The Skoltech Summer School of Machine Learning (SMILES-2023) concluded on September 1. Held in the scenic mountainous area of Belokurikha in the Altai region, the school offered a 12-day intensive course, culminating in poster presentations and defenses of team projects created during the hackathon on AI methods for addressing industrial challenges of sustainable development. The winners received certificates signed by Skoltech president Alexander Kuleshov and the head of the Applied AI Center Evgeny Burnaev.
“We decided to hold the school away from Moscow for a reason: it helped attract students from other regions. Moreover, being far from civilization meant that students had no choice but to focus on classes and project work,” says Professor Burnaev, the director of SMILES-2023. “In our 12-day course, we aimed to cover as much advanced research in deep machine learning as possible and to hone in on topics of particular interest to the global machine learning community. Some of the research first presented at the International Conference on Learning Representations in May was incorporated into our lectures here in Altai. Our overall impression is that the school attracted bright students from all over Russia who produced outstanding projects. Once they return home, we believe they will delve into machine learning with even greater motivation and success.”
Winners of the poster session:
— Alexander Tolmachev presented the poster “Analysis of deep neural networks using the information bottleneck method with lossy compression”. “Recently, neural networks have found extensive use in medicine, unmanned vehicles, natural disaster prediction systems, and other applications. The big question now is exactly how a neural network makes decisions and why we can trust it, especially in such critical cases as diagnosing health problems based on lung images or predicting the spread of forest fires,” Tolmachev notes. “The information-theoretic approach holds much promise when it comes to analyzing the learning process of neural network models. In our research, we proposed a method based on the compression of internal representations for analyzing data shared between the outputs of a neural network’s layers.” Going forward, the researcher plans to enhance the method with the novel normalizing flow technique, which has already led to breakthroughs in various areas of machine learning. “We will also experiment with different neural networks to thoroughly explore their respective information planes. This research will bring us a step closer to answering the key question of how a neural network learns from the perspective of information theory,” he adds.
— Tatiana Zaitseva’s study “Regularity of tile B-splines” explores a special kind of function systems. “Tile B-splines form the core of algorithms for geometric modeling of surfaces. Similar algorithms are used in animation, for example, by Pixar. They help add new details to the modeled figure at each iteration,” Zaitseva explains. She studies the smoothness of B-splines, which affects the quality of the resulting surfaces and the convergence of applied algorithms.
— Artem Gallyamov’s poster “Development of a surrogate model for hydraulic fracture cleanup” describes his project that has enabled reducing the running time of the existing hydraulic fracturing simulator from 3 hours to 0.1 seconds thanks to a machine learning surrogate model. His research will enable massive calculations in hydraulic fracturing simulation and selection of the most appropriate parameters, which will lead to a 20% increase in cumulative oil and gas production. In the future, Gallyamov plans to apply various neural network architectures and improve data quality.
Winners of the hackathon:
— Anatoly Onishchenko and Stefan Maria Ailuro developed a model for predicting ice behavior within a bay in their project “Predicting the dynamics of radar satellite images based on neural network PDE”. The model helps enhance maritime safety in the north in the absence of sufficient satellite imagery. “In the north, it is crucial for seamen to have a clear picture of ice movements in order to avoid threats to navigation. Typically, they go by satellite images. They are not taken on a daily basis, which rules out the use of classical machine learning methods,” Ailuro explains. “Our solution is special in that it helps capture the patterns of both the interaction between neighboring ice areas and weather effects, while not only making up for the lack of satellite images, but also predicting ice behavior. Thanks to this capability, the model can be used wherever real ice data are lacking.” The winners plan to carry on with their project by studying physics-inspired architectures and tweaking modern approaches to make up for the lack of data.
— Igor Udovichenko and Alexander Kolesov’s project “Estimation of the continuous entropy Wasserstein barycenter based on an energy model” focuses on searching for an average distribution in the form of a Wasserstein barycenter. “For years, it has been a purely theoretical challenge due to some limitations that make it impossible to find a suitable practical application,” the project authors explain. “The efficient novel algorithm marks a new milestone in this research area, which not only helps to elaborate the theory but also to discover a wealth of new applications for the algorithm”.
— Daniil Kolotinsky, Bogdan Kirillov, and their online partner Petr Hovental presented the project “Machine learning methods for studying the effect of gas-discharge plasma flow on the behavior of condensed-state microparticles”. The condensed-state microparticles levitating in gas-discharge plasma display non-reciprocal interactions in the plasma environment. “Understanding how plasma and condensed-state microparticles interact is important for many fields, including astrophysics, microelectronics, and plasma medicine,” Daniil Kolotinsky comments. “To gain a deeper insight into the processes occurring in such systems, researchers need fast and efficient ways to calculate the forces acting upon microparticles in a plasma flow. Our approach based on machine learning methods enables fast and accurate prediction of these forces.”
SMILES-2023 was attended by 65 master’s and PhD students from top Russian universities, who are involved in proactive research in the field of machine learning and its applications for modeling engineering and physical systems. The participants worked on their projects using Skoltech’s Zhores supercomputer specifically intended for machine learning and data-based modeling tasks.
Over 200 students joined the school online. Participation was free for everyone. The school’s information partners included the federal project “Artificial Intelligence” of the national project “Digital Economy” and the open platform “Russia is Land of Opportunity”. Check out the photos from the event.
Contact information:
Skoltech Communications
+7 (495) 280 14 81
