Scientific Expedition to Hainan: HSE Scientists Organise Conference on Statistical AI in China

© HSE University

The Statistical AI Conference was held in Sanya, Hainan Island, China, from 24 to 28 August 2026. The international event brought together leading experts in statistics, machine learning, and applied AI. Alexey Naumov, Director of the AI and Digital Science Institute at the HSE Faculty of Computer Science, and Sergey Samsonov, Head of the International Laboratory of Stochastic Algorithms and High-Dimensional Inference, were among the conference organisers.

The conference provided a platform for discussing key issues in statistical theory, which underpins the reliability of modern AI, including how to quantify uncertainty, make decisions under uncertain conditions, and develop theoretically sound algorithms capable of scaling to large volumes of data.

Alexey Naumov

‘Organising this conference was entirely our initiative, from submitting the application and developing the programme to inviting members of the international community. We managed to bring together researchers from eight countries—Russia, China, Singapore, Hong Kong, Japan, the United Arab Emirates, France, and the United Kingdom—which confirms the relevance of the topic and our colleagues’ willingness to maintain direct scientific contacts,’ said Alexey Naumov.

Sergey Samsonov

‘I believe we have been quite successful in building a dialogue with colleagues from the world’s leading research centres. We are already actively conducting joint research with some of them, and the conference provided an excellent platform for exchanging ideas and discussing future academic partnerships,’ noted Sergey Samsonov.

The scientific programme of the conference covered the full range of contemporary research at the intersection of statistics and artificial intelligence. Presentations by researchers from different countries set the agenda for discussions, spanning topics from fundamental questions about the generalisation capabilities of neural networks to applied methods in reinforcement learning and generative modelling. Kenji Fukumizu (Japan) spoke about the unexpected connection between generative models and proximal operators, while Gen Li (Hong Kong) discussed the theoretical foundations of diffusion generative models. Two presentations deserve particular mention: Alex Lamb (China) spoke about compact internal world models in transformers, and Yunbei Xu (Singapore) presented his theory of pointwise generalisation, which received an award at the ICML conference. At the end of the programme, Junpei Komiyama (UAE) presented a method for statistical certification of voting, while Lei Wu (China) discussed functional scaling laws for LLM pre-training.

Vladimir Spokoiny

Vladimir Spokoiny, Academic Supervisor of the Laboratory for Theoretical Foundations of AI Models, presented a new perspective on Gaussian Mixture Models through the prism of manifolds in his talk, ‘Manifold Gaussian Mixture Models’. ‘Our approach addresses a classic problem faced by data analysts: how to work with distributions in very high-dimensional spaces, where traditional methods, such as standard Gaussian mixtures, either break down or require enormous computational resources. We propose using the concept of a manifold—the assumption that real-world data does not lie chaotically across an entire vast feature space but is instead concentrated on a certain low-dimensional manifold,’ said Spokoiny. ‘Our numerical experiments show that the method achieves state-of-the-art performance on real-world tasks, while our theoretical guarantees confirm that component identification is optimal in terms of effective dimensionality. This means that we have not just a heuristic but a mathematically sound tool for working with the most complex datasets encountered in modern business and science.’

Nikita Puchkin

Nikita Puchkin, Head of the Laboratory for Theoretical Foundations of AI Models, spoke in his presentation, ‘Stochastic Optimal Control Approach to Generative Modelling and Schrödinger Potential Estimation,’ about making generative models reliably accurate even when only limited training data is available. ‘We work with generative models based on stochastic processes. To obtain the desired data distribution, we often use what is known as a Schrödinger bridge—an elegant way of connecting two distributions. But there is a catch: the control signal, or potential, depends on the target distribution, which we do not know in advance. We only have a sample at our disposal, and any error in estimating the potential can throw the model off course. We focused on estimating the Schrödinger potential by minimising the empirical KL divergence and showed that, under very mild conditions that hold for a wide class of real-world distributions, it behaves well and has excellent geometric properties. This allowed us to derive non-asymptotic upper bounds on the estimation error. In other words, we can now quantify how much our estimate may deviate from the true value,’ Nikita Puchkin explained.

© HSE University

The conference also featured presentations by researchers from the AI and Digital Science Institute at the HSE Faculty of Computer Science:

 Maxim Rakhuba, Head of the Laboratory for Matrix and Tensor Methods in Machine Learning, presented ‘Linear Algebra Problems and Algorithms in Muon-Based Optimisation Methods’;

 Marina Sheshukova, Junior Research Fellow at the International Laboratory of Stochastic Algorithms and High-Dimensional Inference, presented ‘Gaussian Approximation and Bootstrap for SGD’; 

 Konstantin Yakovlev, Junior Research Fellow at the Laboratory for Theoretical Foundations of AI Models, presented 'Generalisation Bounds for Diffusion Models';

 Denis Belomestny, Leading Research Fellow at the International Laboratory of Stochastic Algorithms and High-Dimensional Inference, presented ‘Statistical Analysis of Inverse Entropy-Regulated RL.’

As part of the workshop, a separate session was devoted entirely to presentations by early-career researchers. The organisers highlighted five presentations by doctoral students and postdoctoral researchers from leading institutions, including Peking University, the National University of Singapore, and HSE University. Nikita Morozov, Junior Research Fellow at the Centre for Deep Learning and Bayesian Methods, and Ilya Levin, Junior Research Fellow at the International Laboratory of Stochastic Algorithms and High-Dimensional Inference, presented the results of their research, covering topics ranging from theoretical aspects of optimisation to applied problems.

The workshop marked an important milestone in HSE University’s collaboration with leading research institutions around the world, including Tsinghua University, SLAI (Shenzhen), CUHK (Shenzhen), the National University of Singapore (NUS), and MBZUAI. 

‘The experience has once again convinced us that the optimal format for a conference is a small one, with 30–50 participants. This number allows us to go beyond formal presentations: there is room for in-depth discussions, an opportunity to engage meaningfully with each presentation, and time to establish direct professional contacts, which often prove more valuable than routine conference sessions. Inspired by the results, we have already outlined the next step: we are considering holding a similar event in Hong Kong or Singapore to expand our geographical reach and attract new partners,’ Alexey Naumov emphasised.