At the AI Centre the theoretical foundations of machine learning are being explored, specialised neuroarchitectures are being developed, and the efficiency of training is being improved, which reduces the costs of computational resources. The aim is to enhance the accessibility and performance of AI technologies for science and business.
The development of new algorithms for training, fine-tuning, and accelerating fundamental and generative models is underway
Tasks
Tasks include the development of new algorithms for training generative flow networks and their application to the task of generating data from a given distribution. The creation of a new method for generative modelling based on the application of stochastic optimal control approaches to the Schrödinger problem is also a focus. Furthermore, the development of a diffusion model in a hierarchical text space that takes into account semantics and the multilayered structure of meaning (where general ideas are represented at a lower level and details and stylistic features at a higher level), as well as a hybrid scheme for text generation, is being pursued.
The complexity of deep learning models is being analysed, along with the exploration of linear algebra, statistics, and optimisation algorithms that play a key role in their creation
Tasks
- Evaluating the accuracy of gradient recovery of the logarithm of the density (the so-called score function) using the denoising score matching method, applied in the training of generative diffusion models, as well as assessing the accuracy of velocity field estimation in the flow matching method.
- Development of new mathematical methods for working with generative flow networks, in particular, the development of a training methodology for non-acyclic generative flow networks.
- Analysis of the effectiveness of combining transformer architectures for sequence learning with graph-based learning architectures in recommendation system tasks.
- Development of algorithms for evaluating large structured matrices and their max-norms.
- Construction of confidence sets in stochastic approximation problems and offline reinforcement learning, which will allow conclusions to be drawn about the reliability of machine learning algorithms.
Modelling time-independent and dependent processes and constructing their surrogates using artificial intelligence methods
Tasks
Adapting methods and approaches of hybrid artificial intelligence for building complex models of intricate systems.
Generative modelling in the task of data synthesis
Tasks
Creating a library of generative models for synthesising data of various natures (tabular data, time series, 2D/3D images).