Artificial Intelligence

March 2018
Proposal title goes here |
 Section title goes here

What is AI? What is Machine

Learning? What is Cognitive
Analytics? How do these
terms relate, or differ, from
one another? In general
terms, AI refers to a broad
field of science encompassing
not only computer science
but also psychology,
philosophy, linguistics and
other areas.

02
 Artificial Intelligence | Contents

Contents

Artificial Intelligence Defined 04

Artificial Intelligence Techniques Explained 10

Applications of AI 16

Five technology trends that leap-frog

Artificial Intelligence 22

AI opportunities for the future 26

Authors 31

Sources 32

03
Artificial Intelligence | Artificial Intelligence defined

04
 Artificial Intelligence | Artificial Intelligence defined

Artificial Intelligence defined

The topic of Artificial Intelligence is at the top of its Hype curve1. And there are
many good reasons for that; it is exciting, promising and a bit scary at the same
time. Various publications are claiming that AI knows what we want to buy, it
can create Netflix series, it could cure cancer and it may eventually take our
jobs or even destroy mankind.2

The problem and at the same time Nevertheless, recent developments in “AI refers to a broad field
opportunity with AI is that it’s not very well AI are impressive and exciting. But also
defined. If we would show the navigation overestimated and misunderstood. In
of science encompassing
system of our car to someone living in order to split hype from reality and help not only computer science
1980, he or she would probably consider it forming a view on this market, we will
as a form of Artificial Intelligence, whereas publish a series of articles explaining the
but also psychology,
we nowadays would probably not. We are world of AI, zooming in to the techniques philosophy, linguistics and
seeing the same with speech and image that are associated with AI, the most
recognition, natural language recognition, appealing business applications and
other areas”
game engines and other technologies that potential issues we can expect.
are becoming more and more common In this first article we will start with the
and embedded in every-day technology. beginning, by explaining AI and associated
terms in five definitions. What is AI? What
On the other hand, various technology is Machine Learning? What is Cognitive
solution providers are taking the Analytics? How do these terms relate, or
opportunity to rebrand their existing differ, from one another?.
solutions to AI, to take advantage of
the huge hype and that the market is Artificial Intelligence (AI)
experiencing and the resulting press In general terms, AI refers to a broad
coverage. If we have a built a machine field of science encompassing not only
learning model that predicts customer computer science but also psychology,
demand, a solution that has been existing philosophy, linguistics and other areas.
for years, we would have called it “data AI is concerned with getting computers
mining” in the past and we now see it to do tasks that would normally require
rebranded as “artificial intelligence”. This is human intelligence. Having said that, there
adding to the confusion and may very well are many point of views on AI and many
lead to inflated expectations. definitions exist. Below some AI definitions
which highlight key characteristics of AI.

1
https://www.gartner.com/document/3380751/
2
http://www.bbc.com/news/technology-37713629
2
https://futureoflife.org/ai-open-letter/

05
Artificial Intelligence | Artificial Intelligence defined

BIG DATA REASONING

Capable of processing massive Ability to reason (deductive or

amounts of structured and inductive) and to draw inference
unstructured data which can based to the situation. Context
change constantly driven awareness of the system.

Capable of analyzing and

Ability to learn based on historical solving complex problems in
patterns, expert input and special-purpose and general-
feedback loop purpose domain

LEARNING PROBLEM SOLVING

Figure 1: Key characteristics of an AI system

Some general definitions: Narrow AI vs General AI

A chess computer could beat a human
•• “Artificial intelligence is a computerized
in playing chess, but it couldn’t solve a
system that exhibits behavior that
complex math problem. Virtually all current
is commonly thought of as requiring
AI is “narrow”, meaning it can only do what
intelligence.”3
it is designed for. This means for every
•• “Artificial Intelligence is the science of problem, a specific algorithm needs to be
making machines do things that would designed to solve it. Narrow AI are mostly
require intelligence if done by man.” 4 much better at the task they were made for
than humans, like face recognition, chess
The founding father of AI Alan Turing
computers, calculus, translation. The holy
defines this discipline as:
grail of AI is a General AI, a single system
•• “AI is the science and engineering of that can learn about any problem and then
making intelligent machines, especially solve it. This is exactly what humans do:
intelligent computer programs”. 5 we can specialize in a specific topic, from
abstract maths to psychology and from
In all these definitions, the concept of
sports to art, we can become experts at all
intelligence refers to the ability to plan,
of them.
reason and learn, sensing and building
some kind of perception of knowledge and
An AI system combines and utilizes mainly
communicate in natural language.
machine learning and other types of data
analytics methods to achieve artificial
intelligence capabilities.

3
Preparing for the Future of Artificial Intelligence, NSTC, 2016
4
6. Raphael, B. 1976. The thinking computer. San Francisco, CA: W.H. Freeman2
5
http://www-formal.stanford.edu/jmc/whatisai/node1.html

06
 Artificial Intelligence | Artificial Intelligence defined

Artificial Intelligence
Ability to sense, reason, engage
and learn
Computer vision Robotics & motion

Natural
language Planning &
processing optimization
Machine Learning
Ability to learn
Voice Unsupervised learning Knowledge
recognition capture
Supervised Reinforcement
learning learning
Methods
Ability to reason
Regression
Decision trees
etc.

Technologies
Physical enablement
Platform
UX
APIs
Sensors
etc.

Figure 2: relation between AI, Machine Learning and underlying methods and infrastructure

Machine Learning Depending on the technique used, an Cognitive Analytics

Machine learning is the process whereby algorithm can improve itself by adding a Cognitive Analytics is a subset of A.I.
a computer distills meaning by exposure feedback loop that tells it in which cases it that deals with cognitive behavior we
to training data6. If for example you want made mistakes. associate with ‘thinking’ as opposed to
an algorithm to identify spam in e-mails, perception and motor control. Thinking
you will have to train the algorithm by The difference with AI however is that allows an entity to obtain information from
exposing it to many examples of e-mails a machine learning algorithm will never observations, learn and communicate.
that are manually labeled as being spam or “understand” what it was trained to do. It
not-spam. The algorithm “learns” to identify may be able to identify spam, but it will not A cognitive system is capable of extracting
patterns, like occurrence of certain words know what spam is or understand why we information from unstructured data by
or combination of words, that determines want it to be identified. And if there is a extracting concepts and relationships into
the chance of an e-mail being spam. new sort of spam emerging, it will probably a knowledge base. For example, from a text
Machine learning can be applied to many not be able to identify it unless someone about Barack Obama, the relations from
different problems and data sets. You can (human) re-trains the algorithm. Figure 3 can be extracted using Natural
train an algorithm to identify pictures of Language Processing. 80% of all company
cats in photo-collections, potential fraud Machine learning is at the basis of most data is unstructured and current Cognitive
cases in insurance claims, transform AI systems. But while a machine learning Analytics systems can search all of it to find
handwriting into structured text, transform system may look “smart”, in our definition the answer to your question.
speech into text etc. All these examples of AI it is in fact not.
would require labeled training sets.

6
Stephen Lucci, 2016, Artificial intelligence in the 21st century : A living introduction

07
Artificial Intelligence | Artificial Intelligence defined

Mexico

Neighbour

USA

President

Barack Obama
Spouse
Michelle Obama
Father

Father
Mother
Mother

Malia Obama Sibbling Sasha Obama

Figure 3: A knowledge base extracted from text

Learning enables the Cognitive System Another important aspect of Cognitive

to improve over time in two major ways. Analytics is the ability to use context.
Firstly, by interacting with humans, and Context enables a Cognitive Analytics
obtaining feedback from the conversation system to infer meaning from language. For
partner or by observing two interacting example, a chatbot can take into account
humans. Secondly, from all the data in the the conversation history to infer who is
knowledge base, new knowledge can be referred to by the word he:
obtained using inference.

User: Who is Obama’s wife?

AI: Michelle Obama.
User: How old is he?
AI: Barack Obama is 55 years old.

