Факультет інформатики

Permanent URI for this community

https://ekmair.ukma.edu.ua/handle/123456789/17507

Browse

Now showing 1 - 20 of 309

3D реконструкція сцени за відео з декількох камер
(2020) Томащук, Вадим; Крюкова, Галина
Робота складається з чотирьох розділів. В першому розглянуто базові поняття виявлення та опису особливостей об’єктів, які є основою для будь якого обраного підходу 3D реконструкції сцени, а також розглянуто найпопулярніші алгоритми для кращого сприйняття цього етапу в процесі розробки. В другому розділі описана теоретична база епіполярної геометрії, вирівнювання та триангуляції, а також згадано про алгоритми відстежування та оцінки позиції об’єкта. Третій розділ повністю присвячений поетапній розробці практичного застосування та частково описується необхідна теорія, така як побудова карт невідповідностей та глибини. В четвертому розділі звернено увагу на недоліки проведеної роботи, оцінено ефективність практичного застосування та простір для подальшого вдосконалення.
Adversarial robustness and attacks in Deep Learning
(2022) Кузьменко, Дмитро; Швай, Надія
The theoretical underpinnings for this field involve the notions of robustness and astuteness, local Lipschitzness, r-separability of datasets, robustness-accuracy tradeoff, and L-inf distance. This work will cover all the preliminaries, explain the choice of CIFAR-10 with L-inf metric space and eps=8/255 as a main dataset for the task, make use of already well-known attacks and defenses, introduce new ones, and try different ensembles on the 3 most robust models available on the benchmark – Adversarial Weight Perturbation, Augmentations and weight averaging, and Self-COnsistent Robust Error (SCORE-based model).
Analysis of Curriculum Learning methods in Reinforcement Learning
(2024) Orel, Danyil; Glybovets, Mykola
The aim of this work is to provide a comprehensive comparison of CL methods in RL across various scenarios and benchmark environments.
Continual Learning Method for image classification in computer vision
(2024) Kreshchenko, Taras; Yushchenko, Yurii
The paper explores the hypothesis that Continual Learning (CL) methods can improve the performance of a deep learning model in a traditional machine learning scenario. By augmenting an existing state-of-the-art ML solution to a problem with CL techniques, this research aims to demonstrate that AI can still achieve more accurate and adaptive performance. This hypothesis is tested on a parking lot occupancy detection problem, a binary classification problem that is well-suited to CL due to the continuous stream of image data. Experiments are conducted to compare the proposed CL-based solution and a contemporary solution that is non-CL based.
Control in Gordon-Newell Networks
(2021) Степанюк, Роман; Чорней, Руслан
As more and more companies are restructuring their business models to be more automized, the question of efficient usage of available capabilities arises. With the help of queuing networks, the only thing the customer has to do to get helpful information regarding the improvement of queueing processes in their company is to input the data that describes the system. As a result, they can be provided numbers, charts, and other data that will drastically improve the working process. By implying strategy improvement procedure on local decision-makers, optimal local strategy can be derived, improving the working process flow in the network.
Cемантична сегментація зображень з використанням Transformer архітектури
(2022) Іванюк-Скульський, Богдан; Швай, Надія
In this work we have presented a model that efficiently balances between local representations obtained by convolution blocks and a global representations obtained by transformer blocks. Proposed model outperforms, previously, standard decoder architecture DeepLabV3 by at least 1% Jaccard index with smaller number of parameters. In the best case this improvement is of 7%. As part of our future work we plan to experiment with (1) MS COCO dataset pretraining (2) hyperparameters search.
Development of a methodology for using microservice architecture in the construction of information systems
(2021) Zhylenko, Oleksii; Cherkasov, Dmytro
In this work will be defined what is microservice architecture, the most important quality attributes and system level requirements. We will gather guidelines grouped by quality attributes that should be used to reduce in future total cost of ownership system under develop. Created methodology will be used in synthetic Java project to demonstrate it on real example.
Development of CI / CD platform deployment automation module for group software development
(2021) Ivanov, O; Cherkasov, Dmytro
