This paper introduces a unified theoretical perspective that views deep generative models as probability transformation functions. Despite the apparent differences in architecture and training methodologies among various types of generative models - autoencoders, autoregressive models, generative adversarial networks, normalizing flows, diffusion models, and flow matching - we demonstrate that they all fundamentally operate by transforming simple predefined distributions into complex target data distributions. This unifying perspective facilitates the transfer of methodological improvements between model architectures and provides a foundation for developing universal theoretical approaches, potentially leading to more efficient and effective generative modeling techniques.
In recent decades, numerous algorithms have been proposed for extracting tables from PDF documents, with most designed as general-purpose table extractors. This paper introduces a novel algorithm specifically tailored for extracting order tables, addressing specific issues not covered by existing general-purpose extractors. Our approach, named OTE, identifies leading rows in an order table through a clustering algorithm and employs heuristics to recognize additional row lines and annotate columns. Through an evaluation of 115 order documents from customers of a medium-sized company, we demonstrate that: i) OTE surpasses general-purpose extractors, ii) accurately identifies over 95
Effective route planning, despite numerous studies, is still a great challenge. Given real-world data and demanding criteria, the problem can be even more complex, with no single optimal solution. In this paper, an interactive system with dedicated Dijkstra-based algorithm is proposed to tackle the problem of efficient bicycle route planning. With the presented web application framework, the user is able to generate a complete cycle of the desired length and surface quality, complementing the selected profile of the bicycle. The results indicate the reliability of the system. The routes adapt well to the chosen profile and fall within the provided acceptable total-length error (5
Innovative educational technologies, the integration of Massive Open Online Courses methodology (MOOC), Challenge-Based Learning (CBL), and Virtual Assistant methodologies in Big Data course represent a dynamic evolution in pedagogical approaches. MOOCs offer scalable access to high-quality educational content, enabling learners to engage with Big Data concepts flexibly. CBL fosters critical thinking and problem-solving skills by immersing students in real-world scenarios relevant to Big Data analysis. Virtual Assistant methodologies leverage artificial intelligence to provide personalized learning experiences, enhancing student support and interactivity. This integrated approach not only cultivates a comprehensive understanding of Big Data but also prepares learners for the demands of a data-driven world. The authors are discussing the methodology and the effectiveness of the implemented course.
This article discusses the importance of effective risk management in Agile projects. The Agile project management philosophy emphasizes open communication, flexibility in product development, and rapid market delivery. Unlike traditional waterfall methods, Agile iterative approach reduces risks by allowing on-going progress assessment. However, Agile alone may not effectively manage risks, especially those stemming from human factors and environmental influences. Existing project management tools lack comprehensive risk identification and management solutions, particularly concerning diverse project components and stakeholder engagement. Decentralization in Agile projects can lead to architectural risks due to documentation inconsistencies and client disengagement. To address these challenges, this article proposes a methodology for identifying risks of Agile project associated with artifacts and attributes of it. By regularly updating attributes' data, potential risks can be unveiled and addressed early, enhancing project team effectiveness and stakeholder clarity through transparent reporting.
This paper investigates the latency challenges in field-edge systems within transport monitoring environments, with a focus on collision detection and object tracking. Leveraging a field-edge network architecture, we explore the integration of image preprocessing at the field device level and complex object detection and tracking models at the edge device level. Our approach is focused on optimizing the balance between processing speed and accuracy, addressing the critical need for real-time responsiveness in dynamic scenarios such as smart city infrastructure and autonomous navigation. We build on the foundation of existing object detection algorithms, i.e. YOLO, and introduce a batch processing strategy with parallel threads to mitigate latency issues inherent in sequential processing frameworks. Our system is tested using a YOLOv8 model trained on a custom Traffic Object Detection dataset, demonstrating the ability to maintain high processing speeds, with performance exceeding 30 frames per second, without sacrificing accuracy. The approach combines advanced object detection and tracking algorithms with network optimization. Our system tests show positive results in reducing latency. This work sets a benchmark for the deployment of field-edge systems in transport monitoring and contributes to future advancements in smart transportation and autonomous systems.
Feature standardization is a widely applied step in data preprocessing that transform the feature values so that the data mean value is equal to zero and standard deviation is equal to one. It allows for easier and more comprehensible comparison between features by making each feature contribute evenly to the process of analysis. The aim of this paper is to investigate the impact of data standardization on the performance of principal component analysis (PCA) and classification using support vector machine (SVM) algorithm. The analysis is performed on two datasets with overlapping feature values and more separately spaced feature values for unstandardized and standardized data. Additionally, balanced and imbalanced datasets are evaluated. Final results showcase that data standardization significantly improves the performance of PCA as well as classification effectiveness of the SVM algorithm.