Figure 4: Example conversation of a cognitive system

08
 Artificial Intelligence | Artificial Intelligence defined

For this simple exercise, the system needs are is vivid if we look at robots developed Since Smart Machines are autonomous and
to be aware of names that represent by Boston Dynamics8 and MIT’s Cheetah II9. intelligent, they might start communicating
people, relationships between people, Other example is Amelia10, an intelligent among themselves. This leads to multi-
gender and the common sense to infer assistant with NLP capabilities. Key aspect agent systems that can make trades
that Obama refers to Barack Obama. All of of robotics is that it combines hardware to improve their utility. The building-
this contextual information is required to (mechanical parts, sensors, screens etc.) inspecting robot can ask a drone to inspect
make the right inferences to answer both with intelligent software and data to the roof for him, trading this favor for
questions. perform a task for which certain level of another favor, like transporting goods or
intelligence is required (e.g. orientation, simply currency.
Since Cognitive Systems are aware of motion, interaction etc.).
context, can understand unstructured A Cognitive System that becomes a Smart
data and reason about information, they Smart Machines Machine can specialize in a specific area,
can communicate with humans as well. The major theme in using the term “Smart becoming an expert in that area. Now,
This enables the system to understand Machines” is autonomy. Smart Machines other Smart Machines can ask it for
a question posed in English, no longer are systems that –to some extend- are able information in that area, and it will be able
requiring the time-consuming process of to make decisions by themselves, requiring to provide more relevant answers more
converting the question into a format the no human input. Cognitive Analytics quickly than a general Cognitive System
computer can work with. For example, systems can be Smart Machines, as well that is not specialized. Information brokers
a call center representative cognitive as robots, or any kind of AI, as long as it like this improve the overall utility of the
system can quickly answer a customer’s adheres to this rule. In the case of a robot, whole network of Smart Machines.
question about camping gear by using autonomy could consists of a capability to
information from product descriptions, plan where it wants to go, what it wants to Conclusion
customer reviews, sales histories, topical achieve and how to overcome obstacles. The terms Machine Learning, Cognitive,
blogs, and travel magazines.7 Cognitive Rather than being human-controlled or Robotics and smart machines are used
Systems can understand and communicate simply following instructions, it could often in relationship to AI, or sometimes
through many mediums, including speech, achieve higher-level goals like getting even as synonyms. AI is a complex field
image, video, sign language, graphs or any groceries, inspecting buildings and so forth. of interest, with many shapes and forms.
combination of these. This is enabled by planning methods, self- Therefore we have tried to shine some
preservation instincts on top of the skills light on the most used terminology. In
Robotics that a normal robot already requires. subsequent blogs, we will dive deeper in
AI is an important enabling factor in design In the case of a Cognitive System, it will techniques behind AI systems, business
and operationalizing smart robots and pro-actively try to learn new facts, gauge applications, some associated technology
other process automation applications. opinions and learn new common sense trends and the top 5 risks and concerns.
In its most simple form, a robot may be a rules by engaging in active conversation
machine that is programmed to perform with humans, asking questions and
a simple task, by following step-by-step double-checking them with data found
instructions. It could consist of a rule-based online. It will also actively inform decision
engine that explicitely tells the system what makers about changes it has observed, for
to do when a certain condition occurs. A example if the opinion of customers on
robot in a car factory Is programmed like social media suddenly makes a swing. It
that and hardly considered “intelligent”. could even act upon these changes, in the
But robotics exist in a variety of much more example engaging with the customers or
intelligent shapes, ranging from unmanned sharing the positive opinions on the social
autonomous vehicles (UAV’s), drones, smart media outlets of the company.
vacuum cleaners to intelligent chatbots and
smart assistants etc. How advanced robots

7
Deloitte, Cognitive analytics™ The three-minute guide
8
http://www.bostondynamics.com/
9
https://biomimetics.mit.edu/research/dynamic-locomotion-mit-cheetah-2
10
http://www.ipsoft.com/amelia/

09
Artificial Intelligence | Artificial Intelligence techniques explained

10
 Artificial Intelligence | Artificial Intelligence techniques explained

Artificial Intelligence
techniques explained

In order to ‘demystify’ Artificial Intelligence, and in some way get more people
involved in AI, we are publishing a series of articles explaining the world of AI,
zooming in on the techniques that are associated with AI, the most appealing
business applications, and potential issues we can expect.

The first blog article explained some of changing problem. Now a greedy heuristic
the most commonly used definitions of would be to always choose the largest
AI. In this second article we will explain denomination possible and repeat this
some fundamental AI techniques used: until we get to the desired value of 7. In
Heuristics, Support Vector Machines, our example, that means that we would
Neural Networks, Markov Decision Process, start with first selecting one 5 cent coin.
and Natural Language Processing. For the remaining 2 cents, the largest
denomination we can choose is 1 cent,
Heuristics leaving us with the situation where we still
Suppose we have coins with the following have to cover 1 cent for which we again
denominations: 5 cents, 4 cents, 3 cents, use 1 cent.
and 1 cent and we need to determine the
minimum number of coins to get 7 cents. So our greedy heuristic gives us a solution
In order to solve this problem we can make of 3 coins (5, 1, 1) to get to the value of
use of a technique called “heuristics”. 7 cents. It can be easily seen that another,
better, solution of only 2 coins exist using
Webster1 defines the term Heuristic the 3 and 4 cent coins. While the greedy
as “involving or serving as an aid to heuristic for the coin changing problem
learning, discovery, or problem-solving does not provide the best solution for this
by experimental and especially trial and particular case, in most cases it will result in
error methods”. In practice, this means a solution that is acceptable.
that whenever problems get too large or
too complex to find the guaranteed best Besides such tailored heuristics for specific
possible solution using exact methods, problems, also certain generic heuristics
heuristics are a way to employ a practical exist. Just like neural networks, some of
method to find a solution that is not these generic heuristics are based on
guaranteed to be optimal, but one that is processes in nature. Two examples of
sufficient for the immediate goals. such generic heuristics are Ant Colony
Optimization2 and genetic algorithms3.
For some problems, tailored heuristics The first is based on how simple ants are
can be designed that exploit the structure able to work together to solve complex
present in the problem. An example of problems and the latter is based on the
such a tailored heuristic would be a greedy principle of survival of the fittest.
heuristic for the above mentioned coin-

1
https://www.merriam-webster.com/dictionary/heuristic
2
https://en.wikipedia.org/wiki/Ant_colony_optimization_algorithms
3
https://en.wikipedia.org/wiki/Genetic_algorithm

11
Artificial Intelligence | Artificial Intelligence techniques explained

Figure 1

1
3

Figure 3 Figure 2

A typical problem where heuristics are the boundary line creates a maximum The two dotted lines are the two parallel
applied to find acceptable solutions quickly separation between the classes. To separation lines with the largest space
is vehicle routing, where the objective is to demonstrate this, we will use the following between them. The actual classification
find routes for one or more vehicles that simple data for our classification problem boundary that is used will be the solid
have to visit a number of locations. (Figure 1). line exactly in the middle of the two
dotted lines.
Support Vector Machines In this example, the green circles and the
The question whether an email is spam or red squares could represent two different The name Support Vector Machine comes
not spam is an example of a classification segments in a total set of customers (e.g. from the data points directly on either of
problem. In these types of problems, the high potential and low potential), based these lines are the supporting vectors. In our
objective is to determine whether a given on all kinds of properties for each of the example, we had 3 supporting vectors.
data point belongs to a certain class or not. customers. Any line that keeps the green
After first training a classifier model on data circles on the left and the red squares on If any of the other data points (i.e. not a
points for which the class is known (e.g. the right is considered a valid boundary supporting vector) is moved slightly, the
a set of e-mails that are labeled as spam line for the classification problem. There dotted boundary lines are not affected.
or not spam), you can then use the model is an infinite number of such lines that can However, if the position of any of the
to determine the class of new, unseen be drawn and 4 different examples are supporting vectors is slightly changed (e.g.
data-points. A powerful technique for presented on top (Figure 2). data point 1 is moved slightly to the left),
these types of problems is Support Vector the position of the dotted boundary lines
Machines4 (SVM). As stated before, with SVM you try to find will change and therefore the position of
the boundary line that maximizes the the solid classification line also changes.
The main idea behind SVM is that you try separation between the two classes. In
to find the boundary line that separates the provided example, this can be drawn
the two classes, but in such a way that as Figure 3:

