Sunteți pe pagina 1din 2

Personal Statement: Department of Awesomeness John Doe

What will come of a stronger relationship between the study of the mind and the study of artificial
intelligence? To grow this relationship, how do we identify and prioritize applicable insights coming from
each discipline? How do we effectively implement and test these insights? With my education in computer
science and psychology, my proven abilities to ask valuable research questions across disciplines and to lead
major research projects, and my experience in developing novel and effective artificial intelligence, I am in
a unique position to engage these questions and to offer a tailor-made background to the cognition program
and to others. This PhD will prepare me to advance the relationship between the studies of the mind and
artificial intelligence as it develops. I would be delighted to do this both in academia or in the industry as a
researcher, consultant, or teacher, based on how my skills and interests develop over the course of my studies
and research.
During my undergraduate psychology studies, my cognitive psychology professor responded avidly to a
question I asked concerning the hippocampus, part of the brain that is active when one accesses long-term
memories. I wondered: could we measure hippocampal activity to validate whether witnesses are accessing
long-term memory when recounting a scene? The total required use of an fMRI machine to explore this
question was outside of his budget, but the moment I saw his reaction to my question I began to consider a
future in research. Encouragingly, I later found a study that – though not via the hippocampus – suggested
that fMRI measurements of brain activity can distinguish between true and false memories [1].
When I joined Big Tech Company’s Research Division as a software engineer, I started a research project
during my free time. This project was based on an idea that I had while learning about artificial neural
networks. I presented a blueprint for the project to my management, and although I didn’t occupy a research
position, my manager granted me the freedom to pursue the research project during 50% of my working
hours.
Neural networks have an immense presence in modern technology – such as face detection – and improv-
ing their reliability is one of the most active pursuits in their development. During the training of a network,
a technique called dropout is employed: a subset of neurons are disabled, allowing other neurons a chance
to “catch up” if they are not contributing enough to the network. Neurons are disabled by multiplying their
contributions by zero – but what if we were to multiply the neurons by a non-zero constant instead? After
implementing and running experiments to answer these questions, I was thrilled to see a non-zero-coefficient
network produce substantially more accurate outputs. I later found that a similar concept had already been
proposed [2, 3]. Knowing that I had asked a valuable question and reached a compelling result was exciting
and encouraging.
The second research project I drove at Big Tech Company resulted in an algorithm that runs 11 times
faster than the fastest known competing algorithm. The algorithm addresses a computationally intensive
matrix search problem that spans a multitude of fields in computer science. The goal of any such algorithm
is to locate a rectangle within a given matrix whose inner sum is larger than that of any other rectangle’s.
The worth of such an algorithm is then typically measured by how fast it is.
The state-of-the-art approach to this matrix search problem was an approximation algorithm, achieving
a speedup over guaranteed-correct algorithms by forfeiting the guarantee to locate the globally optimal
rectangle. To run even faster than this state-of-the-art approximation algorithm, I evaluated only a subset
of the input matrix: select “slices” (single rows and columns) of the matrix. The algorithm was then
operating with even less knowledge of the input data, so I designed the algorithm to account for unknown
slices of the matrix via interpolation between known slices, resulting in a competitive output quality. I gave
a presentation of my work to an artificial intelligence team at IBM by invitation of the team’s manager.
Although my work was ultimately not accepted to CVPR or ICCV, it received very encouraging reviews
from those conferences. Of the total of six reviewers, all six of them specifically complimented my writing,
the process of which I thoroughly enjoy and take pride in. Per the recommendation of an ICCV reviewer, I
will submit this work to BMVC 2020 [ref].
Before continuing, I would like to highlight the following. Due to an unexpected illness during my
undergraduate studies, I was unable to continue with the honors college and I graduated with a 2.6 GPA.
However, I focused on my core computer science classes, wherein I achieved a 3.6 GPA despite being ill.
Thanks to a finalized treatment, I have been healthy and symptomless for over a year and a half.
Personal Statement: Department of Awesomeness John Doe

Expanding my knowledge of the intersection of the human mind and artificial intelligence will enable
me to advance these two areas of science that I find both deeply engaging and foundational to our future.
I’m excited by the prospect of drawing from and furthering School’s interdisciplinary nature in the study
of the mind – including the integration of artificial intelligence (as in Dr. Billy Bob’s work, for example).
Dovetailing with my background and experience, these things have made my decision to pursue this PhD,
albeit momentous, very straightforward.
I was thrilled to see a paper from Dr. Bob’s lab present a visual cortex-inspired neural network that
reproduced an optical illusion consistent with one that humans experience [ref]. Seeing this artificial agent
display such an intricate aspect of human cognition is truly something to marvel at. Discoveries like these
support the mounting evidence that the studies of the mind and of artificial intelligence can meaningfully
inform one another. Given the opportunity to study the underlying computations of vision with Dr. Bob,
my primary objective would be to bring an awareness of both disciplines – visual cognition and artificial
intelligence – to the other, asking and pursuing how they can inform one another. To this same end, I would
also be very excited to work with Dr. Fred Smith in his study of neural mechanisms more broadly. In
working with either of these professors, I will have the incredible opportunity to discover new aspects of the
human mind and to be part of the emerging new era of artificial intelligence.

References
[1] R. Cabeza, S. M. Rao, A. D. Wagner, A. R. Mayer, and D. L. Schacter. Can medial temporal lobe regions
distinguish true from false? An event-related functional MRI study of veridical and illusory recognition
memory. Proceedings of the National Academy of Sciences, 98(8):4805–4810, 2001.
[2] Y. Li, R. Xu, and F. Liu. Whiteout: Gaussian adaptive regularization noise in deep neural networks.
stat, 1050:5, 2016.
[3] X. Shen, X. Tian, T. Liu, F. Xu, and D. Tao. Continuous dropout. IEEE transactions on neural networks
and learning systems, 29(9):3926–3937, 2017.

S-ar putea să vă placă și