Oleh Basystiuk

Retrieval-Augmented Generation (RAG) systems have become a fundamental paradigm for enhancing the factual reliability and contextual grounding of large language models by integrating external knowledge sources into the generation process. However, the effectiveness of such systems strongly depends on the quality of the document retrieval and ranking stages, which must accurately prioritize the most relevant evidence for downstream reasoning. This study proposes a method of dynamic feature aggregation for document reranking in multimodal Retrieval-Augmented Generation systems. The approach integrates heterogeneous relevance indicators into a unified neural representation capable of learning nonlinear dependencies between signals. The proposed framework aggregates semantic, lexical, and structural features through trainable neural architectures designed to capture complex relationships between ranking cues. Two alternative models were implemented and evaluated within the proposed pipeline. The first model is based on a Kolmogorov–Arnold Network (KAN) that employs spline-based functional transformations to approximate multidimensional mappings between input features and ranking scores. The second architecture, referred to as Explainable Network (XNet), uses parametric Cauchy activation functions to enable interpretable nonlinear feature transformations while maintaining model transparency. Experimental evaluation demonstrates that both neural reranking models significantly improve retrieval performance compared to deterministic baselines that combine lexical and semantic scores through static weighting schemes. In particular, the KAN-based aggregation model achieved the best overall performance, reaching NDCG@5 = 0.9158, MRR = 0.9018, and MAP = 0.8866, outperforming the strongest baseline (BM25 + Semantic) by more than three percentage points. These results indicate that dynamic nonlinear aggregation effectively captures complementary relevance signals and improves prioritization of top-ranked documents in RAG pipelines. The findings suggest that adaptive neural feature aggregation can substantially enhance the reliability and precision of document retrieval in multimodal intelligent systems, particularly in domain-specific applications where heterogeneous evidence sources must be integrated for accurate knowledge-grounded generation.

Modern AI development and multimodal data analysis methods are gaining critical importance due to their ability to integrate information from diverse sources, including text, audio, sensor signals, and images. Such integration enables systems to form a richer and more context-aware understanding of complex environments, which is essential for domains such as healthcare diagnostics, adaptive education technologies, intelligent security systems, autonomous robotics, and various forms of human-computer interaction. Multimodal approaches also enable AI models to compensate for the limitations inherent in individual modalities, thereby enhancing robustness and resilience to noise or incomplete data. The study employs theoretical analysis of scientific literature, comparative classification of multimodal architectures, systematization of fusion techniques, and formal generalization of model design principles. Additionally, attention is given to evaluating emerging paradigms powered by large-scale foundation models and transformer-based architectures. The primary methods and models for processing multimodal data are summarized, covering both classical and state-of-the-art approaches. Architectures of early (feature-level), late (decision-level), and hybrid (intermediate) fusion are described and compared in terms of flexibility, computational complexity, interpretability, and accuracy. Emerging solutions based on large multimodal transformer models, contrastive learning, and unified embedding spaces are also analyzed. Special attention is paid to cross-modal attention mechanisms that enable dynamic weighting of modalities depending on task context. The study determines that multimodal systems achieve significantly higher accuracy, stability, and semantic coherence in classification, detection, and interpretation tasks when modalities are properly synchronized and fused using adaptive strategies. These findings underscore the promise of further research toward scalable architectures capable of real-time multimodal reasoning, improved cross-modal transfer, and context-aware attention mechanisms.

Ontologies are a modern direction in the field of artificial intelligence. In the general sense, ontology can be characterized as a special type of knowledge base or as a “specification of conceptualization” of a particular subject area. This implies that within a specific location, based on the classification of basic terms, key concepts are highlighted and the relationships between them are determined. In the future, the ontology can be represented in the form of a graphical model or described using formal languages, which is the process of creating an ontology specification. Ontological knowledge bases are a powerful tool for organizing and processing large amounts of information. They allow you to formalize knowledge in the form of a clear structure, where each element of the base has a clearly defined meaning, properties, and relationships with other components. They are handy for working with data that has a complex hierarchical or relational structure, as is often the case when analyzing thermal portraits of targets.

Depression is one of the most common mental health disorders in the world, affecting millions of people. Early detection of depression is crucial for effective medical intervention. Multimodal networks can greatly assist in the detection of depression, especially in situations where in patients are not always aware of or able to express their symptoms. By analyzing text and audio data, such networks are able to automatically identify patterns in speech and behavior that indicate a depressive state. In this study, we propose two multimodal information fusion networks: early and late fusion. These networks were developed using convolutional neural network (CNN) layers to learn local patterns, a bidirectional LSTM (Bi-LSTM) to process sequences, and a self-attention mechanism to improve focus on key parts of the data. The DAIC-WOZ and EDAIC-WOZ datasets were used for the experiments. The experiments compared the precision, recall, f1-score, and accuracy metrics for the cases of using early and late multimodal data fusion and found that the early information fusion multimodal network achieved higher classification accuracy results. On the test dataset, this network achieved an f1-score of 0.79 and an overall classification accuracy of 0.86, indicating its effectiveness in detecting depression.