4
https://en.wikipedia.org/wiki/Support_vector_machine

12
 Artificial Intelligence | Artificial Intelligence techniques explained

In real life, data is not as straightforward Biological neuron Artificial neuron

as in this simplified example. We
normally work with much more than
two dimensions. Besides having X1

straight separation lines, the underlying

mathematics for an SVM also allows for
X2 Output
certain type of calculations or kernels that
result in boundary lines that are non-linear.
X3
SVM classification models can also be
found in image recognition, like face
recognition or converting handwriting Figure 4: Graphical representation of a biological neuron (left) and an artificial
to text. neuron (right)

Artificial Neural Networks Where in the biological brain learning takes all the inputs are received, processed and
Animals are able to process (visual or other) place by strengthening or weakening the converted to outputs to the next layers.
information from their environment and bonds between different neurons, in the The hidden layers consist of one or more
react adaptively to a changing situation. ANN the learning takes place by changing layers of neurons each passing through
They use their nervous system to perform the weights between the neurons. By inputs and outputs. Finally, the output layer
such behavior. Their nervous system can providing the neural network with a large receives inputs of the last hidden layer and
be modeled and simulated and it should be set of training data with known features the converts this to the output for the user.
possible to (re)produce similar behavior in best weights between the artificial neurons
artificial systems. Artificial Neural Networks (i.e. strength of the bond) can be calculated Figure 2 shows an example network where
(ANN) can be described as processing in order to make sure the neural network all neurons in one layer are connected
devices that are loosely modeled after the best recognizes the features. to all neurons in the next layer. Such a
neural structure of a brain. The biggest The neurons of the ANN can be structured network is called fully connected. Depending
difference between the two is that the into several layers5. Figure 5 shows an on the type of problem you want to solve
ANN might have hundreds or thousands illustrative scheme of such layering. This different connection patterns are available.
neurons, whereas the neural structure of network consists of an input layer, where For image recognition purposes, typically
an animal or human brain has billions.

The basic principle of a neural structure Hidden layers

is that each neuron is connected with
a certain strength to other neurons.
Based on the inputs taken from the
Output layer
output of other neurons (also considering
the connection strength) an output is
generated which can be used as input Input layer
again by other neurons, see Figure 1 (left).
This basic idea has been translated into an
artificial neural network by using weights
to indicate the strength of the connection
between neurons. Furthermore, each
neuron will take the output from the
connected neurons as input and use a
mathematical function to determine its
output. This output is then used by other Figure 5: Schematic of a connected ANN
neurons again.

5
http://neuralnetworksanddeeplearning.com/chap1.html

13
Artificial Intelligence | Artificial Intelligence techniques explained

Convolutional networks are used, where •• Once the MDP is defined, a policy can be -1

only groups of neurons from one layer are trained using “Value iteration” or “Policy
connected to groups of neurons in the next Iteration”. These methods calculate the
layer. For speech recognition purposes, expected rewards for each of the states. +3

typically Recurrent networks are used, which The policy then gives the best action that
allow for loops from neurons in a later layer can be taken from each state. -3

back to an earlier layer. -1 -1

As an example, we will define a grid which

Markov Decision Process can be seen as an ideal, finite world for -10 +10
A Markov Decision Process (MDP) is a a robot7. This example grid is shown in
framework for decision-making modeling Figure 6.
where in some situations the outcome
-1
is partly random and partly based on The robot can move (action) from each
the input of the decision maker. Another position in the grid (state) in four directions,
Figure 6: Example – grid world of
application where MDP is used is planning, namely north, left, right and south. The
a robot
where the planning is optimized. The probability that the robot goes into the
basic goal of MDP is to find a policy for the desired direction is 0.7 and 0.1 if it goes
decision maker, tells him which particular towards any of the other 3 directions. A Natural Language Processing
action should be taken at which state. reward of -1 (i.e. a penalty) is given if the Natural Language Processing, or NLP in
robot bumps into a wall and doesn’t move. short, is a term for everything from speech
An MDP model consists of the Also, additional rewards and penalties are recognition to language generation, each
following parts6: given if the robot reaches the cells that are requiring different techniques. A few of
colored green and red, respectively. Based the important techniques will be explained
•• A set of possible states: for example,
on the probabilities and rewards a policy below, which are Part-of-Speech tagging,
this can refer to a grid world of a robot
(function) can be made using the initial and Named Entity Recognition, and Parsing.
or the states of a door, door open and
final state.
door closed.
Let us examine the sentence “John hit
•• A set of possible actions: a fixed set of Another example where MDP can be the can.” One of the first steps of NLP is
actions a robot for example can take, used is the inventory planning problem, lexical analysis, here a technique is used
such as going north, left, south or west. where a stock keeper or manager each called Part-of-Speech (POS) tagging. With
Or with respect to a door, closing or week has to decide how many units have this technique every word is tagged to
opening door. to be ordered. The inventory planning can correspond to a category of words which
be modeled as an MDP, where the states have similar grammatical properties,
•• Transition probabilities: this is the
can be considered positive inventory and based on its relationship with adjacent and
probability of going from one state to
shortages. Possible actions are for instance related words. Not only words are tagged,
another state. For example, what is
ordering new units or backlogging to the but also paragraphs and sentences. Part-
the probability that the door is closed,
next week. Transition probabilities can be of-speech tagging is mainly done with
after the action of closing the door
considered as the action that will be taken statistical models, which give probabilistic
is performed.
based on the demand and inventory for results instead of hard if-then rules, and
•• Rewards: these are used to guide the the current week. Rewards, or in this case, is therefore more capable of processing
planning. With respect to the robot and costs are typically unit order costs and unknown text. Also, they can cope with the
the grid example, a robot may want to inventory costs. possibility of multiple possible answers,
move north to reach its destination. instead of only one. A technique which is
Actually going north will result in a often used for tagging is a Hidden Markov
higher reward.

6
https://www.cs.rice.edu/~vardi/dag01/givan1.pdf
7
http://artint.info/html/ArtInt_224.html

14
 Artificial Intelligence | Artificial Intelligence techniques explained

Model (HMM). An HMM is similar to the A final technique to be discussed is called S

Markov Decision Process, where each state Parsing (Syntactic Analysis). With parsing,
is a part of speech and the outcome of the text is analyzed for grammar and
the process is the words of the sentence. the arranging of the words so that the N VP
HMM’s ‘remember’ sequences of words relationship between the words is clear.
that came before and can, based on this, This is done by using the Part-of-Speech tag
V NP
make better estimates of what POS a from the lexical analysis and then grouping
word is. For example: ‘can’ in ‘the can’, is these into small phrases, which in turn can
more likely to be a noun than a verb. The also be combined with other phrases or D N
end result is that the words are tagged as words to make a slightly longer phrase. This
followed: ‘John’ as a noun (N), ‘hit’ as a verb is done until the goal is reached: to have
(V), ‘the’ as a determiner (D) and ‘can’ as a used every word in the sentence. The rules John hit the can.
noun (N) as well. how the words can be grouped are called
Figure 7
the grammar and can take a form like this:
Named Entity Recognition or NER, is similar D+N = NP, which reads: a Determiner +
to POS tagging. Instead of tagging words Noun = Noun Phrase. The final result is
with what function the word has in the depicted in the figure 7.
sentence (POS), words are tagged with
the type of entity the word represents. Conclusion
These entities can be for example person, The techniques used within the domain
company, time, location, etc. But also of Artificial Intelligence are, when you
more specialized entities such as gene, dive into them, just advanced forms of
or protein. Although an HMM can also be statistical and mathematical models. All
used for NER, the technique of choice is a these models cleverly put together give us
Recurrent Neural Network (RNN). An RNN tools to compute tasks, previously thought
is a different type of neural network as to be reserved for humans. In subsequent
discussed earlier, but it takes sequences blogs we will dive deeper in business
as input, that is a number of words in applications, some associated technology
a sentence, or complete sentences, trends, and the top 5 risks and concerns.
and remembers the output from the
previous sentence.8 In the sentence we
are looking at, it will recognize John as the
entity ‘person’.

