week in Machine Learning. Aug 5, 2016. "Apple Acquires Turi, the
DARPA Hacker-Bot Challenge and More." https://twimlai.com/apple-acquires-
ml-startup-turi-darpa-hacker-bot-challenge-comma-ais-autonomous-driving-
dataset-twiml-20160805/
4. Learning Machine 101. Feb 23, 2015. "How to Build a Deep Learning
Machine." http://www.learningmachines101.com/lm101-023-how-to-build-a-
deep-learning-machine-function-approximation/
5. Learning Machine 101. Oct 27, 2016. "How to Build a Machine that Can
Learn Anything (The Perceptron)." http://www.learningmachines101.com/
lm101-015-perceptron/
5. Changelog. Sep 9.2016. "219: TensorFlow and Deep Learning with Eli
Bixby (Developer Programs Engineer at Google)." https://changelog.com/
podcast/219
6. DeepLearning.TV. https://www.youtube.com/channel/UC9OeZkIwhzfv-_
Cb7fCikLQ
REFERENCES

Tarun Mehra is a Principal with
Fuld + Co. He regularly advises
Fortune 1000 companies and
PE clients (focus on Tech and
Healthcare) across issues
related to growth strategy,
market segmentation, M&A due
diligence, war gaming, scenario
planning, channel strategy, etc.
His prior consulting experience
includes working with Monitor Group, Accenture and i2
Technologies in a variety of senior roles. Tarun was the President
and CEO of Fuld Omniscope, a strategic research and consulting
assets business that he founded in 2013. He also independently
advises and invests in early stage technology and healthcare
companies. Tarun is based in Boston, MA.
C O P Y R I G H T © 2 0 1 6
ABOUT FULD
Fuld + Company is a leading competitive strategy consultancy
that helps clients anticipate competitive activity, see beyond
market disruptions, and develop or refine robust business
strategies. Through research, analysis, and strategic consulting
we work with the Global 1000 to identify and solve tactical and
strategic challenges.
With over 35 years of experience, and offices on three continents,
Fuld + Company developed many of the competitive intelligence
and strategic analysis techniques used today. Having completed
thousands of projects, we are recognized as an organization of
thought leaders by publications such as Fortune, Fast Company,
The Financial Times, The Economist and Time Magazine.
The right to use this white paper in its entirety, or any portion thereof, remains
exclusively that of Fuld + Company Inc. Upon request, Fuld + Company will,
at its discretion, grant permission to republish any of this material. This report
is not intended to be, and should not be construed as, a recommendation for
purchase or sale of any companies, or securities of any companies, mentioned
herein. The information has been derived from statistical and other sources
which we deem reliable, but their accuracy, and their completeness, cannot
be guaranteed. Opinions expressed herein are based on our interpretation of
available information, are subject to change, and should be considered strictly
as opinions.
US GLOBAL HEADQUARTERS
131 Oliver Street, 3rd Floor, Boston, MA 02110, United States
+1 617 492 5900
EUROPE
20 Conduit Street, London W1S 2XW, United Kingdom
+44 (0) 20 7659 6999
ASIA-PACIFIC
3402 One Corporate Centre, Julia Vargas cor. Meralco Ave.,
Ortigas Center, Pasig, Philippines
+63 2 706 3292
www.fuld.com
ABOUT THE AUTHOR

CompetitiveAdvantageThroughDeepLearning (white paper)

More Related Content

What's hot (11)

Viewers also liked (20)

Similar to CompetitiveAdvantageThroughDeepLearning (white paper) (20)

CompetitiveAdvantageThroughDeepLearning (white paper)