Hibridni događaj

Program događaja

četvrtak, 28.5.2026 9:00 - 13:00, Camelia 2, Grand hotel Adriatic, Opatija 9:00 - 10:45 Optimization, Formal Methods and Image Analysis Predsjedatelj: Darko Huljenić   1. J. Sabljo, M. Đumić (School of Applied Mathematics and Informatics, J.J. Strossmayer University of Osijek, Osijek, Croatia), M. Đurasević (University of Zagreb, Faculty of Electrical Engineering and Computing, Zagreb, Croatia) Analysis of Training Models for Evolving Heuristics in the Unrelated Parallel Machines Scheduling Problem with Additional Resources  The unrelated parallel machines scheduling problem with precedence and additional resource constraints (UPMSP-PR) is an NP-hard problem, which makes exact solution approaches impractical for larger instances. Consequently, heuristic methods are commonly employed to address this problem. The design of such heuristics can be effectively tackled using genetic programming (GP), a hyperheuristic approach aimed at evolving new heuristics with strong generalization capabilities. To achieve improved generalization, GP must be trained using high-quality training models and a diverse set of problem instances. In this paper, we generate a comprehensive set of UPMSP-PR instances and analyze different training models, with particular emphasis on the impact of instance selection during training on overall model performance. Specifically, we investigate the influence of problem instance size and examine the role of a validation set during the training phase. 2. J. Angeloska, D. Usovikj, M. Toshevska, S. Gievska (Faculty of Computer Science and Engineering, Skopje, Macedonia) Benchmarking the Gap Between Theory and Practice in Multi-Agent Reinforcement Learning  Multi-agent reinforcement learning has emerged as a critical framework for addressing complex cooperative tasks across robotics, resource management, and autonomous systems. This study presents a systematic comparison of three representative multi-agent reinforcement learning algorithms across coordination-heavy and exploration-focused cooperative tasks. We evaluate Independent Proximal Policy Optimization, Multi-Agent Proximal Policy Optimization, and Multi-Agent Deep Deterministic Policy Gradient using the BenchMARL framework integrated with the Vectorized Multi-Agent Simulator. Our experiments systematically examine these algorithms' performance on the Balance and Sampling environments with agent team sizes ranging from three to seven agents. This reveals scalability characteristics and the practical value of centralized critics in different task contexts. By looking at learning dynamics, convergence behavior, and training stability across 300,000 environment interactions per trial, we establish task-dependent algorithmic selection guidelines that practitioners can use. This work contributes to bridging the gap between theoretical developments in multi-agent reinforcement learning and practical deployment considerations by providing comprehensive analysis and actionable recommendations for algorithm selection in cooperative multi-agent systems. 3. V. Vaskin (ITMO University, Saint Petersburg, Russian Federation), A. Semenov (Matrosov Institute for system dynamics and control theory of SB RAS, Irkutsk, Russian Federation) Adapting Cube-and-Conquer technology to CircuitSAT  Cube-and-Conquer (CnC) is a powerful hybrid SAT solving strategy that has proven effective on hard combinatorial problems, yet its Cube stage operates exclusively on CNF: applying CnC to CircuitSAT instances requires a prior Tseitin transformation that discards exactly the structural information most valuable for decomposition. In this paper, we propose a circuit-native adaptation of CnC that performs the Cube stage directly on the And-Inverter Graph (AIG) representation, without prior CNF conversion. We define a gate-level lookahead mechanism based on gate-value substitution and structural constraint propagation through the AIG, and introduce gate connectivity efficiency metrics that guide the construction of a SAT-partitioning of the circuit. The resulting sub-problems are dispatched to a standard CDCL solver in the Conquer phase, with the Tseitin transformation deferred to this late stage. We evaluate the approach on miter circuits for integer multipliers - benchmarks well known to be among the hardest instances for SAT solvers. Our method outperforms the strongest CNF-based baseline (Kissat) by up to a factor of 2.1 at 11-12 bits, demonstrating that circuit-aware decomposition yields systematic gains that grow with instance size. 4. G. Oparin, V. Bogdanova, A. Pashinin (Institute for System Dynamics and Control Theory of SB RAS, Irkutsk, Russian Federation) Dynamic Coverage of All Graph Vertices in Multi-Agent Pathfinding  A dynamic Boolean model is proposed for the multi-agent pathfinding problem, involving m mobile agents, a graph with n vertices defined by an adjacency matrix, and constraints on admissible agent transitions from a current vertex to the next during the system's operation. Within this model, the set of graph vertices is dynamically partitioned into m disjoint