"It Is Important to Value Your Time and Effort"
Arina Rak, a student of the Applied Mathematics and Information Science programme, took part in two internships abroad despite the pandemic. She tells us about the difficulties she had to overcome and how she managed to achieve her goal.
Enrolling and studies
My brother is a programmer, so I'm always around a person interested in computer science. But it wasn't because of him that I got interested. When I was in the 8th grade, Anya (my brother's wife) and I started taking an online course in Python together. I liked it. Anya quickly dropped out, but I got interested, finished that course, then others, and finally transferred to a math school. So it was all a happy accident. When I got acquainted with Python and realised that I was interested in it, I went from my school with advanced English to the lyceum of the Bauman University - my brother was studying there and we knew that it gives good mathematical training.
In the 11th grade, I didn't know where I was going to enrol; I had just a rough idea of the study programmes available. I was choosing between Moscow State University, HSE University and Moscow Institute of Physics and Technology. Of course, it was also about the teachers, for whom this is often not their main job; they are practitioners who also teach at the same time. This affects both the average age and the relevance of knowledge. Our teachers are just people, and you can chat with them after the lecture as with a programmer, a specialist in their field, rather than a professor. Applied Mathematics and Information Science programme are closer to me because I am interested in mathematics and I want more mathematics than programming. And, of course, the fact that HSE University is not in Dolgoprudny (laughs).
The experience of the lyceum helped me in my first year when graduates of ordinary schools were intimidated by long lectures and complex words. During the first two years, we had a lot of maths. Often programmers question the need for such an abundance of mathematics. In our linear algebra workshops, we had problems, for example, on matrix decomposition, which we solved in Python. Such applied examples help us understand why we are studying; it motivates us and makes us interested. All the courses, in general, are built to be useful in the future; we didn't have to study for the sake of studying.
I'm majoring in Machine Learning and Applications; I have a bit more experience in natural language processing. When I applied for internships, I wasn't focused on a particular area of deep learning. I wanted to try everything, understand how things work in each area and make a choice of both a career and a master's degree.
The selection process was very long. First, I submitted the documents, a CV and a motivation letter. Every two or three weeks, the school sends heaps of rejections. I applied together with a group of people; there were several waves of rejections, my acquaintances received rejections in the third or fourth wave. After that, if you are lucky, you get into a pool where managers consider you.
When I applied for my first internship at EPFL, my first interview was on the subject of computer vision, but the communication did not go well; the task seemed uninteresting to me. When I was asked about my experience, I casually mentioned that I had a course in natural language processing; and the interviewer simply told me that he had a colleague who needed an intern on this topic and recommended me to him. We talked to this colleague, he had a great assignment, and he invited me straight away. So, after casually mentioning the topic, I found myself a great supervisor.
At the beginning of March, I started to prepare the papers, and in the middle, they cancelled mobility completely. But I was already interested in the task; I liked the supervisor, I told them I wanted to do it remotely. They didn't move it online; I asked for it myself. I still haven't seen the supervisor, although we've been working for a year. Now that I have left for an internship in Karlsruhe, I will finally be able to meet him.
At EPFL, we were doing multi-hop question answering, where there is a question, and to answer it you have to find several relevant pieces of text. In a classic formulation of this problem, you have a question, you have a text, and somewhere in that text, there are one or two sentences in a row that answer that question. And with multi-hop, you have to make several jumps, which means that the answer isn't localized, but there are several areas in the text that are responsible for it.
We mainly focused on the fact that modern models for natural language processing can't handle long sequences, and when a problem is formulated in this way, very long texts are usually given; they need to be "cleared" first so that only the relevant part is fed to these smart models. So we dealt with filtering out the irrelevant in texts. Practical applications of this task are search, dialogue assistants, e.g. if there are several messages and you need to find some answer in them.
The topic of my second internship at the Karlsruhe Institute of Technology is computer vision for biology. Our data are three-dimensional microtomographs of the abdomen. The challenge in this field is to mark up the data - it is very time-consuming and requires medical professionals. Marking a single scan of a single person can take up to four hours. To do a complete abdominal segmentation you have to have a very large amount of data and it is very expensive. For two-dimensional data, for other domains, this task is quite manageable; it is in medicine that it is difficult to do the partitioning.
We wanted to train a model that could mark up the entire abdomen based on several sets of data with different mark-ups. For example, in one place, kidney experts got together and marked up just the kidneys and marked up everything else as background; for another medical task, they also marked up the liver, and so on. We used the contrastive loss to train representations for each pixel from these heterogeneous datasets and then trained small classifiers on these representations. The results of the model can be used for further data analysis: to measure the average size, the average width of an organ; we also had data about tumours: by training on them, you can detect tumours on new data, you can use them simply as medical atlases.
Here I did manage to visit the campus a couple of times, I saw my research supervisor, but I still work remotely from the dormitory.
Going abroad in a pandemic
Getting abroad was surprisingly easy. All the difficulties that came up were related to paperwork and applying for a grant. I was very easily granted a visa for three years, although others find it very difficult to get even to a conference. I think it was because I was sponsored by a German state foundation.
I know English; my German is at a super basic level, just enough to go to the shop. My best result is a dialogue in a coffee shop for eight lines on each side. I used to think that everything had to be said perfectly, now for me the most important thing is to say what you need to say. I don't know how much I follow the rules of the German language, but I don't have a goal of speaking the perfect German.
I was very lucky to leave on time; if I had left a week later, I don't know if I would have been allowed into Germany; because there was a new wave of coronavirus in Russia. I got in at a time when the quarantine for Russia was lifted and just came in with a test. Every week something started to open up, cafes, museums. Testing is quick and free; the test is done in ten minutes and is valid for one day. I was able to go to other cities and even visit my brother in Zurich.
I like walking and take long walks in the city. I also do indoor cycling. I watch series like everyone else. I've seen Friends a million times and I liked The Bridge series very much.
This year I want to work in the IT industry. I already have some idea about academia, but I have no idea about the industry at all. I only worked at Yandex in my second year. In a year, I want to apply for a master's degree based on what I see in the industry. For my master's degree, I am considering Europe - Switzerland, Holland, Germany, Scotland. It would be ideal to apply to EPFL after the positive experience I had during my internship.
Advice for future participants
It is important to value your time and effort. If you undertake something, try to see it through and convey to others that your time is valuable. You have to have a kind of bravado. Normally, I would have just regretted cancelling the internship, especially as no one offered the online option. Here I spent a lot of time and effort, so I managed to get an internship and then turned it into a degree.