Диссертации, представленные на защиту и подготовленные в НИУ ВШЭ

Implicit Neural Representations (INRs) recently emerged as a powerful and compact 3D representation, well fitted for the needs of differential rendering. On the other hand, basic INRs and algorithms for their optimization lack the number of qualities especially important for the tasks of multi-view reconstruction and 3D generation. These qualities include the inability of INRs to represent a large variety of objects simultaneously and excessive rendering time. Also, when combined with the generation approaches, INRs often fail to produce results which are multi-view consistent, or consistent with each other, in the case of the generation of multiple instances. In order to overcome the described drawback, a number of methods is developed and described in the current thesis: a method for training hypernetwork INR in the role of Generative Adversarial Network (GAN) generator; a method for obtaining aligned 3D models, parameterized by a single INR, given a set of text prompts; an algorithm for fast rendering and reconstruction of the implicit surface; an approach for completing the surface, which is observed only partially; a GAN-based image generation method for unsupervised shape/appearance disentanglement, where deformation maps and textures are produced by separate INR generators. Developed methods widen the scope of the application for the Implicit Neural Representations, allowing to model complex structures of the significant variety. They also allow to manipulate existing 3D objects, edit them and complete the missing parts, producing the assets which could be used in computer graphics applications.

This dissertation is devoted to the development and application of geometric and topological methods for analyzing, interpreting, evaluating, and improving deep learning models. The work investigates geometric and topological properties of internal data representations (embeddings) in various Deep Neural Network (DNN) architectures, including convolutional DNNs and Transformers under different training modes (e.g. supervised, self-supervised learning etc). A method for evaluating model performance based on geometric properties of intermediate representations is proposed. The research develops an approach for improving model performance through integration of domain-specific topological data information into embeddings and proposes a modification of the RNN architecture called TonnetzNet, demonstrating performance enhancement. The work also investigates the applicability of the fractal intrinsic dimension of Transformer embedding data manifold for robust detection of content generated by large language models (LLMs) and presents a method for synthetic image recognition. Finally, the capabilities and limitations of various language models are evaluated in tasks of modeling symbolic sequences representing elements from normal closures in free groups. The dissertation provides both theoretical insights and empirical validation of the developed methods. The research results contribute to the fields of AI safety, mechanistic interpretability, and explainable AI. Overall, the work emphasizes the promise and importance of geometric and topological methods not only as analytical tools, but also as a foundation for developing approaches to interpretation, evaluation, application, and improvement of DNNs.

The Large Hadron Collider (LHC), which was built by the European Organization for Nuclear Research (CERN), is the world’s largest collider. The LHCb experiment at the LHC focuses on studying heavy flavor physics, making precise measurements of CP violation, and investigating other effects within and beyond the Standard Model.The LHCb detector consists of several components, including an electromagnetic calorimeter (ECAL). Simulating the expected detector response is crucial for the physical analysis of the collected data and deriving physical results. Using theGeant4 package to simulate detector responses is computationally expensive and resource intensive, so there is a need to speed up this process. This research paper explores the potential of using generative adversarial networks (GANs) to accelerate the simulation process of calorimeter response in high-energy physics experiments. The thesis contribution consists of three essential parts. The model’s performance is highly sensitive to its architecture, so it was improved comparing to the previously published once. The second contribution is a method that helps the model to take particular properties of generated objects into account and improve its generation quality. This method requires increasing models capacity, so an other approach that allows to balance between training stability and expressivity is proposed. By proposing methods to enhance simulation speed and improve calorimeter response accuracy, this work holds significant implications for the advancement ofLHCb and other high-energy physics projects.

This research focuses on developing methods for converting raster images and three-dimensional objects into vector representations using deep learning. Vectorization of objects involves finding object representations using mathematical primitives and relationships between them.To achieve this goal, the following tasks were addressed: data collection, construction of mathematical models, and development of vectorization algorithms. Data collection was performed by processing scanned images of 2D and 3D objects and generating synthetic data. Automatic algorithms using computer vision methods were developed for data cleaning and processing, along with procedures for manual data processing. These algorithms facilitate semi-automatic annotation of data, opening up the possibility to train neural networks using deep learning methods. Various neural network architectures, including convolutional neural networks and transformers, are explored to create models capable of accurately and efficiently vectorizing technical drawings and 3D point clouds. The proposed algorithms demonstrate high accuracy and efficiency in solving object vectorization tasks, with potential applications in computer vision, robotics, and data visualization.

Nowadays, recommender systems are essential components of various consumer services, from e-commerce to social media. They help navigate through a large volume of items empowering user experience. Methods vary from classic matrix completion techniques to modern sequence models inspired by natural language processing. One of the prominent approaches is to consider the recommender system as a link prediction problem on a bipartite user-item interaction graph.The dissertation studies the adaptation of network embedding techniques to recommender systems. In the dissertation, we investigated different aspects of graph machine learning techniques, and their effect on downstream link prediction problem. We proposed an efficient strategy to incorporate node and edge features with structural information to solve the link prediction (recommendation) problem. Also, we developed novel models to preserve temporality and enable graph-based exploration for recommender systems. Proposed approaches were compared with the state-of-the-art open-source benchmarks and showed the efficiency of our approach.