8
More on RNN: http://karpathy.github.io/2015/05/21/rnn-effectiveness/

15
Artificial Intelligence | Applications of AI

16
 Artificial Intelligence | Applications of AI

Applications of AI

In order to ‘demystify’ Artificial Intelligence (AI), and in some way get more
people involved in it, we are publishing a series of articles explaining the
world of AI, zooming in on the techniques that are associated with it, the most
appealing business applications, and potential issues.

In our second blog article, we covered claim. The insurance employee then only
some fundamental AI techniques. In needs to do a quick check before approving
this article we take the next step and the recommendation. This can save a lot of
will discuss 5 applications of AI: Image time and increase the quality of the work.
recognition, Speech recognition, automatic This is just one example. In the remainder
translation, Q&A and Games. of this article we will discuss 5 exciting
applications in which we will see great
Introduction development in the coming years:
For a lot of tasks, computers are becoming
•• Image recognition
or are already better than humans.
The calculator on your cellphone for •• Speech recognition
example, can calculate the most difficult
•• Translation
multiplications in an instance that would
take people a few minutes and a piece •• Q&A
of paper and pen. It has already been
•• Games
a long time since the world champion
of chess was able to beat the strongest
These developments will make applications
chess computers. In the past few years
cheaper and more accurate, opening the
however, computers started doing
door for business to use them.
typically human things such as seeing and
hearing, which they do as good as, if not
Image recognition
better than people. Google can recognize
Recognizing images is an easy task for
exactly which things are present on an
most of us. We don’t have any trouble
image and Siri can understand what
differentiating a car from a tiger or
you are saying, to name a few examples.
recognizing that a car is still a car when
In the coming years the intelligence
you observe it from the front instead of
of these applications will increase in a
from the side. This task has been proven
rapid pace, due to developments in the
considerably more difficult for computers,
field of neural networks and increased
but recent progress in image recognition
computational power.
accuracy has resulted in interesting
applications. Because different vendors
The increased accuracy, availability,
like Google and IBM are offering their
and ease of implementation of artificial
preprogrammed algorithms open source
intelligence methods creates opportunities
and software libraries like Tensorflow1
for companies to apply them in their
make it possible to construct your own
business. As an example, insurance
algorithms, visual recognition is becoming
companies use AI to read claims from
more accessible for the public.
their clients, understand if the claim
is difficult or easy, and it can give a
recommendation on how to handle the

1
https://www.tensorflow.org/

17
Artificial Intelligence | Applications of AI

Well-known applications of image and tested it on a large publicly available to observe their behavior as an effect of
recognition are Google’s shopper app2 dataset, which they describe in a pre-print product displays, or the traffic flow in the
or facial recognition for security cams. of the article5. The vision of IBM is that at store. This is a familiar concept in web
Such applications are already using image a certain point medical staff can send a shops, where A/B testing can be used
recognition on a daily basis, however there picture of skin irregularities to Watson, the to see which website display yields the
has been a lot of development in other same way that they send blood sample to best profits, however this suggests that
areas over the last few years. IBM Watson, the lab. facial recognition tools can be used to
which we know from playing Jeopardy3, orchestrate these tests live in a store.
has developed its image recognition skills Facial recognition, which we mainly know Different software development
in the field of medicine. IBM Research has from security cameras, has also been companies are offering facial recognition
been working on deep learning techniques developed in other areas6. A survey of 150 for retailers7/8. They apply specific
for computer vision4 that could be used to retail executives by Computer Services algorithms that use facial landmarks to
recognize whether skin irregularities are Corporation, which was held in the UK in recognize and distinguish between faces,
melanoma. They created an ensemble of 2015, suggested that a quarter of all British which can be saved and later matched
methods that can segment skin lesions shops use facial recognition software. to enhance customer experience and
and methods that can detect the area The software is used for security, as one personalize service.
and surrounding tissue for melanoma might expect, but also to track customers

Figure 1: A mannequin of Saks Fifth Avenue, an early adaptor of the facial recognition software in retail22.

2
https://google-shopper.en.softonic.com/android
3
http://www.techrepublic.com/article/ibm-watson-the-inside-story-of-how-the-jeopardy-winning-supercomputer-was-born-and-what-it-wants-to-do-next/
4
https://www.ibm.com/blogs/research/2016/11/identifying-skin-cancer-computer-vision/
5
https://arxiv.org/abs/1610.046623 http://artint.info/html/ArtInt_224.html
6
https://www.theguardian.com/cities/2016/mar/03/revealed-facial-recognition-software-infiltrating-cities-saks-toronto
7
https://www.scnsoft.com/case-studies/facial-recognition-for-retail
8
https://www.facefirst.com/industry/retail-face-recognition/

18
 Artificial Intelligence | Applications of AI

Speech recognition
Voice Word Accuracy Rates Improving Rapidly...
Speech recognition is an AI application that +90% Accuracy for Major Platforms
recognizes speech and can turn spoken
words into written words. It is hardly Word Accuracy Rates by Platform*, 2012 – 2016
used on its own but it is largely used as *Word accuracy rate definitions are unique to each company...see footnotes for more details

an addition to Chatbots, virtual agents

100%
and mobile applications. Well known 90%
examples are Apple’s Siri, Google Home 80%
Word Accuracy Rate (%)

and Microsoft’s Alexa. Speech recognition 70%

started already in 1952 with ‘Audrey’ 9. 60%

Audrey was able to recognize digits spoken 50%

40%
in a single voice, which is quite impressive
30%
given the computers back then. Today we
20%
have applications on our phone and in our 10%
car that can respond to our voice. 0%
3
Baidu 1 Google 2 Hound Voice Search
(2012 - 2016) (2013 - 2015) & Assistant App
Not only the amount of applications with (2012 - 2016)
voice recognition capacity has increased, Source: Baidu, Google, VentureBeat, SoundHound
Note: *Word Error Rate (WER) definitions are specific to each company. Word accuracy rate = 1 - WER. (1) Data shown is word accuracy rate on Mandarin speech recognition on one of Baidu's speech
tasks. Real world mobile phone speech data is very noisy and hard for humans to transcribe. A 3.5% WER is better than what most native speakers can accomplish on this task. WER across different

also the accuracy of voice to words has

datasets and languages are generally not comparable. (2) Data as of 5/15 and refers to recognition accuracy for English language. Word error rate is evaluated using real world search data which is
extremely diverse and more error prone than typical human dialogue. (3) Data as of 1/16 and refers to recognition accuracy for English language. Word accuracy rate based on data collected from
thousands of speakers and real world queries with noise and accents.

improved dramatically over the last few

Figure 2: A table from the KPCB report of internet trends about the incline in
years according to KPCB10. accuracy of speech recognition10

One of the business applications that has

gained quite a lot of ground is the use of
Translation The usage of deep learning has had a
speech recognition in health care11. A lot of
A different topic with large business significant impact on the quality of machine
physicians are working with an electronic
implications is automatic translation. translations by completely shifting the
health record (EHR) to document patient
This topic can be defined as the process paradigm. Rather than working in a rule-
information, however this has been said
of translating text from one language to based way, powered by human decision
to delay the consults and to restrict the
another by using software. Traditionally, making, translation using a neural network
patient narrative. Using speech recognition,
translation was done by substituting each is completely based on mathematics. On
patient documentation can be recorded
word by its closest counterpart in the other relatively basic texts, the GNMT system
in a flexible and fast manner, which allows
language. While this works reasonably translations approach the quality of
the physician to pay more attention to the
well for single words, a pair of words or human translators. An experiment even
patient. This solution is already offered by
sentences are generally harder to process showed that when you translate English
different vendors such as Nuance12 and
correctly due to the fact that relations to Korean and subsequently translate
M*Modal13.
between words are important for the English to Japanese, the model is able to
meaning of a sentence, but such nuances translate Korean to Japanese reasonably
The rapid improvement of speech
cannot be captured when each word is well, without any prior training focused on
recognition accuracy offers a lot of
analyzed separately. the formal link between the two languages.
opportunities in the near future. Having
One article even asked the question: “have
all our soft- and hardware voice controlled
computers invented their own internal
might not be as far away as many people
language?”8.
think.

