Abstracts

Опыт Сбера в Цифровой трансформации бизнеса Клиентов в рамках службы DTaaS (Digital Transformation as a Service). DTaaS в Сбере 4 года, задача - трансформировать бизнес Клиентов с учетом опыта Сбера. За 4 года вертикаль выросла до 500 человек, а количество проектов с крупными и крупнейшими клиентами, что мы ведем превысило 7,5 тыс. Важную часть в работе с Клиентами по ЦТ занимает GenAI и др. передовые технологии и решения на их базе. Доклад посвящен подходу команды DTaaS Сбера к приоритезации трендов, объяснению как они влияют на бизнес модели и почему Клиенты должны ускорять их внедрение в свои процессы. А также основы уникальной методологии ЦТ, которую Сбер применяет в проектах с Клиентами. Структура портфеля проектов и практические кейсы.

We present a model of three Kuramoto oscillators with higher-order interactions whose strengths are described by the May-Leonard model. The observed dynamical regime of the system is deterministic chaos with sequential triadic synchronization. We establish that the system dynamics can be reduced to one-dimensional Poincaré maps for the phases and the couplings. The obtained results shed light on some of the mechanisms occurring in the neural networks of the brain, and can be used in machine learning as applications.

В докладе исследуется антропологическое измерение экспансии больших языковых моделей (LLM). Опираясь на философию, теорию внутренней речи, диалогическую философию и актуальные эмпирические исследования последних лет, авторы концептуализируют четыре зоны экзистенциального риска («ловушки»), формирующих новый антропологический ландшафт. Эпистемологическая ловушка коренится в архитектуре LLM, где связность текста систематически вытесняет истинность. Лингво-экзистенциальная ловушка связана с утратой языка как «дома бытия» и экстернализацией мышления, при которой внутренняя речь атрофируется, уступая место готовым алгоритмическим формулировкам. Феноменологическая ловушка описывает трансформацию субъективности через «семантический протез», когда внутренний опыт утрачивает легитимность без машинной обработки, порождая феномен расщепленного «Я». Социальная ловушка раскрывается через анализ взаимодействия с LLM как с «квази-Другим» — симулятором интерсубъективности, ведущим к атрофии навыков подлинного диалога и атомизации общества. В заключении формулируются принципы «антропологической безопасности» (семантическая гигиена, языковой суверенитет, непередача экзистенциальных функций) и обосновывается необходимость перехода от регулирования технологий к защите человеческой автономии.

This study presents an interpretable machine-learning approach to examining structure-function relationships in aphasia. Drawing on a large sample of more than 300 Russian-speaking patients with chronic post-stroke aphasia and dysarthria, we derived quantitative measures of gray- and white-matter damage and combined them with core demographic characteristics. Two goals guided the analysis: building and validating a reliable model for predicting both overall severity and syndrome category based on Luria's classification system, and pinpointing which neuroanatomical features contributed most to these predictions. After comparing several machine-learning algorithms and introducing a new metric — the Robust Generalization Index — we selected TabPFN as the best-performing and most generalizable model, confirming its performance on a separate validation sample (n = 41). Results showed that severity and syndrome classification behaved quite differently. Severity prediction was strong overall, especially when reduced to a mild versus non-mild distinction, with model outputs interpretably tied to time since stroke and the extent of gray- and white-matter damage. Predicting distinct aphasia syndromes was harder, yielding only moderate accuracy. Even so, this analysis offered useful theoretical insight: the patterns the model relied on aligned with central principles of Luria's framework, while also revealing how common mixed and atypical lesion profiles are in real clinical samples. Overall, this work contributes a practical computational tool for classifying aphasia status and offers new, data-driven perspective on how brain structure relates to language function after stroke — showing that structural MRI alone can yield meaningful predictions even given the complexity and variability of these disorders.

A pattern concept is a pair , where B is the maximal common description of all objects in A and A is the set of all objects sharing description B. Descriptions can be of different types, e.g. vectors of numerical intervals or sets of graphs with labeled vertices and edges. We show how this model can be used to produce short interpretable classifiers that perform on par with mainstream machine learning models in terms of standard learning metrics.