In the presented work we reviewed the main CI/CD principles and delivery workflow. We provided definition and benefits for each of part of the CI/CD. Latter we covered definition of CI and CD. Provided tools analysis and narrowed audience for expected module. We have chosen the platform base and picked up clouds for developing solution. After that we developed modules for automatics deployment into cloud as easiest for user as possible and created all necessary scripts with ability to integration in bigger system or launching out of the box. In the end test runs were done for every script and compared result and difficult of managing. Based on achieved results we provided the feedback.
Development of the system for plagiarism checking of Ukrainian texts
(2022) Bikchentaev, Mykola; Hlybovets, Andrii
So, the aim of this work is to review two machine learning models called BERT and Word2Vec, determine how can they be used in plagiarism detection, and develop an application where users can check texts for plagiarism.
Distributed system technical audit
(2020) Zhylenko, Oleksii; Hlybovets, Andrii
In this coursework will be defined what is distributed systems, review Monolithic, Microservice and serverless architecture. Also, we will deep dive into technical audit process, specify what aspects of system must be considered during audit. Then will iterate over checklists item in order to provide guidelines based on best practices in industry that helps to prepare for system audit.
Emitter Position Estimation Using Time Difference of Arrival
(2022) Musiiaka, O.; Hlybovets, Andrii
As a result of this work, the estimator convergence speed was improved by using a local linear transform. This new approach can be considered an adaptation of the Newton-Raphson algorithm with the Hessian matrix replaced with a statistically approximated value that guarantees the algorithm convergence and reduces the amount of computation. Performance testing of the optimization algorithms has shown that the proposed algorithm outperforms the steepest gradient descent 28.2 times and “momentum” modification 7.8 times on the CPU implementation.
Euclidean Algorithm for Sound Generation
(2021) Laiko, Artem; Horokhovskyi, Semen
This course work aims to research possible ways of Euclidean algorithm application and influence for sound generation process, as well as mathematical basis of sound and sound waves. The practice part applies obtained knowledge to develop a VST3 plug-in for sound wave morphing with the Euclidean algorithm application.
Evolutionary art generation using genetic algorithms
(2021) Moroz, Andrii; Horokhovskyi, Semen
This course work explains the concept of generative art and especially concentrates on evolutionary art. Evolutionary art is a part of generative art that uses genetic programming to produce art. Genetic programming for art generation works by creating an initial population, picks out ones that do not fit, and forms new ones combining images that went through selection. Also, this course work is going to explain how to generate images using genetic programming and show results from the developed program.
Impact of adversarial sparsity as an auxiliary metric in adversarial robustness
(2023) Кузьменко, Дмитро; Швай, Надія
The purpose of this research is to investigate adversarial sparsity in computer vision models and introduce a more efficient method for adversarial sparsity estimation. To fulfil this objective, the following tasks have been undertaken: To implement and evaluate an n-Ary search algorithm as an improvement over the conventional binary search method used in adversarial sparsity estimation. To benchmark and compare the performance of the proposed n-Ary search algorithm against the traditional binary search algorithm. To explore the implications of adversarial sparsity on the robustness of machine learning models.
Investigation of the relationship between software metrics measurements and its maintainability degree
(2020) Shapoval, Oleksandr; Hlybovets, Andrii
The goal of this thesis was to practically learn methods of empirical engineering software, algorithms for data collection and data analysis. Results include software measurement, analysis and selection of direct and indirect metrics for research and identification of dependencies between direct and indirect metrics. On the basis of received result were built dependencies between software metrics and software expertise properties. Metrics and properties selected by individual variation. Relationship between metric and expertise includes building direct relationships between the metric and expertise, indirect metrics and expertise. Additionally, was determined whether they have common trends of the relationship between those direct metrics and expert estimates, indirect metrics and expert estimates.
Iтерацiйний пiдхiд до необумовленого оптимального вибору для певної категорiї в роздрiбнiй торгiвлi
(2024) Мироненко, Роман; Дрiнь, Свiтлана