9
http://www.pcworld.com/article/243060/speech_recognition_through_the_decades_how_we_ended_up_with_siri.html
10
https://www.slideshare.net/kleinerperkins/2016-internet-trends-report/119-KPCB_INTERNET_TRENDS_2016_PAGE119Voice
11
http://www.healthcareitnews.com/news/speech-recognition-proving-its-worth
http://www.cleverbot.com/
12
https://www.nuance.com/healthcare/physician-and-clinical-speech/dragon-medical.html
13
https://mmodal.com/speech-solutions/

19
Artificial Intelligence | Applications of AI

The impact of quality translations in a something that can be seen in many

global economy are enormous. With business applications. After an intent is
business translations originally dominated distilled from the command of a user, it
by conversions between European can be linked to a specific follow up action.
languages, the need for translations This action can range anywhere from
to Chinese, Japanese, and Korean is asking a return question to the retrieval
increasing9. A simple example is one that of information from the internet. In the
Uber was investigating, where automatic area of customer service, Chatbots are
translation takes place between you quickly becoming the norm, one example
and your local Uber driver, who can only being IPsoft’s Amelia12. Standard queries
communicate in Chinese. are already handled automatically, with
only the difficult ones being forwarded
Question answering to human decision makers. Question
Q&A agents or Chatbots are another answering has also made an introduction
example of applying AI to language. in the field of Law, where lawyers can
When talking about the ability of having pose questions in natural language to an
conversations, distinctions are made in intelligent assistant about legal cases. The
the domain and the way of generating assistant can respond to the query with the
an answer of the agent. A chatbot can relevant passage, drawn from high quality
be focused on answering questions legal documentation13.
in an open or closed domain. When it
operates in an open domain, it should Game/ Solver
be able to answer general questions that One of the most exciting applications of AI
can concern any topic (see for example lies in playing games. Playing a game well
cleverbot). This is generally harder than requires you to not only know the rules, but
a closed domain, which concerns only a to calculate the next possible moves within
limited amount of topics. Closed domains, these rules, and finally make a careful
however, have very good business judgement on which move would give you
application such as answering questions to best chance to win. If computers can
at helpdesks. A couple of years ago, there play games as well as human players, there
was a breakthrough in question answering are no reasons why they cannot learn any
interest, when IBM Watson beat humans in other difficult task that people do in their
a game of Jeopardy, a well-known American daily work (although human supervision
quiz show10. More recently another probably remains needed).
breakthrough was made by Google,
who can now give chatbots the ability to Recently there was a big step forward in the
have a short term memory, which gives field of games when the world Champion
the chatbot the ability to mimic real-life of Go was beaten by a computer for the
conversations more realistically11. first time. Go is a game that cannot be
brute-force calculated, since the number of
The ability of machines to recognize intent possible moves is higher than the number
(or the purpose) from a question and of stars in the universe. The top Go players
to answer it in a variety of ways, is again of the world rely for a large part on their

20
 Artificial Intelligence | Applications of AI

intuition to come to the best moves. large datacenters. The AI controls over
Google’s AlphaGo (a neural network based 120 variables in Google’s datacenters,
go-engine), however, learned how to play such as the windows, fans and cooling
like a top human player by studying millions systems, optimizing for energy usage while
of human games. It then became even keeping computing performance up. The
stronger by playing against another version optimization potentially lowers Google’s
of itself millions of times, which finally energy bill for hundreds of millions over
enabled it to beat the world champion. several years. Another application lies in
If computers can beat human players healthcare, where an app of Deepmind
in one of the most complicated games saves nurses over two hours of time per
that currently exists, then where do the day by warning about upcoming acute
possibilities for AI stop? kidney failure. These are two applications,
but since this is a newly developed field
One big advantage people still have there is a huge potential for more. Think
over computers, is that we can take our about predicting stock prices, optimizing
knowledge and training in one area, and the layout of distribution centers.
apply it to a new task or area. For example, Imagination and available data are the limit.
good go players can apply their way of
thinking to solve their daily problems in Conclusion
their jobs. AlphaGo cannot do this: it is We have discussed a few applications
only good at playing go and nothing else. of AI that will become more advanced
When you make it learn something else, the coming years and that will probably
like chess, it will lose its ability to play go. become more and more common in our
Recently however, a first step was taken in daily lives. In the series of AI blogs we
overcoming this problem: neural networks have tried to shed light on applications of
are now able to remember to most AI that everyone will encounter at some
important knowledge from one game, and point and give some examples of where
at the same time learn a new game (link). these applications might lead to in the near
Google Deepmind wrote a new algorithm future. In the next blog we will address 5
that allowed a neural network to learn 10 trends that leapfrog AI adoption.
Atari games at the same time, and play
them with human performance.

Once this field will be more developed,

computers will be able to perform series
of difficult tasks that at the moment only
people can perform. Google themselves
use it to lower the energy bills of their

21
Artificial Intelligence | Five technology trends that leap-frog Artificial Intelligence

22
 Artificial Intelligence | Five technology trends that leap-frog Artificial Intelligence

Five technology trends that leap-frog

Artificial Intelligence

In our first two articles about AI terminology1 and techniques, we have

discussed the definitions and techniques that make up AI. But developments in
AI do not stand on its own. The broader field of technology development have
a huge interrelationship with the adoption of AI, making it more accessible on
one hand and driving the need to use AI on the other hand

In this article, we will elaborate on a few HavenOnDemand allow you to quickly build
broader technology developments that are and run applications. With cloud ‘Analytics
leap-frogging the adoption of AI. Trends as a Service’ (AaaS) offerings, organizations
or factors of growth that bring AI to more can experiment with AI and begin to build
people and organizations, of which some of intelligent applications without harness
the more important trends are described existing IT infrastructure.
in this article.
2. Big Data
1. Cloud Second amongst the recent accelerators
One of the first trends that enables the for AI is big data, or more specifically:
rapid growth of AI is cloud computing. large, fast, and/or unstructured data.
As explained in our previous article2, Think of all the information in images, text,
AI techniques are based on complex sensor data or other data generated by
mathematical models and require large for example mobile devices. Right now,
amounts of training data (examples) to 80% of all company data is unstructured
learn their intelligent capabilities. Therefore and increases much faster in size than its
building, improving and running AI structured counterpart4. In recent years,
applications requires immense computing technology has become widely available
power. Cloud technology offers that in to capture, store and process that data.
a flexible and scalable environment at Therefore, many companies have invested
relatively low-cost and without huge in building “Data Lake” platforms manage
initial investments.3 their Big Data.

In addition, the IT infrastructure of large The potential for applications of this

corporates is often too big and too rigid unstructured data is huge and yet largely
to experiment with AI applications across untapped. AI techniques make it possible
and within the enterprise platforms. AI to process and analyze unstructured
cloud services such as Amazon AWS AI, data, allowing business to obtain valuable
Microsoft Cortana, IBM Bluemix/Watson, insights from the information and improve
Google Cloud Machine Learning and HPE their decision making. It can potentially

