Skoltech recently hosted a hackathon on the topic “Applying AI Methods to Address Sustainable Development Challenges,” held as part of the “Bayesian Machine Learning Methods” course in the Data Science master's program. The hackathon involved 38 students working on individual research projects under the guidance of Skoltech AI Center research engineers — Alexander Kolesov, Sergey Kholkin, and Grigory Shutov.

The awards ceremony was led by Professor Evgeny Burnaev, Director of the Skoltech AI Center. Evgeny congratulated the participants, presented the winners with branded hoodies featuring Skoltech and AI Center logos, and shared insights about the future development of artificial intelligence technologies.

Key Hackathon Projects

Mikhail Lukanov
Adapted the continuous normalizing flows algorithm in the Flow Matching paradigm using an equivariant neural network. He successfully obtained structures of small organic molecules with acceptable geometric properties.
"It was fascinating to combine chemistry knowledge with skills gained in the course to solve such a challenging task. Everything at the cutting edge of science today requires an interdisciplinary approach, so it’s essential to step outside your specialization and try something new!" — shared Mikhail.

Ramil Khaffizov
Developed a method to improve image generation quality with a limited number of “denoising” steps. His approach demonstrated high-quality results with just 1, 2, 4, and 16 steps, compared to traditional diffusion models that require over 100 iterations.

Alexander Zaytsev and Vasily Kakurin
Compared two methods — Image-to-Image Schrödinger Bridge (I2ISB) and Diffusion Schrödinger Bridge Matching (DSBM) — to evaluate their efficiency in restoring corrupted images.
"Modern diffusion models require significant computational resources, which limits their use on devices with constrained capabilities, such as smartphones," the authors noted.

Anna Borisyuk, Arseny Ivanov, and Vladislav Gromadsky
Proposed a way to accelerate the denoising of Flow Matching models using the Shortcut approach.
"We implemented Shortcut Flow Matching and compared it with the conventional Flow Matching model to evaluate improvements in inference speed and the quality of generated images," said Arseny.

Alexander Sharshalvin and Elfat Sabirov
Developed a method for latent code processing in deterministic diffusion using the Denoising Diffusion Implicit Model.
"We successfully identified an operation that preserves the object’s original style with minimal losses," said Alexander.

Timur Nabiev and Egor Miroshnichenko
Applied Discrete Flow Matching technology to the task of image resolution enhancement. The method produced excellent results by increasing the resolution of black-and-white images from 32×32 to 64×64 pixels.

Yaroslav Abramov, Ekaterina Filimoshina, and Marina Sheshukova
Adapted the Diffusion Rejection Sampling (DiffRS) approach for significance-based diffusion sampling, significantly reducing computational costs.

Aysel Mirzoyeva and Svetlana Lukina
Investigated the use of neural optimal transport with a Truncated Diffusion Model enhanced by a Variational Autoencoder (VAE).
"We visualized the transport map and generated samples. The results confirmed the effectiveness of the truncated diffusion model enhanced with VAE," emphasized Svetlana.

Congratulations to all participants, and we wish them success in their future research and projects!