У данiй роботi ми розглядаємо сучаснi методи оптимiзацiї попиту для групи товарiв та дослiджуємо iтерацiйний пiдхiд для знаходження оптимальної кiлькостi товарiв у певної заданої категорiї.
Language model optimization using pruning, distillation and quantization techniques for NLP tasks
(2024) Petrenko, Mykhailo; Marchenko, Oleksandr
The dominant approaches to quantizing neural net- works with billions of parameters focus primarily on weight quantization due to accuracy considerations. However, activation quantization remains a significant bottleneck for inference speed. Building upon the foundational research of GPTQ and Qual- comm, we propose GPTAQ, a novel framework that introduces activation quantization for large language models (LLMs) while attempting to balance out activation-induced error with the following enhancements: Eigenvalues of the Hessian sensitivity matrix, although our experiments reveal this approach yields mixed results. Cross-Layer Equalization (CLE), which balances weight scales across layers to prevent channel suppression. Bias Correction, to correct the effects of CLE. We demonstrate the effects of our approach through exper- iments on the Facebook OPT model using the C4 dataset for calibration. Our results show that RTN and Token-wise activa- tion quantization combined with CLE achieve the best trade- off between model efficiency and accuracy. GPTAQ introduces activation quantization while maintaining low perplexity scores, indicating minimal performance degradation given the limited experimental setup. Our framework offers a comprehensive solution for effective activation quantization, enhancing the deployment efficiency of large language models and providing valuable insights for future research, such as further Hessian Eigenvalues tuning to decrease introduced error, expand and switch calibration dataset, and remaining ablation study.
Modeling Distributed Generalized Suffix Trees For Quick Data Access
(2022) Діденко, Віра; Глибовець, Андрій
The aim of this work is to distribute generalized suffix tree construction, so the process is efficient in terms of time complexity and memory consumption. A distributed approach to constructing the suffix tree will allow working with large alphabets and very long strings that exceed the available memory capacity. In this work, an efficient and highly scalable algorithm for constructing generalized suffix trees on distributed parallel platforms was modeled. The experimental results proved that the modeled algorithm’s efficiency is no less than the before known Elastic Range algorithm (ERa) while out-performing ERa on specific data.
A multicriteria competitive Markov decision process
(2021) Левченко, Іларія; Чорней, Руслан
The course work is devoted to A multicriteria competitive Markov decision process; proposed software implementation of their solution. The work consists of an introduction, the main part that consists of six sections, a conclusion, a list of used sources and an appendix. Relevance. The modern-day world makes people face more and more complicated problems which require a solution and the price of mistake for them can be really high. Besides that, nowadays there is so much data that making a decision based on that intuitively and without analysis and math is not an option anymore. The multicriteria Markov decision process is much more similar to reallife than some other common games and decision models – choosing one of the available actions without knowing action chosen by the opponent as well as having vector reward rather than single reward are both much more common in a real application. However, solving such problems as they are is complicated. Therefore in this paper considered algorithm to transform them into linear programming problems, which have more well-known solution algorithms. The object of the study is a multicriteria Markov decision process. The subject of the study is an algorithm for solving the multicriteria competitive β-discounted Markov decision model. Purpose to study multicriteria competitive Markov decision games and algorithm to solve them. Theoretical research methods were used in the study; information from various scientific sources is analyzed, compared and summarized.
Object feature extraction for YOLO detectors
(2023) Абашкін, Олександр; Швай, Надія
The main goal of the research: To create an architecture that can surpass in quality and speed the solutions of that time such as the deformable part models (DPM) that were using the sliding window approach where the classifier is used for each evenly spaced location, and a the R-CNN that were using a network for generation potential bounding boxes and as a second stage applies a classifier on this regions.

Browse

Browsing Факультет інформатики by Title

Results Per Page

Sort Options