Vadym Mukhin

The subject of this research is dependency parsing of Ukrainian-language texts within the Universal Dependencies framework, specifically the improvement of accuracy by combining the outputs of independently trained parsers. The purpose of the work is to develop a two-stage voting algorithm that extends prediction-time parser combination from unlabeled attachment to labeled dependencies, specifying the order in which the head and label decisions are taken. The objectives are to evaluate such ensembles against individual parsers, to compare homogeneous and heterogeneous ensembles, and to quantify the prevalence of structurally invalid trees. The methods include training 200 transition-based and graph-based parsers with deep contextualized word representations, exhaustive enumeration of three-model combinations, evaluation in terms of UAS and LAS, and significance testing with McNemar’s test. The results show that ensembles outperform individual parsers at p<0,001, the best configuration improving UAS by 1.10% and LAS by 1.26%. Heterogeneous ensembles show higher mean values than homogeneous ones, but this advantage is not statistically significant. Unconstrained voting yields between 2.15% and 3.44% structurally invalid trees, below half of the lowest rate previously reported for ensembles without contextualized representations. The scientific novelty is that a two-stage voting algorithm has been developed for the first time in which the label is chosen only among parsers agreeing on the head; the measurement of structural validity in parser ensembles has been improved; and the question of residual architectural diversity has been further developed. The practical significance is that the algorithm operates on finished parses and requires no retraining.

The article describes a model of cloud-native AI pipelines designed for continuous optimization of computing infrastructure and real-time anomaly detection. The developed model combines modern approaches to observability, machine learning (ML), and auto-scaling based on load forecasting. The methodology is based on the use of LSTM models, autoencoders, and convolutional neural networks (CNN) integrated into Kubernetes environment with support for Prometheus, Kafka, and Grafana. Load changes are simulated, and the system's response to critical events is evaluated. The results demonstrate a significant improvement in anomaly detection accuracy (up to 93%) and resource efficiency (up to 26% cost reduction compared to traditional approaches). The proposed model can be used in AIOps systems that require a high level of automation and reliability.

Efficient resource distribution in heterogeneous distributed computing systems requires intelligent node selection mechanisms capable of adapting to dynamic system conditions. This research presents a comparative evaluation of Convolutional Neural Network (CNN) and Multilayer Perceptron (MLP) architectures for binary node suitability classification in distributed task scheduling environments. The study employs five synthetic datasets ranging from 100 to 2000 instances, with each node characterized by four critical attributes: performance, security level, baud rate, and reliability. Experimental results demonstrate that the MLP architecture achieves validation accuracy between 91% and 98.8% with high consistency across dataset sizes, while the CNN architecture shows fluctuating performance ranging from 85% to 94%. The key finding reveals that MLP architectures outperform CNNs for tabular node data due to better alignment with unstructured attribute relationships, as fully connected layers naturally handle unordered features without imposing spatial locality assumptions. This empirical analysis provides practical guidance for architecture selection in AI-based resource schedulers, demonstrating that simpler architectures can outperform more complex ones when appropriately matched to problem structure. The findings contribute evidence-based recommendations for distributed systems practitioners implementing neural network-based scheduling solutions.

The article discusses the solution to the issue of information systems architecture, which makes it possible to separate identification flows associated with the processes of managing the identification of digital entities and data flows associated with the processes of managing data describing the properties of digital entities. At the same time, a new architecture of connections between digital entities and services is proposed, which makes it possible to create a flexible system for processing the properties of identifiers that can be described in an irregular and unstructured form. In this case, the nomenclature of the parameters of the properties of a digital entity can be customized and expanded as a separate entity, which always maintaining a connection with the identifier. This allows, within the framework of the information system, on the one hand, to adapt any identification systems and ensure the solution of convergence requirements, to ensure compliance and solution of the requirements of recognition, immutability, stability in conditions of processing large volumes of data without implementing generally accepted principles of a global unique permanent identifier, and on the other hand, build flexible connections between digital entities and services.

Subject matter: this research focuses on the use of multimodal transformers for high-quality sound synthesis. By integrating heterogeneous data sources such as audio, text, images, and video, it aims to address the inherent challenges of accurate modality alignment. Goal: the primary goal is to conduct a comprehensive analysis of various modality alignment algorithms in order to assess their effectiveness, computational efficiency, and practical applicability in sound synthesis tasks. Tasks: the core tasks include investigating feature projection, contrastive learning, cross-attention mechanisms, and dynamic time warping for modality alignment; evaluating alignment accuracy, computational overhead, and robustness under diverse operational conditions; and benchmarking performance using standardized datasets and metrics such as Cross-Modal Retrieval Accuracy (CMRA), Mean Reciprocal Rank (MRR), and Normalized Discounted Cumulative Gain (NDCG). Methods: the study adopts both quantitative and qualitative approaches. Quantitative methods entail empirical evaluations of alignment precision and computational cost, whereas qualitative analysis focuses on the perceptual quality of synthesized audio. Standardized data preprocessing and evaluation protocols ensure reliability and reproducibility of the findings. Results: the analysis reveals that contrastive learning and cross-attention mechanisms achieve high alignment precision but demand considerable computational resources. Feature projection and dynamic time warping offer greater efficiency at the expense of some fine-grained detail. Hybrid approaches, combining the strengths of these methods, show potential for balanced performance across varied use cases. Conclusions: this research deepens understanding of how multimodal transformers can advance robust and efficient sound synthesis. By clarifying the benefits and limitations of each alignment strategy, it provides a foundation for developing adaptive systems that tailor alignment methods to specific data characteristics. Future work could extend these insights by exploring real-time applications and broadening the range of input modalities.

