Кафедра математики

Permanent URI for this community

https://ekmair.ukma.edu.ua/handle/123456789/17501

Browse

Now showing 1 - 8 of 8

Developing a Hybrid AI model for Financial Market Prediction
(2025) Войтішин, Микита; Кузьменко, Дмитро
Over the recent years, financial time series modeling has presented a significant challenge due to market stochasticity and volatility. The stock market is influenced not only by market data such as price and volume but also by a wide range of additional external factors, including macroeconomic indicators, seasonality, fundamentals, and market sentiment. The increasing availability of diverse financial data, combined with the rapid advances in artificial intelligence (AI), has opened up new possibilities for analyzing and understanding how stock markets behave. These technologies have the potential to capture more complex nonlinear patterns that traditional statistical and machine learning models often fail to detect. This research examines how combining various model architectures and feature sets with domain - specific knowledge from the financial sector can enhance uncertainty quantification, a crucial aspect of making informed decisions and investments in financial markets.
Dropout for Neural Networks Pruning
(2025) Семенець, Дарина; Швай, Надія
In this study, the hypothesis is examined whether Dropout masks can be used for structural pruning without further evaluating the importance of individual filters or weight. It was decided to compare a Dropout-based approach, which is based on the utilization of binary Dropout masks, with a classical L2-Norm-based pruning method. For this task, we manually designed an architecture of a convolutional neural network with a custom Dropout. The research undergoes the following phases: designing a mask generation mechanism, preprocessing data, training of model, implementing of pruning algorithms, and conducting experiments using the Imagenette2 dataset. Our idea is to determine whether Dropout pruning can offer a reliable alternative to traditional methods, especially under different levels of sparsity and stochasticity.
Enhancing Temporal Smoothing in Dynamic Neural Radiance Fields
(2025) Вербицька, Марія; Кузьменко, Дмитро
In this work, we conduct an end-to-end training and fine-tuning process for the Neural Radiance Field (NeRF) model [1] and introduce 4 experimental cases with filtering techniques [2] designed to strengthen the rendering performance. We evaluate our modifications on synthetic image data of the articulated objects. For this project, we chose the architecture of the Knowledge NeRF model [3]. It includes an original PyTorch NeRF implementation [4] alongside a projection module for dynamic scenes extension. Incorporating the rendering step adjustments allows for better results without requiring complete model re-training. Our study covers the theoretical basis of the 3D scene reconstruction problem [5] alongside the NeRF architecture, such as radiance field, volume rendering, the concept of coarse and fine networks etc. [1], provides a trained and fine-tuned model for one object of a specified motion type, and suggests four methods to handle postprocessing in Knowledge NeRF better.
Image embeddings with Kolmogorov-Arnold networks
(2025) Юрченко, Артур; Кузьменко, Дмитро
This research aims to evaluate performance of Kolmogorov-Arnold networks (KAN) in image embedding tasks. It focuses on modifying existing state-of-the-art architectures - CNN and ViT, replacing their MLP segments with KANs, aiming to improve their computational performance and embedding quality. Training and evaluation methodology is fully described in sections 4 and 5.
Neural Architecture Search for Neural Decision Trees
(2025) Микитишин, Артем; Швай, Надія
Neural Decision Trees (NDTs) have recently gained attention as hybrid models combining the interpretability and structured decision-making of classical decision trees with the representational power of neural networks. Despite promising initial results in specialized applications, certain assumptions underpinning their architecture remain largely unverified. Specifically, Roy and Todorovic (2016) hypothesized that deeper nodes within NDTs would exhibit simpler architectures, requiring fewer convolutional and fully connected layers. To empirically investigate this claim, this thesis employs Neural Architecture Search (NAS) as an unbiased and automated method to explore node complexity across different tree depths, using the CIFAR-10 dataset as a benchmark. Our comprehensive experimental evaluation finds no empirical evidence to support the hypothesis that node complexity systematically decreases with increasing tree depth. These results suggest that assumptions derived from specific applications, such as monocular depth estimation, may not generalize to other domains, underscoring the importance of empirical validation and careful search-space design in neural decision tree research.
Photo-realistic image restoration algorithms
(2025) Засядько, Матвiй; Крюкова, Галина
In this work, a new algorithm to reconstruct the facial images from degraded inputs is proposed with the visual high-definition reconstruction as its goal. The approach utilizes edge map information in a generative adversarial network (GAN) framework to be able to restore more delicate local structures and semantic content. The architecture is consisting of three parts: a DeblurEncoder which takes a blurred face image and its corresponding edge map, a Generator which recovers high resolution, and a Latent Encoder which supervises in latent space using the consistency loss terms. Training is performed end-toend all the while using a combined loss function that includes L1 loss, LPIPS perceptual loss, SSIM-based structural similarity loss, total variation loss, and a latent alignment term. Our approach was evaluated on the CelebABlur dataset and achieved comparable results in terms of numerical evaluation and visual quality. The study also compares with some recent state-of-the-art methods such as StyleGAN-based latent optimization, Posterior-Mean Rectified Flow and DiffIR. An advantages of this method are the combination of edgeinformation and latent-space constraints, which results in the improved quality of generated images, and that all three model components are trained simultaneously, what provides more consistent learning across the latent and pixel spaces enhancing both visual fidelity and structural coherence.
Метаевристичнi алгоритми для обрiзки нейронних мереж
(2025) Котляренко, Анастасiя; Швай, Надія
This work studies neural network pruning with metaheuristic optimization methods. Pruning was formulated as an optimization problem with a target function that is a weighted sum of neural network accuracy and sparsity. This problem was solved with stochastic metaheuristic methods (Genetic Algorithm and Particle Swarm Optimization) that generate binary masks. Obtained results demonstrate that pruning with metaheuristic methods is comparative with 𝐿2 pruning when finetuning is possible and is significantly more performant when no post-pruning finetuning is available.
Сooperation of partially observed agents in ad-hoc open teams
(2025) Вiнокур, Євгенiй; Кузьменко, Дмитро
The aim of research: Systematic comparison of eight decentralized training baselines. We are inspired by the research of , where authors tested choosing best clearing action with Deep Learning on fire spread simulation. However, authors provide limited choice of algorithms with limited metrics and encounter non-stationairty issues due to common reward. Focus of our research is evaluation through extensive benchmarking of Independent, value-decomposition, central-critic, and agent-modeling methods proposed by Papoudakis et. al evaluated under common hardware/runtime constraints. Our work considers constraints of partial observability, generalization and mixed teams. Results promote insights on beneficiary features of baselines to assist further researches in selecting or developing effective algorithms for decen- tralized planning and control. Our contribution transfers Wildfire benchmark, created by Tran Research Group to PettingZoo library to promote verification of our results.

Browse

Browsing Кафедра математики by Subject "bachelor`s thesis"

Results Per Page

Sort Options