1
Part 1: Artificial Intelligence defined - https://www2.deloitte.com/nl/nl/pages/data-analytics/articles/part-1-artificial-intelligence-defined.html
2
Part 2: Artificial Intelligence Techniques Defined: https://www2.deloitte.com/nl/nl/pages/deloitte-analytics/articles/part-2-artificial-intelligence-techniques-
explained.html
3
The Cloud: technical advantages - https://www2.deloitte.com/au/en/pages/deloitte-private/articles/cloud-computing-deloitte-private.html
https://www.nuance.com/healthcare/physician-and-clinical-speech/dragon-medical.html
4
Dark analytics: illuminating opportunities hidden in unstructured data: https://dupress.deloitte.com/dup-us-en/focus/tech-trends/2017/dark-data-
analyzing-unstructured-data.html
23
Artificial Intelligence | Five technology trends that leap-frog Artificial Intelligence

find patterns and complex relationships by 3. APIs 4. Open Source

shifting through billions of observations. Perhaps the easiest way to start building Although APIs are a great way to start
For example, in the process of assessing intelligent applications is by using building AI applications, they are very
insurance claims, intelligent AI applications Application Programming Interfaces specifically aimed at achieving a certain
can automatically understand natural (APIs). An API is a piece of out-of-the-box task. When you need to perform a machine
language from texts and analyze images functionality that can be called from learning task that isn’t available through
such as photographs. By using these another program or App. If for example, an API, you can build one yourself. This
techniques, AI applications have the your app requires face recognition, you requires knowledge, complex algorithms
potential of detecting fraud earlier, improve can call an API rather than program it and frameworks.
the quality and consistency of the claim yourself. Many of the large technology
assessment and make the process more firms offer APIs in the field of computer Today, increasingly more AI algorithms and
efficient. This is only one out of many vision, speech recognition, and natural frameworks are available as Open Source,
example use cases. language processing (NLP) or other meaning they are publicly available and
cognitive domains on their cloud platforms. often at no license cost. Consequently,
While on the one hand AI is the solution Intelligent APIs are pre-trained and pre- developers of AI applications can rely on
to analyze large amounts of unstructured configured models for a certain task and the knowledge and previous work of a large
data, AI on the other hand needs big data serve as gateways to AI applications. user base. That makes this trend our fourth
in order to become ‘intelligent’. As already This can be illustrated with the Visual accelerator of the AI uptake. An example
mentioned in the previous section and our Recognition6 API from IBM. When this API of open source AI software is TensorFlow
previous article, AI applications need to be is called and receives an image of a car, from Google.
trained with many examples. Think of this it recognizes the car, and perhaps other
as personal development: during your life, objects that are on the provided image. The TensorFlow is an open source machine
you learn from experience (examples given recognized objects are returned to the user learning library with many different
by your environment), which enable you with a certain amount of confidence for algorithms and frameworks7. Open source
to perform certain tasks and get better, each class of objects. training sets are available these days which
leading to more expertise. This can be allow for loading Intelligence directly into
illustrated by an example of Tesla’s Fleet Furthermore, APIs do not need to be used your AI application8.
Learning5, where all Tesla’s continuously as standalone services. They can serve as
share the data with the central intelligence building blocks for combined intelligent A simple comparison between an API and
system, such as map data: where is applications. For example: building a the underlying open source framework
driven when - and at what speed. Once speech translator assistant requires a can be illustrated by the following example.
a significant number of cars report on a speech-to-text API, a translation API and a Google’s Speech API can be used to
changing condition (due to i.e. poor road text-to-speech API to return the translation develop applications that receive audio
conditions, works, etc.) the system is in the other language. This modular from its user and convert it to plain text.
updated allowing others cars to learn to characteristic of APIs makes them very To do this, Google’s API uses techniques
anticipate on the changed condition. useful for a wide range of AI applications. such as deep learning and is trained with
Hence the relationship between AI and big Using cognitive APIs is the easiest and millions of examples. Next to this API,
data is two-fold: big data is a prerequisite quickest way to start building and Google also made the source code for
for AI and AI is the solution to process integrating an AI application. the (deep learning) neural network itself
unstructured data and derive insights publicly available. However, with only the
from it. source code, no audio can be converted

5
PTesla’s Fleet Learning - http://gas2.org/2017/02/14/fleet-learning-capacity-key-teslas-lead-self-driving-race/
6
Visual Recognition by IBM - https://www.ibm.com/watson/developercloud/visual-recognition.html
7
TensorFlow by Google - https://www.tensorflow.org/
8
Open Data for Deep Learning - https://deeplearning4j.org/opendata

24
 Artificial Intelligence | Five technology trends that leap-frog Artificial Intelligence

to text. The model needs to be trained blocks for standardization in the AI field are
with many examples, but can be set very the JSON data format and REST principles
specifically according to the needs of for APIs12.
the user. This can gain the competitive
advantage organizations are looking for. JSON (JavaScript Object Notation)13 is a data
format that is easy to write and understand
As a final note on open source, AI for humans and easy to generate and
applications can also use a combination parse for computers. REST stands for
of APIs and open source technologies Representational State Transfer and is an
to rapidly build up an application from architecture/approach for web services
multiple pre-built modules, which speeds which allows easy evolution of API driven
up the development process and uses services.
the best of both worlds. Altogether, open
source AI technologies broaden the user With more data generated by IoT devices
base and the set of use cases for AI, and and easier development and integration
create competition for large commercial AI due to standardized formats and principles,
vendors so that prices remain low. At the AI solutions are emerging more broadly.
same time, some of the large vendors take
Open Source AI code and package it into Conclusion
a commercial application, adding support, Recent technological trends drive the
maintenance, training etc. broad adoption of AI. With the cloud as
platform and APIs as building blocks for
5. IoT and standardization intelligent applications, AI is available for
The last important accelerator for AI is more people and organizations than ever
Internet of Thing (IoT), meaning all the before. The rise of unstructured data
(mobile) devices, vehicles or sensors opens opportunities for AI technologies
that are connected to the internet. which can give companies a competitive
All these devices together generate a advantage. Also, AI applications are better
massive amount of fast, semi-structured than in the past thanks to better availability
data, which can ‘feed’ and improve AI of training data. Moreover, with open
applications. A few examples of these are: source AI technologies, customized AI
self-driving vehicles, homes with intelligent solutions can be developed quickly using
thermostats9, intelligent pacemakers pre-built modules and using wisdom of the
from patients that provide doctors with crowd. Finally, IoT will drive standardization
real-time insights or a parking garage that further which allows for even more data
recognizes your car10. As more and more becoming available and better integration
devices are connected, standardization of devices.
of data flows, -formats and -services is
needed, so that these devices can interact With the recent technological trends
properly. Standardization in IoT is still in described in this article, we believe that the
development11. Two important building right time has come for organizations to
start developing their first AI use cases.

9
Combining AI with IoT - https://www.techemergence.com/artificial-intelligence-plus-the-internet-of-things-iot-3-examples-worth-learning-from/
10
The Edge: the smartest building in the world - https://www.bloomberg.com/features/2015-the-edge-the-worlds-greenest-building/
11
2016 : A year of advancing Internet of Things (IoT) applications: https://innovation-in-manufacturing.deloitte.com/2016/12/28/2016-a-year-of-advancing-internet-of-
things-iot-applications/
12
Open Connectivity Foundation (OCF): https://openconnectivity.org/
13
JSON - http://www.json.org/

25
Artificial Intelligence | AI opportunities for the future

26
 Artificial Intelligence | AI opportunities for the future

AI opportunities for the future

As mentioned in our previous blogs, Artificial Intelligence (AI) is said to be

one of the most disruptive technologies impacting several industries and
businesses. Researchers and practitioners compare the emergence of AI
with the industrial revolution of the last century1. AI comprises of a set of
technologies that will change business as we know it. More and more tasks
will be automated, but the largest impact will follow from brand-new business
models and the rise of intelligent services that did not exist before.

To conclude our series of blogs on AI, this edition will highlight 5 exemplary business
opportunities enabled by use of AI. We provide these examples to have a discussion
around questions such as: what does incorporating AI in your business can offer? What
value can be delivered by AI? And how to measure the value of AI? We know that impact of
AI can be expressed in different ways, e.g. as depicted in table below:

Value Example

Cost reduction High level of intelligent automation in call centers for insurers

Efficiency in healthcare by reducing contact moments

Optimal service between patients and general practitioners (implementing
efficiency smart chatbots to perform triage and answer most common
questions)

New flows of Cross-sell and up-sell of products in e-commerce via intelligent

revenue recommendation systems and via high-level of personalization

Customer New services such as grocery shops providing health advise to

satisfaction customers

We will now go through some examples across various industries and examine the nature
of AI implemented in these cases.

https://medium.com/mmc-writes/the-fourth-industrial-revolution-a-primer-on-artificial-intelligence-ai-ff5e7fffcae1
1