This study compares and evaluates different methods for malware detection. It specifically examines the effectiveness of ensemble learning, which combines predictions from multiple models (such as Decision Trees (DT), Support Vector Machines (SVM), and Random Forest (RF)) using a voting approach, against a advanced deep learning algorithm, the Deep Neural Network (DNN). The evaluation encompasses a rigorous assessment of these models using the widely recognized CIC-MalMem-2022 dataset. To optimize the model performance, a strategic feature selection process is employed, reducing the number of input variables and consequently streamlining training time. This consideration is pivotal in achieving efficient and practical detection solutions. The study covers both binary and multi-class classification scenarios, adding a layer of complexity to the evaluation process. By addressing multiple classification types, the research accounts for diverse real-world applications and extends the understanding of model behavior across different scenarios. In the context of performance evaluation, various metrics are employed to gauge the models' effectiveness, including accuracy, precision, recall, F1-Score, and ROC-AUC. These metrics provide a comprehensive overview of the models' predictive capabilities. Notably, in two distinct scenarios, the ensemble learning approach demonstrated substantial promise by attaining commendable accuracy rates, outperforming the DNN model.
Hybrid learning, which integrates in-person and online learning experiences, presents specific delivery challenges that affect the quality and effectiveness of instruction. These challenges vary depending on the context and available technologies. This paper explores different issues in delivering hybrid learning, focusing on synchronous and asynchronous learning methods, technology integration, and finding the right balance between in-person and online interactions. It also highlights technical issues such as internet connectivity, platform compatibility and audio and video quality. In addition, it discusses the advantages of hybrid learning, including cost-effectiveness and increased accessibility, as well as challenges such as reduced face-to-face interaction and technological differences between students. Addressing these issues requires comprehensive support, equal access to resources, and effective training for educators and students. Ultimately, the effectiveness of hybrid learning depends on a well-designed curriculum, teacher training, and seamless integration of technology.
This research explores the security challenges in leveraging Blockchain technology within the FinTech sector, particularly focusing on the control of digital trustworthiness in combating money laundering. Despite Blockchain's promise in enhancing the security and integrity of digital transactions and identities, the study identifies critical security gaps in technical and operational security controls over FinTech domains. The findings emphasize the necessity for a concerted effort among researchers, technology developers, regulatory authorities, and financial institutions to address these challenges.
Research shows that the development of artificial intelligence technology brings with it many new opportunities and challenges. In recent years, generative models such as Generative Adversarial Networks (GANs) have become increasingly important for generating realistic images, sounds and text. In this review article, authors propose new GAN architecture enhanced with CLAHE algorithm to improve process of generating grayscale images. We employ CLAHE to preprocess the X-ray images, which enhances the local contrast and highlights the important features necessary for accurate diagnosis. The preprocessed images are then fed into a GAN architecture designed for image analysis classification. We evaluate our approach on a medical dataset of chest X-ray images and achieve a remarkable accuracy of 93.61
The online learning activities have gained prominence during the past three years, and an appreciable amount of significant learning components were adapted for online learning patterns. Pandemic enhanced the conceptual and practical preoccupations concerning the online learning processes, with an emphasis on the overall development of the networking techniques, software tools and services. There are various learning materials, such as open papers, books, coding examples, dictionaries and online trainings, which are dedicated to specific didactic processes, or transmitted through animation-based approaches. Thus, the existing scientific literature proposes new machine learning and natural language processing-based mechanisms, which are related to, for example, materials dedicated to learning foreign languages. The consideration of such learning methods may be efficiently used in order to enhance the study of structured learning scopes, such as computer science, mathematics, or other subjects. Consequently, the relevant learning processes are re-structured and enhanced. This work addresses approaches that relate to the combination of classical learning techniques like recognition, intercalation, recall time, and written notes, with proper tools and improved educational processes. This paper presents an integrated online learning paradigm, which is validated considering the didactic interaction with relevant student groups. Additionally, this article discusses the real-world benefits of the proposed learning model, and on the open research questions that will be approached by future research projects.
As a result of the information analysis of various approaches to the types of interaction of robots in a group, the following tasks are solved: an environment model and the algorithm for mobile robot autopilot in the conditions of information uncertainty and a multi-agent simulation model and algorithm for focus group behavior of a team of robots in the framework of inter-agent communication were implemented. Based on the proposed simulation model the software has been developed that allows studying the effectiveness of group robots behavior algorithms.
