‘We Facilitate High-Speed Car Crashes and Study How Car Engines Work Based on Photos of Flying Debris’

Nikita Kazeev

Photo courtesy of Nikita Kazeev

Nikita Kazeev holds a Candidate of Sciences degree (Russian equivalent of a PhD) in Computer Science and a PhD in Physics. He is a Research Fellow at the LAMBDA Laboratory and works at CERN. In an interview with HSE News Service, he talked about what it was like to defend his dissertation in a double doctoral degree programme at HSE University and Sapienza University of Rome, what it is like to conduct research in Geneva, and why it is imperative to communicate with colleagues.

Double Doctoral Degree Programme

I met my HSE supervisor Andrey Ustyuzhanin at Yandex when I was studying at the School of Data Analysis. I was invited to intern with his group, which was engaged in machine learning for high energy physics at the European Organization for Nuclear Research (CERN). I met my dissertation advisor in Rome under rather unusual circumstances. The collider operates around the clock, and researchers visiting CERN do monitoring shifts. During their eight-hour shifts, the scientists keep track of how the experiments go. During one shift I met Barbara Sciascia, and later she invited me to the INFN Frascati laboratory.

In Russia I work at the LAMBDA laboratory. The lab’s main areas of research are based on the fact that machine learning methods have made a quantum leap in the past ten years. They can be used not only for already routine tasks, for example, to distinguish cats from dogs in photographs on social media, but also for more non-trivial ones, such as making discoveries in the natural sciences. The lab’s flagship area involves collaborating with the LHCb experiment at CERN, and this was the area in which I conducted my dissertation work. In addition, the laboratory optimized the design of the SHiP detector, computed evasive maneuvers for spacecraft, proposed a new black box optimisation at the leading NeurIPS conference, and, since 2015, they have been holding an annual summer school on machine learning.

Of course, the dual doctoral degree programme was not without problems: for example, unlike HSE, at the University of Rome it was not possible to defend a collection of articles instead of a dissertation manuscript. You had to write a full-fledged PhD dissertation, supplemented by classical standards with a manual on machine learning and the hardware of the LHCb experiment. They believe that a good researcher should be able to communicate their results to a reader who is not an expert in the field. In addition, HSE had its own peculiarities in terms of document flow, and as a result, my Italian supervisors had to send reviews of my dissertation by mail. On the other hand, in Rome they got acquainted with my work in detail and asked good and interesting questions. In addition, thanks to this programme, I increased my Italian vocabulary from 5 to 15 words.

The pandemic did not particularly affect my work. At CERN they traditionally work remotely due to the large number of collaborators in different countries. I was in constant communication with the Italian group when I was working on a joint project with them, and before the pandemic, I spent less than a month total in Rome, as the main place of work for everyone was CERN in Geneva. We had originally planned for my supervisors and committee members to meet for the defense in Moscow, but this idea had to be abandoned.

Projects

Roughly speaking, our work at the Large Hadron Collider is like this: we facilitate high-speed car crashes and, using photographs of flying debris, we study how car engines work. Before starting to draw conclusions about the physics, it is necessary to determine the types of components in the photographs. My work involves determining the types of particles in a detector using machine learning. In total, my dissertation included four projects.

The first is global particle identification. Different detector components collect different information about the particles passing through them. If we go back to the example of cars, then we basically use cameras from different angles with different light filters to shoot a ‘collision’. The information obtained allows us to determine the type of particle. To solve this problem, we developed an algorithm based on machine learning (CatBoost), which handles this task better than the previous solution based on a simple neural network.

Photo courtesy of Nikita Kazeev

The second task is the identification of muons. Among other charged particles, muons are especially interesting in that they have a high penetrating power—they can pass through calorimeters and iron sheets of absorbers. We developed an algorithm that can quickly determine if a particle was a muon.

Another project involves the use of machine learning on noisy data. While working on the problem of muon identification, we ran into a problem. Our algorithms were trained on real, non-synthetic data, so their labeling was inaccurate—not all particles that were identified as muons were actually muons. But for each of them, we knew the probability that the label was wrong. We developed a method to train machine learning algorithms on such data. Since noisy data is used not only in physics, but also in most applications of machine learning in the real world, we are now actively looking for where else a similar noise model has to be used. If you are reading this and you have suitable data, please feel free to write to me.

In addition, I am engaged with fast simulation of Cherenkov detectors. Simulated data is required to create and validate algorithms that parse detector data. To do this, we collide virtual particles on a virtual detector using a special computer program. As a result of such an experiment, we have accurate information about what actually happened. The problem with simulated data is that it requires large computing resources—about one second for one event, and tens of millions of these events are required. I built an algorithm based on generative adversarial networks (GANs) that simulates the response of a detector to particles passing through it. This approach is two to three orders of magnitude faster than the simulation described earlier.

Almost all of these projects are interesting because they serve as tools for fundamental research.

We don’t know in advance what new physics we will discover, or how and when it may come in handy

This may sound a little sad, but in fact the opposite is true: out of pure scientific curiosity, mankind has already acquired electricity, radio waves, semiconductors and many other achievements that surround us. Well, from the obvious benefits, in addition to being used in a specific LHCb experiment, the developed methods can form the basis of algorithms for other physics experiments.

Remote Dissertation Defense

Initially, the defense was planned as a big party, which would have brought together foreign committee members and the dissertation advisors, but because of the pandemic, this did not work out. On the other hand, the online format made it possible for a large number of my friends to attend, not all of whom could have come in person. We considered an option in which those who wish could meet in person, and the rest could attend online, but this proved to be technically difficult. There was also even an idea to conduct the defense at the Italian embassy, ​​but, again, the pandemic made its own adjustments.

The defense itself passed predictably and according to plan. All the committee members sent their reviews in advance. I was pleasantly surprised by the quality of the discussion, questions, and suggestions. After the defense, we celebrated with a buffet reception and a live action role-playing game.

In conclusion, I would like to say that the main thing is to not give up when plans don’t work out (due to the pandemic), to love your work, and not to miss out on interacting with colleagues. Who knows—your future supervisor could be sitting next to you right now.