'I Would Like to Leave a Lasting Impact on Science'

Aibek Alanov pursues his own scientific research and leads two teams of scientists, one at HSE University and the other at AIRI. In this interview for the HSE Young Scientists project, he explores the parallels between today's AI researchers and early 20th-century physicists, discusses generative models, and shares his passion for bachata partner dancing.

How I Chose a Career in Science

At the start of my career, I explored various paths. During the third year of my bachelor's, I secured an internship at Yandex Research, a division of Yandex where scientists work. It was then that I realised what researchers actually do and discovered that science offers more freedom in choosing tasks, with no strict guidelines on what to do or how to do it.

A year later, I interned as a programmer and developer at WorldQuant, where the focus was more on programming in the financial sector, such as predicting stock prices and similar tasks.

This experience proved valuable, but I realised that I am more drawn to research. I would rather focus on writing and publishing scientific papers and presenting at conferences.

I also joined the Bayesian Methods Research Group led by Professor Dmitry Vetrov. In the fourth year of my studies, he became my academic supervisor. From him, I learned how to set objectives, conduct scientific research, formulate and test hypotheses, and write academic papers. 

I then worked as a researcher with Professor Vetrov for several years at the Samsung AI Centre. From there, I moved on to the Artificial Intelligence Research Institute, AIRI, where I currently lead the Controllable Generative AI group. 

New tasks arise all the time, which motivates me to generate ideas, collaborate with colleagues, and write articles. If these articles are accepted at international conferences, it serves as further proof that our work is valuable and relevant. 

What Are Bayesian Methods 

Bayesian methods are an approach used in statistics and machine learning. They help minimise uncertainty. For example, let's say we want to predict a company's stock price. This value is influenced by many factors, and Bayesian methods allow for more accurate predictions based on the information collected. If aggregated correctly, this information can reduce uncertainty to a minimum range, ensuring that the prediction accurately reflects reality.

The Subject of My Research

Our centre conducts a variety of studies, ranging from fundamental research to more practical applications. They primarily focus on generative models. One example is diffusion models, which enable the generation of new data. Stable Diffusion and DALL-E are widely discussed today—these models can generate realistic images from text descriptions. There is also a distinct type of models that generate text, such as the well-known ChatGPT. We work with all these classes of models: develop methods to improve them, explore their properties, and explain why they work—or sometimes don't.

In Russia, we have locally developed solutions in the field of generating images from text descriptions, such as Kandinsky from Sber and Shedevrum from Yandex. They are now actively used in a variety of applications.

My Personal Research Focus

I primarily work with generative models for images, using them for various image manipulations. For example, I use these models to edit real images. We can take a photo or a drawing and instruct the model to modify a specific part of it. For example, if the image is of a person, we can change their hairstyle, clothing style, or the style of the image itself. 

There is also personalised generation, which is used when we want to generate specific entities rather than arbitrary objects.

What I Take Pride In

Over the past two years, I have published several significant papers. They did not go unnoticed: at the end of last year, I received the Yandex ML Prize for achievements in science. The first publication in 2022 was accepted to the prestigious NeurIPS Conference, one of the leading international conferences in artificial intelligence. The paper was titled Universal Domain Adaptation for Generative Adversarial Networks.

It presented an approach for training a generative model when data is limited. For example, if we want to generate photorealistic faces, we can collect a large number of photorealistic face images and train a high-quality generative model on them. 

However, if we want to generate faces in the style of a particular artist, we won't be able to collect a large dataset. There are not many paintings in that style, so we won't be able to train a high-quality generative model with such a limited number of examples. A task arises: how can I use a generative model trained on a large dataset of photorealistic faces to generate faces in the style of a specific artist? 

This task is called 'domain adaptation,' where we adapt the model to a new domain with limited images. In that paper, we proposed a model that performs this task effectively, uses several thousand times fewer parameters than existing approaches, and still achieves the same level of quality.

In our next paper, we further developed this method and explained how to make it even more effective. It was accepted to another prestigious conference, ICCV, focused on computer vision and artificial intelligence, and was published last year. 

I have also published papers on sound generation, where we address the task of improving sound quality and removing noise.

My Dreams

I believe every scientist has a dream of creating something that will transform the field in which they work. Write a paper that will remain relevant for years to come. Develop a new approach. Leave a lasting impact on science.

If I Hadn't Become a Scientist

I think I would have become a film director. Science is very similar to art in the sense that both require inventing something new. Writing an academic paper is like creating a complete work of art through which the author aims to convey a message. That's why I can understand artists who create, and I can relate to their process. If I hadn't become a scientist, I would have pursued art—making films or theatre productions.

Scientists I Would Like to Meet

There are two. The first is Isaac Newton. He founded many key areas in mathematics and physics and was one of the first to view the world from a rational perspective. His scientific approach enabled him to propose ingenious hypotheses that proved to be true, and he was able to test and describe them using advanced mathematical language. It would be fascinating for me to understand his thought process and how he viewed the world. 

The second scientist is Alan Turing, the founder of programming and computer science. He introduced the concept of the Turing machine, which is fundamental to our field. 

My Interests besides Science

I go to the gym three times a week in the evenings, and it really helps clear my mind. I take long walks on weekends and spend time with friends. I've been practicing bachata partner dancing for a while. This also helps me relax.

Advice for Budding Scientists

Try different tasks to discover what resonates with you. Also, find a good academic supervisor with whom you feel comfortable and inspired. If you manage to find such a supervisor, it greatly increases your chances of succeeding in science, discovering your own research topic, and writing high-quality papers.

My Favourite Location in Moscow 

Neskuchny Garden. It makes me feel nice and relaxed. I enjoy taking walks there, either alone or with friends. It feels particularly pleasant in the evening.