27
Artificial Intelligence | AI opportunities for the future

Application 1: AI in disease diagnosis of GAN is that the generator generates There are also other examples. Fore
and illness treatment images of which the discriminator instance AI driven chatbots equipped with
Cost reduction by AI seems to be an believes them to be real (using an iterative NLP and image recognition capabilities can
important driver to counter the ever approach). The generated images are so interact with customers in an intuitive way
increasing cost of healthcare. However the alike to the training set, they would fit right and advise them on products. Imagine you
opportunity to improve the effectiveness in. Using this technique a GAN can act as taking a picture of your shirt and sharing
of healthcare by AI driven diagnostics and a true fashion designer by creating brand it with the chatbot who then comes up
treatment plans is much bigger. new items, in the same style as the style with advise on newest trendy sneakers
of the items on which it is trained. The matching your outfit.
Opportunities range from drug design Amazon team Lab126 has already started
to patient diagnosis and to create to test the possibilities in this domain4. For
personalized treatment plans (almost on this specific example, in terms of value,
DNA level). greater customer satisfaction and even
competitive advantage can be achieved.
Just as an example, Infervision uses
Deep Learning techniques on patient
data derived from X-Ray, CT, MRI, text
descriptions of symptoms and diagnostic
reports, to construct automatic diagnostic
recommendations2. Multiple top-notch
hospitals in China already work with this
recommendation system, identifying (lung)
cancer in an early stage.
Another example is the use of Deep
Learning to classify skin cancer.
Researchers already developed a system
that can classify skin cancer with a
precision comparable to dermatologists3.
Transferring this technology to mobile
devices can make dermatologic care
available for more people than ever before.

Application 2: AI in fashion design and

customer interaction
Not the first topic that comes to mind
regarding AI, but the technology is
expected to influence the fashion industry
as well. A long term, but disruptive,
opportunity is made possible by GANs:
Generative Adversarial Networks. The
algorithm uses two neural networks: a
generator and a discriminator. The logic

2
https://www.forbes.com/sites/jenniferhicks/2017/05/16/see-how-artificial-intelligence-can-improve-medical-diagnosis-and-
healthcare/#68a5ef462239
3
http://www.nature.com/nature/journal/v542/n7639/full/nature21056.html?foxtrotcallback=true
4
https://www.technologyreview.com/s/608668/amazon-has-developed-an-ai-fashion-designer/?utm_content=buffer4008a&utm_
medium=social&utm_source=twitter.com&utm_campaign=buffer

28
 Artificial Intelligence | AI opportunities for the future

Application 3: AI in cyber crime and

fraud detection
The impact of cybercrime on global
business is enormous. Next to intangible
damage for a company (e.g. brand damage),
costs can be huge. It is estimated by Forbes
that the global costs for cybercrime will be
6 trillion dollars by 2021. A large proportion
of this cost is in credit card fraud and Figure 1: Targeted Advertising Demo
although fraud detection techniques exists
for years, they are still not sufficient enough on photos taken at the entrance of the Voice ordering
to prevent this from happening. New AI festival5. Not only facial recognition can England largest online grocery shop Ocado
techniques, like Recurrent Neural Networks help in tracking down criminals, emotion is a true pioneer when it comes to AI-
that were previously not available can analysis can deliver extended value technology in a customer’s daily shopping
be the solution to detect fraud in earlier regarding this opportunity. Being already routine. With a cutting edge robotics
stages8. Equipped by trained RNN’s, fraud in use, NTchLab created software that system and an autonomous delivery
detection systems can scan thousands identifies if someone is feeling stressed or system in store, they recently released a
of transactions instantly and predict/ angry. Put into use in e.g. a grocery store, new Amazon feature: voice control7. When
classify them into buckets (e.g. ranging detecting a stressed person will alert in possession of an Amazon Echo (Alexa)
from high to low probability of fraud). This security to pay extra attention6. one can simply call out new items and
type of system could help save time by Alexa will add them to an existing shopping
focusing the cases where probability is Application 4: AI in personalized list. Although this feature is already on the
high for fraud. A word of caution in this advertising and support market, its added value will be something
case is also applicable. Utilizing these type Imagine walking into a store and your for the future. Currently, a shopping list
of algorithms requires a solid governance favorite brand or product is at discount has to be initiated manually and Echo only
system to monitor the output and limit the same day. This can be a coincidence listens to full and scripted sentences. So a
potential bias as much as possible. It’s and your lucky day, but it’s an advantage phrase such as: ‘He Echo, we’re out of milk!’
very well known that due to pre trained for the store owner as you’re probably is still a long way off.
character of the algorithms, bias can be more willing to buy the product. All of this
serious flaw in the system. is possible with AI techniques like face-
recognition. A personalized advertising
Criminal identification application as described above is trained
With the recent launch of the new iPhone with many examples, knowledge of people
X, Apple brings face recognition to many and their shopping behavior. It results in
of us worldwide. In the next years iPhone an intelligent application that can identify
users will be able to unlock their iPhone the type of customers, their emotions
by looking into the front camera. But and possible shopping preferences.
authentication of personal content is not Personalized advertising is a proven
the only application for face recognition. method to generate more revenue. Since
Governments and security services use these techniques are already available
face recognition to identify citizens and today, this opportunity is a high value/short
track down criminals. Recently the police term win.
in China captured twenty-five wanted
criminals at a Chinese beer festival based

5
http://www.dailymail.co.uk/sciencetech/article-4851564/Facial-recognition-detects-criminals-beer-festival.html?ITO=1490&ns_
mchannel=rss&ns_campaign=1490
6
http://www.dailymail.co.uk/sciencetech/article-4494994/New-technology-reads-emotions-potential-terrorists.html
7
https://www.engadget.com/2017/06/29/cargopod-ocado-oxbotica-driverless-delivery-van/

29
Artificial Intelligence | AI opportunities for the future

Application 5: AI in logistics and on- Conclusion

time delivery We have discussed some of the
With immense growth of e-commerce applications of AI and can conclude that
and on-line shopping, one of the biggest that there is a major opportunity for the
challenges for ecommerce and many business to improve, gain efficiency and
logistics companies is to reduce the define new business models by utilizing the
costs of delivery in the “last mile” while power of AI.
maintaining quality and service. Some
of the most innovative companies are This was the 5th and last edition of our
going through a paradigm shift using series on AI. We started the series on AI
AI. Some great examples are out there by providing an overview of terms and
where driverless cars are being utilized definitions of AI9. In the 2nd blog we have
for autonomous delivery and real-world discussed some of the fundamental
tests are right now on-going (7). There are techniques which are at the heart of AI10.
also more practical areas where AI plays In the 3rd article we went on and gave
an essential role in the logistics. Think five applications of AI: Image recognition,
about on-time delivery, planning and Speech recognition, automatic translation,
time-estimation used by logistic companies Q&A and Games. Then we continued and
through their track-and-trace systems. analyzed five technology trends that leap-
A good example of how AI is utilized frog AI. This was published in our 4th blog.
is an algorithm developed by Deloitte And then finally, in this 5th release we have
Netherlands to predict timeslot of delivery. provided some practical applications of AI
This model is able to make predictions for in day-to-day life.
future deliveries based on delivery
route patterns. These patterns are found AI is and will remain a dynamic and broad
using AI techniques. Proven cost reduction filed encompassing several areas of
and significant improved customer expertise, ranging from computer science
satisfaction are the big wins here. to mathematics, neuroscience all the way
up to philosophy and even biology. We
hope that the topics discussed in our 5x5
on AI edition will help the reader to get
an end-to-end overview of AI in a more
practical sense rather than a bullet proof
complete scientific exercise.

8
https://www.forbes.com/sites/theyec/2017/07/13/the-true-cost-of-cybercrime-for-businesses/#61c4114a4947
9
https://www2.deloitte.com/nl/nl/pages/deloitte-analytics/articles/part-1-artificial-intelligence-defined.html
10
https://www2.deloitte.com/nl/nl/pages/deloitte-analytics/articles/part-2-artificial-intelligence-techniques-explained.html
11
https://www2.deloitte.com/nl/nl/pages/deloitte-analytics/articles/part-3-applications-of-artificial-intelligence.html
12
https://www2.deloitte.com/nl/nl/pages/deloitte-analytics/articles/part-4-five-technology-trends-that-leap-frog-artificial-intelligence.
html?id=nl:2sm:3li:eng_da_corp:AI4

30
 Artificial Intelligence | Authors

Authors

Stefan van Duin

Director
Tel: + 31 88 288 4754
Mobile: + 31 61 234 4457
Email: [email protected]

Stefan van Duin is Director at Deloitte Consulting and leading the Data Analytics team, a team of
over 100 professionals specialized in Big Data, Data Visualisation, Machine Learning and Artificial
Intelligence. Stefan is convinced that more and more companies consider data analytics as a core
competency, to develop better products, services and business models, or to personalize the client
experience. Stefan has made it his mission to help companies on their journey.

