Indian Institute of Management Kozhikode

Final Examination-2019-20 (ePGP-11/EPGCITMA-07)

Name: ____________________________________________________

Roll No: ___________________________________________________ Date: __________________

General Instructions:
a. TYPE YOUR NAME AND ROLL NUMBER IN THE SPACE PROVIDED ABOVE.
b. OPEN-BOOK Examination.
c. Answer all questions.
d. Figures on the right hand margin indicate marks and word limits(W).
e. This question-answer sheet consists of seven subjective type questions and one Case Analysis.
f. Total number of pages after the answer typed should not exceed 16 pages, otherwise the extra sheets
(page number 17 onwards if any) will not be evaluated.
g. No Negative marking.
h. The answers should be typed in MS Word in Times New Roman font: size 11 with at least 18 pt. line
spacing.
i. All your answers must adhere to the maximum word limit and the space provided for answers. Wherever
required draw/add figures, tables etc. to increase the clarity in your answers.
j. You need to upload the answer sheet within the stipulated time as announced by the ePGP personnel.
k. Your answer-sheets will be processed through our Institutes antiplagiarism software.
l. Any copying instance among yourselves or from any published sources (internet, books etc.) may attract
punishments up to the extent of subject F-Grade.

1 of 8
Q. No. 1. Games like checkers (draughts) and chess provide the ideal working environment for an
AI program. The kind of competence required to play these games is extremely specialized. The
micro-worlds that games present are ones of strict rules, uncomplicated environments and
predictable consequences. AI thrives on these properties, and as a result, game-playing machines
are very successful.
This victory of machine over man is widely cited, and for good reason. It demonstrates an important
lesson: the capabilities of an artificial agent are not necessarily restricted by the capabilities of the
designer. Samuel’s program plays better checkers than he does.
Most game-playing machines work by constructing a symbolic representation called a game tree.
From the starting position, the game tree details all the possible ways the game can unfold. The
representation is symbolic: it might use a symbol to represent a white piece, and another symbol to
represent a black piece.
Unlike a human, a computer can easily generate the game tree and keep it in memory. Using this
internal representation, the computer can then look ahead to see the precise consequences of its
actions.
In a chess, the space of all possible games is enormously high. IBM’s tailor-made chess computer
Deep Blue defeated Garry Kasparov, the most highly ranked player in the world. This was a
landmark for AI.
Once Garry Kasparov, perfectly expressed the difficulty in chess, “The number of potential chess
moves exceeds the number of atoms in the universe. It’s a number beyond any possible
calculations”. Thus, looking ahead even a moderate number of moves becomes intractable – the
number of combinations becomes too large to contemplate. The game tree for chess cannot fit into
the universe, let alone a computer’s memory.
a. What kind of natural difficulty seems to be in tree representation that even constraints always
victory notion of fastest chess playing computers every time it plays with the human chess player
(i.e., not every time chess playing computer becomes winner)? (1.5)
(W:100)
b. How the problems in (a) above are addressed in a chess playing computer, so that you can
justify the statements in (a)? (2)
(W:100)
c. Justify the statement, “Touting Deep Blue as a success amounts to Al putting its hands up and
admitting lack of progress in replicating anything even approaching human cognition”. (2)(W:100)
Answer:

2 of 16
Q. No. 2. If we look at the biological systems that support cognition, we see brains of varying sizes
built from collections of neurons. The human brain has approximately 100 billion neurons, and on
average, each one of those neurons is connected to around 10,000 other neurons by cable-like
structures called axons.
As discussed in the class, collections of neurons can act as computing devices, and the work of
McCulloch and Pitts tells us that these configurations of neurons can compute the same class of
calculations as a Turing machine.
Connectionist models usually take the form of artificial neural networks, referred to as neural
networks. Neural networks are groups of artificial neurons configured to perform some calculation.
Neural networks are becoming increasingly well known.
(a) Draw the nodes and node connections for a fully connected feed-forward network that
3 of 16
 accepts three input values, and has one hidden layer of five nodes and an output layer
containing four nodes.
(1) (Figure)
(b) What’s the sole function of the nodes in the input layer? (1)(W:50)
(c) Suggest how this network can be extended to be called as a true deep learning neural
network. (2)(W:100)
Answer:

4 of 16
Q. No. 3. Below three predictive scenarios are given. You need to assess whether AI as practiced
today is applicable to these scenarios and explain it. (1.5*3 = 4.5) (W:100 each)
(a) Scenario 1: Predicting short-term retail and economic activity based on satellite images to
track the movement and number of ocean-going container ships.
(b) Scenario 2: Predicting new prices of a portfolio during Covid-19 pandemic which was on an
unexpected event that completely disrupted the foundations of the global economic system.
(c) Scenario 3: Predicting the effect of genomic research on costs of healthcare spending after
25 years.
Answers:

5 of 16
Q. No. 4. Explain the term complexity as used in an NLP application. Indicate how the variation of
the values in each of the axes/dimensions can quantify the complexity of a situation or problem. In
the present juncture of time companies analyse the user reviews on Amazon, and various social
media sites to understand the sentiments of the users. The Sentiment analysis methods analyse all
text messages, and assess the user opinions on the subject to be positive, negative, or neutral.
On the other hand, to make the customers more beneficial, normally the reviews are tagged by its
topic that can allow customers to quickly find the information they want. This is known as the text or
document classification.
Now analyse the complexity of document classification vis-à-vis sentiment analysis. (4)(W:100)
Answer:

6 of 16
Q. No. 5. Transfer learning in essence works by transferring the knowledge for solving one task to
another task. This increases the training speed of the new task in the presence of relatively very
less data. This essentially describes the reusability of already trained models. Now answer the
following:
(a) Can the weights learned during the training stage of a model be reused to model architecture of
another network? Explain, why or why not? (2)(W:50)
(b) Can the “learned” weights be used on any different task? (2)(W:50)
(c) State, in which of the following cases transfer learning may or may not produce desirable
results with proper reasoning: (3*1=3)(W:50
each)
i. Pre-trained model for types and varieties of fruits to be used on a new dataset for types
and varieties of vegetables.
ii. The pre-trained model in (i) is employed on a new dataset consisting of makes and
models of trucks and vans.
iii. A model trained on a massively diverse set of classes of images of the order of tens of
thousands of image classes. How can you use this on any arbitrary new dataset and what
are its limitations?
Answer:

7 of 16
Q. No. 6. RPA is a class of software “robots” that mimics exactly how humans operate (by logging
into a system, entering data, executing workflows, etc.). They work with business applications, such
as ERP and CRM systems and many other applications. Simply speaking, since these software
robots replicate human activities, to a large extent the business applications work exactly as they
always have before, without significant human involvement. The term “robot” signifies the software-
driven capability to replace or enhance a human task. RPA primarily attempts to memorize the
activities humans do at the “presentation layer” (the user interface) of business applications.
The RPA alone can deliver significant benefits; however, to make automation sustainable and to do
real enterprise-wide transformation, AI capabilities need to brought into the mix. Product vendors
have latched onto the buzzwords and have been marketing cognitive capabilities in their RPA tools
in a big way. (1.5*4 = 6) (W:100 each)
(a) What characteristics of RPA is it’s greatest strength as well as the greatest weakness?
(b) Give an example of this situation that satisfies the point in (a).
(c) Suggest a solution how you can incorporate that into the RPA so as to address the issues in (a).
(d) RPA is a bot and so also many chatbots extensively used across industries. Explain in what
aspects RPA is different from chatbots.
Answer:

8 of 16
Q. No. 7. Artificial Intelligence (AI) is slowly becoming part of our daily life. AI based
computational photography becoming a norm in smart phones with multiple cameras. Its now is the
new buzzword in photography. Simon Fitzpatrick, Senior Director, Product Management
at FotoNation says “It's about taking studio effects that you achieve with Lightroom and Photoshop
and making them accessible to people at the click of a button.” 
Recently, an AI camera captured the attention of media, when Inverness’ Caledonian Thistle FC
(ICTFC) incorporated an AI camera to cover their football games. This game was conducted during
the Covid-19 pandemic, so the Inverness club had decided not to employ human camera operators
and so opted for this advanced automated camera system to follow the action. This camera is
programmed to follow the moving ball and can provide a dashing real-time experience to all fans.
Before the match ICTFC club announced "ICTFC will utilize live footage supplied by the newly
installed Pixellot camera system installed at Caledonian Stadium. The Pixellot system uses
cameras with in-built, AI, ball-tracking technology to produce live HD footage of all home SPFL
Championship matches at Caledonian Stadium, which will then be broadcast directly to ICTFC
Season Ticket holders and those purchasing a PPV match via a centrally operated streaming
platform at StreamAMG,".
However, during the match between Inverness Caledonian Thistle and Ayr United hosted in
Inverness in the distant Scottish Highlands at the Caledonian Stadium, the viewers started
commenting that they keep on missing their team's goals due to the slapdash performance of the AI
based robot cameras. The robot camera continuously tracked a bald head as it could not
differentiate between the ball and the bald head of the linesman, with the referee’s head glistening
9 of 16
in the glorious Scottish sun in the late afternoon. What exactly happened that when centre-forward
Nikola Todorov scored the team’s first goal, the camera kept on mistaking the ball for the bald head
on the side lines, denying viewers of the real action while focusing on the linesman instead. Until
then the club authority never dreamt of the situation of such hilarity.
Although the match ended a 1-1 draw, more than the team’s effort, it was the gaffe that stole all the
attention online. As many started to tag the camera company to complain and demand answers, a
spokesperson in a tweet said: “We are aware of the issue with linesman and we are already
working on improving this for the next game. We hope that this will not happen again and that your
viewing experience will be better. Thank you.” (W:100 each and diagrams extra)
(a) Explain, what way AI is used to enhance camera photography? What AI technologies used to
identify the content of an image? (1.5)
(b) Traditionally expert photographers use to have mastery over DSL camera to create various
types of images. However, nowadays AI has use to facilitates all the photography effects and
also empowers non-professional photographers the capabilities of more advanced
photography. During last two decades in the digital photography domain Adobe with its
Photoshop and Lightroom features enjoys a good market share. With the advent of AI camera
is it going to drive Adobe out of the market or at least Adobe’s Photoshop and Lightroom will
soon be out of fashion? Give your views on this.
(2)
(c) What could you suggest to avoid this type of confusions by today’s AI Technology? (1.5)
(d) This clearly shows the failure how technology can fail in an embarrassing way. Can you think
off a solution how deep learning can be used to build a true digital camera. (2)
(e) How can you use a Q-learning method to provide an alternative solution to this ball tracking
problem? (2)
Answers:

10 of 16
11 of 16
Q. No: 8. Case Study: Machine Translation of Natural Languages
In 1999, Marco Trombetti and Isabelle Andrieu decided to start a business together. Trombetti was
studying physics, and Andrieu was a linguist. To use their skills, the young couple came up with the
idea of building an internet-based translation company. They called it Translated, and launched it with
an initial investment of $100 for the internet domain Translated.net. In the following 20 years,
Translated would become the largest online professional translation company in the world, growing
organically without any external investment.
Translated focuses on professional translations, like medical journals and engineering user manuals. Such
translations are traditionally performed by professional translators contracted by companies that act as
brokers. Matching a body of text with a translator is not trivial: many documents need both deep language
skills and subject matter expertise. For instance, translating a paper about transcatheter aortic valve
implantations from English to German would need a cardiologist skilled in both languages. Translation
brokers that own a database of translators can take advantage of this and charge large premiums to
provide clients with the right translators. The larger the translation company, the larger the challenges and
inefficiencies in matching translators with jobs. This was an opportunity that Trombetti and
Andrieu set out to exploit.
The first challenge that Trombetti and Andrieu had to face was building a large database of
translators. From the outset, the duo invested in search engine optimization (SEO) so they could land at the
top of Google searches for professional translation services. These being the early days of the internet,
their investment quickly paid off, and Translated started receiving hundreds of resumes per month.
The early success led to a new challenge: many of the translators sending their resumes didn’t meet the
quality standards of the company, and sorting professional translators from non-professional ones required
a great deal of human effort. Translated risked falling into the same inefficiencies as other translation
brokers.
The inspiration for how to address this challenge came to Trombetti in 2002, when he read
the very same Paul Graham essay “A Plan for Spam”. Trombetti thought that the same method
used to discriminate between spam and not-spam email could be used to make a first assessment on
whether a resume they received was from a professional or nonprofessional translator, as depicted in
figure 1.

12 of 16
 Figure 1 Translated’s ML algorithm helped the company automatically identify professional and
nonprofessional translators.
The algorithm was first introduced in 2002, when the company started receiving hundreds of translation
requests per day, and it has been improved ever since. The next evolution of the algorithm was
to turn it into an automatic translator-document matching AI, training it using the matchings made by
Translated’s project managers (PMs) over the years. Once deployed, at any new translation request,
the algorithm would look at the document and at the company’s pool of translators and calculate a
match probability for each translator. Translated’s PMs would then validate the matches and assign the task.
This process allowed the young startup to handle many clients and translators without compromising
delivery speed (a factor often crucial for clients) and to keep their overhead low by having a smaller
team of PMs.
Today, Translated acts as a double-sided platform: from one side, translators apply to offer their
services; and from the other, customers submit the documents they need to have translated. In the
backend, Translated matches the right translator for each job and delivers the final result to the
client, without them needing to interact with each other. The current version of their algorithm is
branded as T-Rank.
The algorithm takes into account up to 30 factors from the source document, including the
translator’s resume (area of expertise) and previous projects (quality and timeliness of previous jobs).
Using this data, T-Rank can rank the translators in the company’s portfolio according to their fit
with the work and provide the best professional for each task, as in figure 2. As in the first
versions of the algorithm, the final recommendations are vetted by expert PMs. In case the PM rejects
the recommendation, the feedback is recorded and used to fine-tune the algorithm to continuously
improve its performance.
Success in the market brought Translated both cash and data. A consistent cash flow enabled
Trombetti and Andrieu to run the company without seeking external investments and to keep complete
ownership. On the other hand, Translated started collecting the translations made by professionals
through its computer-aided translation system called Matecat, improving its matching algorithms and pushing
the company toward a new mission: becoming the leader in machine-aided translation.
To evaluate the performance of its ML-generated translation, Translated has always been tracking the
percentage of suggested words that are changed by the professional translator. In the early days,
translators had to change 45% of suggested words.