This work is aimed at developing intelligent systems capable of automatically classifying types of educational materials. This will allow students to find the resources they need faster, and it will make it easier for teachers to manage content in educational platforms. The solution of the problem of recognition of information objects using fuzzy output systems and neural networks is considered. This approach combines the advantages of neural machine learning with the flexibility and efficiency of fuzzy logic, making these systems effective tools for solving problems related to fuzzy or uncertain data. An information model of a neural network for classifying information objects in e-learning systems has been developed. Experimental testing of the proposed approach was carried out on a data set, which consists of information objects from real e-learning systems, namely such as manuals, lectures, syllabuses, and textbooks. Design an adaptive mechanism that utilizes fuzzy neural networks to personalize content recommendations for students based on their learning progress and preferences, thereby enhancing the effectiveness of e-learning systems. Improve the efficiency of classification algorithms by fine-tuning neural network parameters and fuzzy logic rules, ensuring high accuracy and computational efficiency when processing large-scale educational datasets. The results of experimental studies have shown that a neural network built based on fuzzy logic is able to classify various information objects efficiently and correctly in e-learning systems. It is shown that the integration of neural networks based on fuzzy logic into e-learning systems to improve the processes of classification of information objects makes it possible to increase the efficiency of educational resources management, ensuring accuracy and flexibility in processing various data.

This paper examines the network-centric approach for enhancing the efficiency of mobile agent control. The primary distinction of this approach is the use of a "wandering center." If communication with the active center is lost, a new center is appointed to continue control, ensuring the system's survivability and improving the overall efficiency of mobile agent management. An experimental study was conducted to evaluate the parameters and effectiveness of network-centric control of mobile agents. The study utilized a developed model of mobile agent interaction based on network- centric control, along with a simulation model of agent behavior in a dynamic environment. Additionally, a neural network was created that accurately predicts the probability of hitting the target (%) under changing dynamics. The use of this neural network also helped identify input parameters with minimal impact on the outcome.

This work is aimed at developing intelligent systems capable of automatically classifying types of educational materials. This will allow students to find the resources they need faster, and it will make it easier for teachers to manage content in educational platforms. The solution of the problem of recognition of information objects using fuzzy output systems and neural networks is considered. This approach combines the advantages of neural machine learning with the flexibility and efficiency of fuzzy logic, making these systems effective tools for solving problems related to fuzzy or uncertain data. An information model of a neural network for classifying information objects in e-learning systems has been developed. Experimental testing of the proposed approach was carried out on a data set, which consists of information objects from real e-learning systems, namely such as manuals, lectures, syllabuses, and textbooks. The results of experimental studies have shown that a neural network built based on fuzzy logic is able to classify various information objects efficiently and correctly in e-learning systems. It is shown that the integration of neural networks based on fuzzy logic into e-learning systems to improve the processes of classification of information objects makes it possible to increase the efficiency of educational resources management, ensuring accuracy and flexibility in processing various data.

The study solves the problem of improving the methodological and algorithmic support of the decisionmaking process by developing a model of the preference criterion for interval evaluations of alternatives. The aim of the study is to improve the efficiency of decision-making based on interval expert data under conditions of uncertainty and risk by developing a criterion for the preferences of interval evaluations of overlapping alternatives. The object of the study is the decision-making process based on the classical efficiency matrix with interval elements, the subject is the

In recent times, there has been significant progress in the application of deep learning, particularly using convolutional neural networks (CNNs), to address image denoising tasks, driven by their exceptional performance in image processing. However, it is noteworthy that CNNs are predominantly used for dealing with Gaussian noise, and there is currently limited work on effectively reducing salt-and-pepper (SAP) noise. One of the strategies to tackle the issue of noisy images is the development of efficient deep learning models specialized in handling SAP noise. This paper explores the potential enhancement of the proposed SeConvNet model, designed specifically for reducing noisy images contaminated with SAP noise. Salt-and-pepper noise, characterized by random occurrence of black and white pixels, is a common type of noise found in images. Given the current relevance of noise reduction in images, particularly in the absence of sufficient models dedicated to SAP noise, this paper introduces a block aimed at potentially improving the performance of the existing model. The results of this study demonstrate the promise of advancing the direction of enhancing the original model, which in turn could be beneficial for a wide range of applications, including medical diagnostics and any domain where image processing is crucial for precise outcomes. Implementing the proposed improvements could have a positive impact on processing images with SAP noise, considering the scarcity of models addressing this specific problem. The model was trained on the well-known BSD68 dataset, ensuring an objective evaluation of the results. Additionally, the paper presents an analysis of existing models that target SAP noise reduction, providing insights into the current landscape of techniques in this domain.