Naser Bakhshi
Senior Manager
Tel: + 31 88 288 3874
Mobile: +31 61 201 1665
Email: [email protected]

Naser Bakhshi is Senior Manager Advanced Analytics in the Deloitte Dutch Analytics service line.
His primary focus is on Artifical Intelligence, Machine Learning and Advanced Analytics where he
focuses on AI driven solution for monitoring, predictions and process optimization. Moreover, Naser
is an experienced transition manager with experience in defining advanced analytics strategy (incl.
business case definition), analytics roadmap, implementation of Analytics Centre of Excellence (incl.
defining operating models for ACE, end-to-end processes and RACI, drafting Analytics architecture
and data acquisition and quality check processes. Combining hard analytical skills with a strong feeling
for organizational change and sensitivity make Naser suitable for supporting clients in becoming
insight driven.

Co-authors
Chapter 2: Titus Sloet tot Everlo & Hicham el Bouazzaoui.
Chapter 3 : Carmen Wolvius en Sjors Broersen
Chapter 4 : Johan van der Veen en Thomas Heeneman
Chapter 5: Joanne Lijbers, Carmen Wolvius en Hicham El Bouazzaoui.

31
Artificial Intelligence | Sources

Sources

Chapter 1:
1. https://www.gartner.com/document/3380751/
2. http://www.bbc.com/news/technology-37713629
https://futureoflife.org/ai-open-letter/
3. Preparing for the Future of Artificial Intelligence, NSTC, 2016
4. 6. Raphael, B. 1976. The thinking computer. San Francisco, CA: W.H. Freeman
5. http://www-formal.stanford.edu/jmc/whatisai/node1.html
6. Stephen Lucci, 2016, Artificial intelligence in the 21st century : a living introduction
7. Deloitte, Cognitive analytics™ The three-minute guide
8. http://www.bostondynamics.com/
9. https://biomimetics.mit.edu/research/dynamic-locomotion-mit-cheetah-2
10. http://www.ipsoft.com/amelia/

Chapter 2:
1. https://www.merriam-webster.com/dictionary/heuristic
2. https://en.wikipedia.org/wiki/Ant_colony_optimization_algorithms
3. https://en.wikipedia.org/wiki/Genetic_algorithm
4. https://en.wikipedia.org/wiki/Support_vector_machine
5. http://neuralnetworksanddeeplearning.com/chap1.html
6. https://www.cs.rice.edu/~vardi/dag01/givan1.pdf
7. http://artint.info/html/ArtInt_224.html
8. More on RNN: http://karpathy.github.io/2015/05/21/rnn-effectiveness/

Chapter 3:
1. www.deloittesmartreports.com
2. https://www.tensorflow.org/
3. https://google-shopper.en.softonic.com/android
4. https://www.techrepublic.com/article/ibm-watson-the-inside-story-of-how-the-jeopardy-winning-supercomputer-was-born-and-what-
it-wants-to-do-next/
5. https://www.ibm.com/blogs/research/2016/11/identifying-skin-cancer-computer-vision/
6. https://arxiv.org/abs/1610.04662
7. https://www.scnsoft.com/case-studies/facial-recognition-for-retail
8. https://www.facefirst.com/industry/retail-face-recognition/
9. https://www.pcworld.com/article/243060/speech_recognition_through_the_decades_how_we_ended_up_with_siri.html
10. https://www.slideshare.net/kleinerperkins/2016-internet-trends-report/119-KPCB_INTERNET_TRENDS_2016_PAGE119Voice
11. http://www.healthcareitnews.com/news/speech-recognition-proving-its-worth
12. https://www.nuance.com/healthcare/physician-and-clinical-speech/dragon-medical.html
13. https://mmodal.com/speech-solutions/
14. https://techcrunch.com/2016/11/22/googles-ai-translation-tool-seems-to-have-invented-its-own-secret-internal-language/
15. http://www.cleverbot.com/
16. https://www.ipsoft.com/amelia
17. https://www.ibm.com/blogs/watson/2016/01/ross-and-watson-tackle-the-law/
18. https://en.wikipedia.org/wiki/AlphaGo_versus_Lee_Sedol
19. https://deepmind.com/blog/innovations-alphago/
20. https://arxiv.org/abs/1712.01815
21. http://www.telegraph.co.uk/technology/2017/03/15/googles-deepmind-ai-learns-like-human-overcome-catastrophic/
22. https://www.bloomberg.com/news/articles/2016-07-19/google-cuts-its-giant-electricity-bill-with-deepmind-powered-ai
23. http://www.wired.co.uk/article/deepmind-nhs-streams-deal

32
 Artificial Intelligence | Sources

Chapter 4:
1. https://dupress.deloitte.com/dup-us-en/focus/tech-trends/2017/dark-data-analyzing-unstructured-data.html
2. http://gas2.org/2017/02/14/fleet-learning-capacity-key-teslas-lead-self-driving-race/
3. https://www.ibm.com/watson/services/visual-recognition/
4. http://www.healthcareitnews.com/news/speech-recognition-proving-its-worth
5. https://www.nuance.com/healthcare/physician-and-clinical-speech/dragon-medical.html
6. https://mmodal.com/speech-solutions/
7. https://www.tensorflow.org/
8. https://www.techemergence.com/artificial-intelligence-plus-the-internet-of-things-iot-3-examples-worth-learning-from/
9. https://www.bloomberg.com/features/2015-the-edge-the-worlds-greenest-building/
10. https://innovation-in-manufacturing.deloitte.com/2016/12/28/2016-a-year-of-advancing-internet-of-things-iot-applications/
11. https://openconnectivity.org/
12. http://www.json.org/

Chapter 5:
1. https://www.forbes.com/sites/jenniferhicks/2017/05/16/see-how-artificial-intelligence-can-improve-medical-diagnosis-and-
healthcare/#68a5ef462239
2. http://www.nature.com/nature/journal/v542/n7639/full/nature21056.html?foxtrotcallback=true
3. https://www.technologyreview.com/s/608668/amazon-has-developed-an-ai-fashion-designer/?utm_content=buffer4008a&utm_
medium=social&utm_source=twitter.com&utm_campaign=buffer
4. http://www.dailymail.co.uk/sciencetech/article-4851564/Facial-recognition-detects-criminals-beer-festival.html?ITO=1490&ns_
mchannel=rss&ns_campaign=1490
5. http://www.dailymail.co.uk/sciencetech/article-4494994/New-technology-reads-emotions-potential-terrorists.html
6. https://www.engadget.com/2017/06/29/cargopod-ocado-oxbotica-driverless-delivery-van/

33
Deloitte refers to one or more of Deloitte Touche Tohmatsu Limited (“DTTL”), its
global network of member firms, and their related entities. DTTL (also referred
to as “Deloitte Global”) and each of its member firms are legally separate and
independent entities. DTTL does not provide services to clients. Please see
www.deloitte.nl/about to learn more.

Deloitte is a leading global provider of audit and assurance, consulting, financial

advisory, risk advisory, tax and related services. Our network of member firms in
more than 150 countries serves four out of five Fortune Global 500® companies.
Learn how Deloitte’s approximately 264,000 people make an impact that matters
at www.deloitte.nl.

This communication contains general information only, and none of Deloitte

Touche Tohmatsu Limited, its member firms, or their related entities (collectively,
the “Deloitte network”) is, by means of this communication, rendering professional
advice or services. Before making any decision or taking any action that may
affect your finances or your business, you should consult a qualified professional
adviser. No entity in the Deloitte network shall be responsible for any loss
whatsoever sustained by any person who relies on this communication.

© 2018 Deloitte The Netherlands