Graduate Talks: Polina Kirichenko
In 2018 Polina graduated from Faculty of Computer Science’s Applied Mathematics and Computer Science bachelor programme with honours. While studying, she worked for three years in Bayesian Methods Research Group. Last year Polina began PhD programme at the School of Operations Research and Information Engineering, Cornell University (USA). She is currently working with Professor Andrew Gordon Wilson’s research group. Last summer the laboratory has been relocated to New York University. Now Polina is studying at the NYU Center for Data Science’s doctoral programme. In this interview, Polina talks about her research, PhD studies, and her decision to do research work.
What do you remember best about your education?
Unlike the majority of Russian universities, HSE University and Faculty of Computer Science pay much attention to your activity during the semester – your homework, tests, and colloquiums. I remember well how we all worked together on problem-solving and exam preparation. We had a group of interested people striving for perfection in every project and homework. We helped each other to learn more and understand better. This was cool.
When I first started to study for my PhD, I lacked it. The atmosphere and mentality here are different. It is uncommon to discuss your tasks with others. I think that we had this sort of academic responsibility back home but we also worked much closer together, inspiring each other.
What do you do now?
I started my PhD studies at Cornell right after I graduated. After my freshman year, our professor and head of the laboratory moved to New York University with the laboratory itself. Cornell and NYU doctoral programmes differ slightly. At Cornell we were focused on the studies, I took many applied mathematics classes. It was harder because we had more practice and programming than mathematics at the Faculty of Computer Science. Now, at NYU I don’t have much to study, it’s mostly research.
While studying at HSE University, I worked with Professor Dmitry Vetrov, who was my academic supervisor. At first, I worked with deep learning projects, and then I switched to Bayesian and neural Bayesian methods. These methods for neural networks teaching not only find parameters to suit a singular task but also approximate parameter distribution, which allows us to account for the uncertainty from the limited data. This way we can estimate and calibrate the uncertainty of neural network predictions, which is important in such applications as self-driving cars or medicine. I still work with Bayesian and probabilistic methods of deep learning.
Now we are employing probabilistic neural networks for anomalies detection or out-of-distribution detection. For example, we want our network to classify cats and dogs. We teach it with images of cats and dogs. If, while testing, we put in a duck image, the network must tell that it is neither cat nor dog. Since the network does not know what it is, it must give the most uncertain prediction. The network must recognise this input as an anomaly in relation to the known data set.
How are your PhD studies?
American PhD studies differ from European ones. European doctoral studies are closer to Russian ones since they imply that a PhD student has both bachelor’s and master’s degrees. The focus of the European PhD is a research project, which lasts three or four years.
American PhD is longer and most students enrol right after receiving a bachelor’s degree. From the European perspective, it combines master and doctoral studies. For two years, one has to take a lot of classes. If you study Computer Science or Machine Learning, you need to take statistics, machine learning, algorithms, databases, and other related courses – to prove that you indeed are an expert in the field. Sometimes you also have to teach relevant courses. On top of this, you have to do research full-time, which slows your progress. In two years, when you are done with studying and teaching, you can work on projects, write articles, give speeches, get research internships and so on.
Besides, an American PhD is more flexible. You study at the faculty, not at the particular laboratory. During freshman year you can change your academic supervisor or laboratory if you want to. European PhD is different, it’s more like a job. You are tied to a laboratory, which finances your research so that you cannot change your project or your supervisor.
You could have worked in the IT industry. Why did you choose to do research instead?
I had two internships at Google while studying for my bachelor’s. These were software engineering internships where I learned how this works in the industry. Simultaneously, I did research work under Professor Dmitry Vetrov. Suddenly I realised that I am more interested in research.
Both industry and research have their advantages and disadvantages. Research attracts me because you never know the answer. In engineering, you know your task, the steps you need to take, and whether it is possible to solve it. In research, you can ask a question, not knowing whether it is a right question, and realise that it is wrong after months of work. This lack of clarity can be overpowering, but it is inspiring too. In your search, you do all sorts of things – you code, you read, you analyse, you discuss.
What difficulties do you have now?
A number of them. One most relevant for machine learning is an astoundingly quick development of the field. Other fields of computer science and mathematics are somewhat different; articles are published in journals, one project can last for years. In machine learning, several relevant articles are published every week. Someone might have already realised an idea you were working with for months.
This maddening pace is very overpowering because you have to work hard and learn about every new development in the field. You also have to maintain the quality of your work. There is a gentlemen’s agreement that you have to have at least three published articles to get a PhD. It is not easy at all.
At the laboratory, we often have team projects when several students write one article. This allows for better strain distribution – someone experiments, someone tries a new approach or improves the model. Such projects work smoother. When you do everything alone, it is hard to keep up.
You often present at the conferences. Could you tell us how it helps you?
It’s important to get acquainted with your colleagues. Sometimes it is much more useful to talk to them than to sit for weeks unable to solve some problem. There are many people out there developing machine learning, and the chances are high that others tackle similar issues. It’s useful to get to know them, to learn something new, and to share your own experience. Besides, I think that at the conferences master students can meet people they might work with in the future, or find a research internship there. During the last conference, I met a lot of people and I was interviewed by several companies because I was looking for a summer internships. I had some five interviews in five days.
Presentation experience is vital too. To work at the office and write an article, which may be accepted for the conference, is one thing. It is very different when you need to present your idea and to demonstrate to your colleagues that your method and your problem are important and interesting and should be developed.
What skills and knowledge from your bachelor studies are now useful to you?
I understand more and more that bachelor’s education is about the base. Faculty of Computer Science has built the base of my mathematical and coding skills, and I build on it.
It was crucial for me to learn how to learn during bachelor’s studies. To discover new fields and to structure new information. This skill was not easy to get. Faculty students have everything ready-made – the professor with the expertise and the course. They only need to digest this prepared knowledge. HSE has this rigid and simple deadline structure, which helps to understand how to consume information. When you go out in industry or research, you have to rebuild this structure on your own. You need deadlines, internal motivation to develop and adapt. I think that the Faculty has helped me to develop this structure too.
Could you give some advice to a freshman or sophomore student who wants to go and get a PhD, as you did?
It all depends on the specific field of computer science. As for my field, I can say that machine learning is getting more popular. When I enrolled in doctoral school, I didn’t have any articles, just my work experience with Professor Vetrov and Google internships. I think that you need one or two articles to get to a top doctoral school these days.
There is another way. It may be hard to write an article during a bachelor’s studies. It seems to me that internships are more accessible, especially research internships. It has suddenly become trendy at the Faculty to apply for industrial internships to Google and Facebook. I learned about research internships only in the fourth year. If a student wants to work for some company like Google after graduation, they should take an internship in Google. There is an internal interview for ex-interns, and the chances to get to a full-time job are high. If a student wants to do research, they should go for a research internship first. Switzerland has a number of programmes at EPFL, ETH Zurich, and Austrian IST. I was looking for such a programme during the fourth year of my studies, in case I didn’t get to doctoral school for the first time. In general, it is a very interesting experience, which lets you understand whether you want to do this for the next four or five years.