Text simplification involves reducing complexity while retaining important information. This is important for improving accessibility for a wide range of readers, including those with cognitive disorders, non-native speakers, and children. We report experiments on text simplification for Lithuanian with a focus on simplifying texts, written in an administrative style, which is not easy to understand for the general public. We chose mT5 as a foundational model and fine-tuned it in 3 different ways. We evaluated models' outputs with 'unseen' data and automatic evaluation metrics accompanied by qualitative analysis. Though automatic evaluation scores were higher for the model fine-tuned with translated data, qualitative analysis revealed that the model that was fine-tuned with specifically for this task prepared Lithuanian data performed simplification task significantly better.
In the management of kidney stones, identifying the stones' composition is crucial for determining appropriate treatment methods, as different compositions may require different treatment strategies. This study introduces a novel approach utilising Fourier Transform Infrared (FTIR) spectra data to determine kidney stones' composition. By using non-negative matrix factorisation (NMF) for dimensionality reduction and Gaussian mixture model (GMM) for Bayesian soft classification, this method aims to offer a faster and non-destructive alternative to other traditional methods, such as wet chemistry. Initial findings indicate that this approach can effectively analyse kidney stones' composition, showing promise even in cases with limited sample sizes. The application of NMF and GMM to FTIR data could significantly improve the efficiency and accuracy of kidney stones' composition analysis, potentially leading to more effective treatment options.
Proportional representation is one of main families of voting methods for electing several candidates. It is further subdivided into party list based proportional representation and single transferable vote (STV). It is known that it is possible to simplify the calculations needed for implementation of single transferable vote by using one of the methods based on party lists- largest remainder method, if the preference lists are made of relatively long non-intersecting party lists (possibly listing one party list after another in the preference list, but keeping the order in each party list). This paper discusses a way to produce the simulated party lists that would make it possible to use the same algorithm without explicitly constructing preference lists from the party lists.
In view of the fact that modern 21st century technologies have become greatly dependent on radio signals, it stands to reason that their security has moved to the front bon of priority. The paper under review is largely concerned with the popular types of radio hacking, it gives a detailed description and demonstration of some of the types and it also offers an in-depth analysis of GPS spoofing, including its risks and the measures that can be taken to prevent it. Another, main topic is ADSB system of ADS-B. This system periodically sends vital flight information units such as an aircraft location and an altitude. The work reviews both forms of spoofing technique in detail by noting their potential consequences for the navigational application safety and airspace integrity. In addition, it provides numerous case studies to demonstrate the changing landscape in navigation and aviation cybersecurity threats.
The work is about the development of a web-based website selection decision support system, which involves the development of the system's data model, algorithm, and user interface. The work proposes evaluation criteria for websites based on search engine optimization metrics. The system's algorithm ranks websites based on many criteria. This allows the user of the system to use the evaluation criteria offered by us and, in case of desire, to increase the number of evaluation criteria. The algorithm is partially expert because the evaluation criteria proposed by us are not evaluated by a human-expert, their value will be determined by the system for the corresponding website, and as for the evaluation criteria added by the user, it is evaluated by a human-expert. This system will enable users to select the best websites for their business quickly and easily, without excessive costs.
Dating manuscripts, particularly when dealing with handwritten materials, poses a distinct challenge, especially compared to identifying authors in anonymous holographs. The wide array of differences in handwriting styles among various authors, coupled with the tendency for these differences to evolve over time, can often obscure more subtle nuances within the handwriting of a single author. In response to this challenge, our research endeavors to explore a diverse set of methodologies and approaches aimed at precisely dating the undated holographs of Galaktion Tabidze, a prominent Georgian poet from the 20th century. This article outlines two distinct approaches utilized in our study, delving into the experiments conducted to assess the efficacy of each proposed dating method. Through the investigation of these methodologies, our objective is to offer valuable insights and enhance the accuracy of dating historical holographs, thus contributing to a deeper understanding of Tabidze's work and the broader historical context in which it was created.
The motor with permanent magnets used for field excitation, including the BLDC motor, currently finds numerous applications, particularly in electric vehicle propulsion. These motors are powered by an electrical system called an electronic commutator. It is justified to investigate the efficiency of converting electrical energy into mechanical energy depending on the method of supplying the motor windings. Speed regulation requires changing the average voltage supplied to the motor windings. The average voltage supplied to the motor can be changed in two ways: by pulse width modulation (PWM) implemented on the transistors of the electronic commutator or by using an additional DC/DC converter connected between the power source and the electronic commutator. The article presents a research setup examining the total energy losses in the studied BLDC motor depending on the method of power supply, load torque, and rotational speed.