Despite considerable progress in recent years, our understanding of the fundamental principles and mechanisms that govern complex brain function and cognition remains insufficient. Network neuroscience presents a novel perspective to tackle these persistent challenges by explicitly embracing an integrative approach to investigating the structure and function of the brain. In this lecture, we will discuss advanced network analysis methodologies tailored for complex neurosystems. Specifically, we will examine network analysis for the Parkinsonian brain utilizing functional magnetic resonance imaging (fMRI) data, alongside network analysis for the epileptic brain using electroencephalography (EEG) data. In addition, we will explore the mathematical and computational paradigms of network controllability as applied to the brain, offering insights into systemic management and dynamic control of neurological states.

While coherence–incoherence patterns have been studied extensively in systems of coupled oscillators, much less is known about the generic mechanisms underlying their emergence and the nature of the associated chaotic dynamics in coupled excitable systems. In this talk, I will first present the mechanisms responsible for the onset and reveal the relationship between two important classes of symmetry-broken states—unbalanced periodic two-cluster states and solitary states—in nonlocally coupled excitable media. Using arrays of FitzHugh–Nagumo units with competing attractive and repulsive interactions as a representative example, I will show that two classes of solitary states inherit their dynamical properties from unbalanced cluster states known from globally coupled networks, while the interplay between local excitability and nonlocal interactions also gives rise to a distinct class of solitary states unrelated to unbalanced clusters. I will then introduce a new class of coherence–incoherence patterns, termed “patched patterns”, whose self-organization is characterized by the emergence of spatially continuous domains (patches) composed of units phase-locked by a 1:2 ratio of their average spiking frequencies. Depending on the balance between attractive and repulsive interactions, patched patterns exhibit periodic, quasiperiodic, or chaotic dynamics. In the chaotic regime, they can additionally develop interfaces separating neighboring patches, where units display intermediate average spiking frequencies. Finally, I will demonstrate that bumps—a characteristic class of localized patterns in coupled excitable systems—can emerge through a supercritical scenario following a localized bifurcation of Turing patterns. These results provide new insights into the mechanisms of symmetry breaking, pattern formation, and the emergence of spatiotemporal chaos in coupled excitable systems.

Existing deep neural networks for decoding brain activity prioritize performance over interpretability, failing to link the decision rule to cortical sources and the dynamic properties of their electrical activity. Conversely, traditional neuroimaging identifies neural substrates behind behavior-specific brain states but relies on oversimplified models unable to capture the complexity of brain activity variations. Our approach integrates interpretable neural networks with source-level cortical dynamics, bridging these gaps to reveal physiologically meaningful patterns that differentiate complex brain states, enabling us to build compact yet powerful decoders and mine potentially novel neurophysiological knowledge.

Neuromorphic technologies represent a key interdisciplinary trend in modern science, bridging neuroscience, materials science, and computer engineering. Unlike conventional artificial neural networks (ANNs), neuromorphic systems emulate the analog spiking dynamics and synaptic plasticity of biological neural circuits. A particularly promising hardware implementation relies on memristors — electronic devices with tunable resistive states — offering energy consumption several orders of magnitude lower than traditional digital processors. However, the biological plausibility of spiking neural networks complicates the use of standard learning algorithms like backpropagation, posing a major challenge for practical engineering. Despite this, hybrid neuromorphic architectures are actively being developed both in Russia and internationally.
A compelling application domain for neuromorphic computing is biomorphic robotics, including humanoid robots, robodogs, and bioinspired platforms mimicking animal locomotion. Recent advances in compact actuators and power sources have enabled autonomous systems capable of complex motor behaviors. In particular, imitation learning via neural networks has solved key problems of stable walking, running, and even dynamic actions such as dancing or wrestling. These robots effectively function as neuromorphic cybernetic systems, where artificial neural networks serve as control architectures embedded in a biomorphic "body."
This talk will discuss both spiking network models and the development of biomorphic robotic solutions inspired by fish swimming and bird flight, highlighting the synthesis of neural control principles with biomechanical design.