13 of 16
 Figure 2 Translated is a double-sided platform: customers send documents, translators send their
data,and the company finds the best translator in its portfolio for each document to translate.
Together with AI-assisted matching of translators, this was enough to deliver translations quicker than the
competition. In 2018, the company could count on more than 1.2 million professional translations, and
percentage of words changed by the professional translator dropped to 24% on the eight most-
spoken languages, 30% lower than other translation engines.
To further improve the experience and productivity of translators, the company has invested in a
technology branded as ModernMT, the first adaptive neural machine translation. Traditionally, ML
algorithms are trained on a large amount of data (training phase) and are static while they’re
used (inference phase). If a translator is presented with a poor suggestion and fixes it, this fix won’t be
taken into account by the machine translation (MT) algorithm until training is repeated with the new
data. This means that while a translator is working on a document, the same fixes might have to
be applied multiple times. The result is a frustrating and inefficient experience for the translator.
As illustrated in figure 3, ModernMT learns and adjusts in real time while the translator is working
through the translation. If the same phrase occurs more than once in a document and the translator
fixes it once ModernMT will learn the correction the first time and avoid repeating the mistake.

Figure 3 Translated’s ModernMT translation engine proposes a translation that can eventually
be corrected by the translator. The translator’s correction is then learned by ModernMT.
As of 2019, Translated receives 2,000 resumes a month from translators. The company has brokered
1.2 million professional translations for more than 130,000 customers, and the translations are performed in
151 languages by more than 180,000 professional translators. The algorithms built on this dataset allow
state-of-the-art machine translation performances on the 27 most common languages. Translated’s
14 of 16
algorithms facilitate, but don’t replace, the work of professionals. They provide a first automatic
translation of the sentences that just need to be refined an d fine-tuned: today, translators need to
change just 24% of the words suggested by the company’s platform. This is 30% better than any
other solution currently on the market. Translated reported revenues of €21 million in 2018, with 51
employees: 12 developers, 35 in operations, with an expected growth of 40% for 2019.
Translated is still investing in improving its translation engine. The goal is not to get AI to produce
perfect translations. In fact, if a translation performed by a professional translator is reviewed by
another professional, that person would still change about 11% of the words. This is because of the
extreme expressiveness of language: the same concept can be expressed with different words to
express slightly different meanings and nuances. According to Trombetti, this is the most
important and fulfilling task required for a translator: making sure that the richness of the
document is expressed in the translation. His goal is therefore improving the company’s AI algorithm to
allow translators to focus on the 11% of the words they need to change to convey the original
message of the document in the most expressive way possible.
Trombetti and Andrieu want to allow translators to perform a more fulfilling and higher-quality job, and
democratize translation along the way. Here’s what they said during our interview with them:
Our technology is creating a lot of opportunities not just for us but also for the translators, who are
finally not working on correcting again and again the same tedious stuff but are spending their time in
being more human, communicating what the text actually means and being more persuasive and
creative.
Analyze this case in the line of the questions given below: (5*2 = 10)
(a) Analyse the risks during the first application and how the company implemented successfully its
first application in NLP. What characteristics of this use case made it a good fit for AI? (W:100)
(b) Analyze the complexity of the machine translation task in Translated Co. using the concept of
depth vs. width. How this complexity is handled in this case? (W:100)
(c) What was the first neural network application the Co. developed and what was the key element
of this development? (W:100)
(d) How do you see AI’s impact on job loss across industries in the same line of Translated’s
adoption of AI? (W:100)
(e) What is your take on the major reason for Translated’s success in implementing the AI
strategy? (W:200)
Answer:

15 of 16
16 of 16
--------GOOD LUCK--------

17 of 16