Sustainable agriculture is pivotal to global food security and economic stability, with plant disease detection being a key challenge to ensuring healthy crop production. The early and accurate identification of plant diseases can significantly enhance agricultural practices, minimize crop losses, and reduce the environmental impacts. This paper presents an innovative approach to sustainable development by leveraging machine learning models to detect plant diseases, focusing on tomato crops—a vital and globally significant agricultural product. Advanced object detection models including YOLOv8 (minor and nano variants), Roboflow 3.0 (Fast), EfficientDetV2 (with EfficientNetB0 backbone), and Faster R-CNN (with ResNet50 backbone) were evaluated for their precision, efficiency, and suitability for mobile and field applications. YOLOv8 nano emerged as the optimal choice, offering a mean average precision (MAP) of 98.6% with minimal computational requirements, facilitating its integration into mobile applications for real-time support to farmers. This research underscores the potential of machine learning in advancing sustainable agriculture and highlights future opportunities to integrate these models with drone technology, Internet of Things (IoT)-based irrigation, and disease management systems. Expanding datasets and exploring alternative models could enhance this technology’s efficacy and adaptability to diverse agricultural contexts.

In our interconnected society, prioritizing the resilience and sustainability of road infrastructure has never been more critical, especially in light of growing environmental and climatic challenges. By harnessing data from various sources, we can proactively enhance our ability to detect road damage. This approach will enable us to make well-informed decisions for timely maintenance and implement effective mitigation strategies, ultimately leading to safer and more durable road systems. This paper presents a new method for detecting road potholes during low-light conditions, particularly at night when influenced by street and traffic lighting. We examined and assessed various advanced machine learning and computer vision models, placing a strong emphasis on deep learning algorithms such as YOLO, as well as the combination of Grad-CAM++ with feature pyramid networks for feature extraction. Our approach utilized innovative data augmentation techniques, which enhanced the diversity and robustness of the training dataset, ultimately leading to significant improvements in model performance. The study results reveal that the proposed YOLOv11+FPN+Grad-CAM model achieved a mean average precision (mAP) score of 0.72 for the 50–95 IoU thresholds, outperforming other tested models, including YOLOv8 Medium with a score of 0.611. The proposed model also demonstrated notable improvements in key metrics, with mAP50 and mAP75 values of 0.88 and 0.791, reflecting enhancements of 1.5% and 5.7%, respectively, compared to YOLOv11. These results highlight the model’s superior performance in detecting potholes under low-light conditions. By leveraging a specialized dataset for nighttime scenarios, the approach offers significant advancements in hazard detection, paving the way for more effective and timely driver alerts and ultimately contributing to improved road safety. This paper makes several key contributions, including implementing advanced data augmentation methods and a thorough comparative analysis of various YOLO-based models. Future plans involve developing a real-time driver warning application, introducing enhanced evaluation metrics, and demonstrating the model’s adaptability in diverse environmental conditions, such as snow and rain. The contributions significantly advance the field of road maintenance and safety by offering a robust and scalable solution for pothole detection, particularly in developing countries.

The steady growth of multimodal data requires effective methods for managing and processing this type of information, including focusing on developing a system for detecting anomalies. This research mainly focuses on detecting anomalies and improving data quality in multimodal datasets. By integrating different data types, such as text, metadata, and images, this research examines the effectiveness of three fusion methods: early, late, and hybrid. Through a comprehensive analysis, the study demonstrates that late fusion significantly improves the accuracy of data processing and achieves the highest processing accuracy of 0.947, which is, on average, 6% higher than the early and hybrid fusion systems. The findings emphasize the potential of multimodal data systems to improve information systems’ reliability, accuracy, and efficiency, thereby supporting more informed decision-making in dynamic environments.

Nowadays, there are more and more web applications of all kinds. Each of them solves a specific problem and makes life easier for its users. Web applications come in many different types: from a platform for learning courses and watching movies to an online store selling goods. The best systems are those that make things as easy as possible for the user, behave like old friends who know the behavior and tastes of their users and can predict their next move. It would be useful to integrate such system behavior into an online store system, as nowadays, a huge number of people prefer to buy goods online, saving time and effort. Thus, recommender systems have become an important tool for improving the efficiency of e-commerce stores and ensuring customer satisfaction. This study analyzes the main approaches to the application of recommender systems for online stores, substantiates the advantages and feasibility of the selected technologies for the implementation of an online store information system using neural networks.

The constant growth of data volumes requires the development of effective methods for managing, processing, and storing information. Additionally, it is advisable to apply multimodal approaches for knowledge aggregation to extract additional knowledge. Usually, the problem of efficient processing of multimodal data is associated with high-quality data preprocessing. One of the most critical preprocessing steps is synchronizing multimodal data stre- ams to analyze complex interactions in different data types. In this article, we evaluate existing approaches to synchronization, focusing on strategies based on real-time classifiers, which are based on comprehensive platforms for data integration and management. After the synchronization of multimodal sets, the key stage is data fusion, data identification in different channels, such as text, video, and audio. The results demonstrate the feasibility of the proposed synchronization approach for revealing subtle relationships between various data sets. An architectural solution was also suggested to integrate the proposed method into existing multimodal data processing pipelines. This work contributes to developing synchronization tools for multimodal data analysis in dynamic real- world scenarios.