Personal Statement

Michał Dereziński NSF application

Growing up in an academic household with my father being a professor of physics, I have always appreciated
the beauty in logic and science. My interest in computer science began with the famous question Can
machines think? proposed by Alan Turing. Although there is still no clear answer or even a formal description
of this problem, there is a significant progress in the way computers can understand and analyze abstract
concepts used by humans. This inspired me to study modern natural language processing and knowledge
representation techniques. When I started my research on structuring information into concept hierarchies,
I decided to focus my graduate work on artificial intelligence.

My research on artificial intelligence began when I started attending a seminar on natural language processing
at the University of Warsaw. While studying methods for unsupervised extraction of knowledge from text, I
learned about applications of Formal Concept Analysis to fuzzy data sets. These methods were promising but
inefficient for my task, so I developed a new algorithm for unsupervised concept hierarchy generation, which
eliminated some of their problems. This became my master’s thesis. After defending my thesis I continued
to work in this area. I devised a semi-supervised technique for improving concept hierarchy through active
learning. I presented this work in a talk entitled Active Semi-Supervised Concept Hierarchy Refinement at the
First International Workshop on Learning with Weak Supervision in Singapore in November 2012. I intend
to publish these results in a near future. My other interests in machine learning include the recent advances
in deep neural networks and probabilistic approaches to classification. In September 2012 I attended the
European Conference on Machine Learning in Bristol. This allowed me to make myself familiar with the
current state of research and the scientific community of machine learning.

I studied mathematics in parallel with computer science. In my opinion, it enhances my computer science
skills. It is worth noting that mathematics plays a significant role in machine learning. However, I am
interested in math also for its own sake. I enrolled in various advanced courses covering all major areas of
mathematics, with very good results. My master’s thesis research was in the area of harmonic analysis. The
problem that I was given withstood traditional mathematical approaches, so I had to think outside of the
box to find the solution. My first attempt, was a computer algorithm which involved analyzing millions of
possible cases. Even though this first approach did not turn out to be fruitful, I did not get discouraged and
decided to go about the problem in a completely different way. I think it is important to take risks and use
unusual methods to pave the way for scientific advances. This allowed me to make significant progress in
my research in both mathematics and computer science.

My first work experience was in Santa Clara, California, where I was a software engineering intern at Nvidia
in the Summer 2010, 2011 and 2012. While I was in Silicon Valley, I joined an Internet-based start-up, where
I was assigned a project that involved natural language processing and data mining. I was glad that I could
use my theoretical knowledge in a practical scenario, even though this was a very different experience than
my research work.

I have just began my Ph.D. studies at University of California, Santa Cruz, but I have already involved
myself in a very innovative new research project on face detection and gaze tracking. The goal of the project
is to help students learn more efficiently while reading scientific books and documents on their computers,
by automatically analyzing their gaze patterns and facial features to find out what parts of the text they
find most interesting, challenging or important. With this information a computer program would be able
to highlight the parts of text which the student should revise most carefully.

With my passion for science, as well as expertise in both artificial intelligence and mathematics, I think that
during my graduate study as an NSF fellow I can make valuable contributions to computer science.

1
 Proposed Research
Michał Dereziński NSF application

The motivation for my current research came from my own practical need. As a student, I have to read a
lot of research papers, digitized textbooks and other scientific documents. However, just reading them once
is often not enough. Students need to review the material before exams, look up facts and definitions in
previously read papers. My goal is to create software algorithms and interfaces which would allow us to
find information in previously read documents faster and more conveniently, based on the information about
our behavior during the first time we read those documents. The main concept which made me believe
such a project would be valuable is gaze tracking. It is a computer science research field whose aim is to
calculate the trajectory of points on a computer screen that is followed by the eyes of the person looking at
the screen. The gaze behavior of a person reading a document provides a vast amount of information about
which sections they find most important or interesting, and which keywords ought to be remembered mostly.

Gaze tracking has many potential applications that are explored by researchers and technology companies
all over the world. In most cases it requires expensive specialized hardware to be performed accurately. The
eye trackers sold by Tobii, one of the industry leaders in the field, allow scientists to, for example, develop
new innovative human-computer interaction methods ([1]) or to analyze advertisement engagement ([2]).
However, the ideal scenario for gaze tracking to become ubiquitous in modern technology would be to use
the cameras built into the modern laptops and tablets. Unfortunately, the best approaches so far have not
been very accurate ([3]). I intend to overcome this problem by narrowing the task of gaze tracking to the
specific scenario of reading text documents. When we read, the gaze trajectory follows a very predictable
pattern. This should allow for a significant noise reduction of measurements, if we employ machine learning
to predict the most likely trajectory based on the given aproximate data. This approach requires developing
a new machine learning algorithm, which could potentially have other applications for predicting trajectories,
and it would also be an interesting expansion of theoretical learning theory.

Eye tracking is not the only relevant information that can be obtained about a person’s behavior when
reading. Other facial features can be a good indicator of whether a person is deep in thought, confused
or uninterested. Matching these reactions with the specific section of the text being read gives us a better
assessment of what is the most relevant information in the document. Face recognition is easier than gaze
tracking, so there are many software products providing such functionality, like the one created by Visage
Technologies.

Developing an application that would be practical for students presents other challanges, like designing a
good interface or creating algorithms for analysing the structure and contents of scientific documents. I hope
that I will get a chance to work alongside other talented researchers, whose espertise would be invalueable
in completing such a project. I think that the results of my work could benefit many students and scientists
in their pursuit of knowledge.

References
[1] Debeljak, M., Ocepek, J., Zupan, A. Eye Controlled Human Computer Interaction for Severely Motor
Disabled Children. Lecture Notes in Computer Science. Springer Berlin Heidelberg, pp. 153–156, 2012
[2] Hervet, G., Guérard, K., Tremblay, S., Chtourou, M.S. Is banner blindness genuine? Eye tracking internet
text advertising. Applied Cognitive Psychology 25, 708–716, 2011.
[3] C. Holland, A. Garza, E. Kurtova, J. Cruz, and O. V. Komogortsev. Usability Evaluation of Eye Tracking
on an Unmodified Common Tablet. Proceedings of the ACM CHI, Paris, France, 2013, pp. 1-6.