subsets as the agents move, such that each agent operates only within its assigned subset. The objective is for the agents to collectively cover all vertices of the graph along their trajectories within a specified discrete time interval. The conditions for dynamically covering all graph vertices using Hamiltonian cycles and paths are examined. This problem is more general than its static counterpart, as it requires the temporal coordination of all agents' actions. The agents' movement is modeled using a k-th order implicit Boolean network, where the parameter k represents an upper bound for the discrete time steps. Local solutions of this Boolean network (if they exist) define both the composition of the disjoint subsets and the trajectories of the agents within them. The search for local solutions is reduced to a Boolean satisfiability (SAT) problem derived from the dynamics equations governing the implicit Boolean network. Examples of solving the dynamic coverage problem are provided for grid graphs and Knight's tour graphs. 5. H. Nuić, D. Vršnak, S. Lončarić (University of Zagreb Faculty of Electrical Engineering and Computing, Zagreb, Croatia) Automatic Exposure Control Guided by No-Reference Image Quality Estimation  Automatic exposure (AE) algorithms in automotive cameras must balance competing constraints such as noise, motion blur, saturation, and perceptual image quality. In this work, we propose a framework for exposure parameter selection based on visual quality assessment. Using a dataset of synchronized HDR and SDR image pairs with exposure metadata, we generate candidate SDR frames corresponding to different combinations of exposure time, analog gain, and digital gain. These candidates are evaluated using a no-reference image quality assessment model, which provides a perceptual quality score for each configuration. The predicted quality scores are then used to rank exposure parameter combinations and identify exposure settings that maximize perceived image quality. Unlike traditional AE methods that primarily rely on heuristic brightness level objectives, the proposed approach directly optimizes perceptual quality and enables offline analysis of AE behavior without repeated field testing. The framework is designed to support rapid prototyping of AE strategies, can be extended to task-aware objectives or hardware ISP constraints, and enables perceptually driven automatic exposure control in automotive imaging systems. 6. R. Franetović, M. Subašić (University of Zagreb Faculty of Electrical Engineering and Computing, Zagreb, Croatia) Comparison of the General Linear Model with Classical Color Correction Models for Color Calibration  Image signal processing pipelines, found in all digital cameras, use a color correction matrix as a simple and efficient model of color correction. In the case of color calibration, a reference object is placed in the image which allows direct calculation of the color correction matrix coefficients. Trough experimental application of the general linear model (GLM) [1], we noticed a significant performance improvement over classical color correction models for the task of color calibration and decided to perform a detailed comparison of the color correction models for the task of color calibration. We compare the performance of the GLM with the standard methods in the case of color calibration, demonstrating significant performance benefits at low computational cost. We explore and compare the robustness of the GLM to detection errors and influence of noise. Additionally, we perform a statistical analysis of the results. Finally, we analyze the computational cost and potential hardware implementation of the general linear color correction model. 7. I. Reljić, S. Seljan, I. Dunđer (Faculty of Humanities and Social Science, Zagreb, Croatia) 3D Reconstruction Using Artificial Intelligence: Photographic Data Requirements, Sensitivity and Model Robustness  Access to virtual representations of real-world objects expands the possibilities of their use, from scientific research to promotional content. They can be created through the process of 3D digitization, which requires efficient data collection for specific applications, such as 3D scanning of cultural heritage with limited resources. Creating high-resolution digital twins of physical objects has significantly different requirements compared to simplistic object representations. Additionally, when scanning a large number of objects, efficiency is of utmost importance. The goal of this research was to determine the minimum amount of photographic data needed to create a “useful” 3D model, which is defined as accurately shaped representations without large holes or surface artifacts that would negatively affect the viewer’s experience. In this research, two approaches based on two cultural heritage objects were explored – standard photogrammetry and neural radiance fields (NeRF), which use artificial neural networks. From the initial complete photographic datasets, progressively smaller datasets were created by reducing the photographic data, creating a series of 3D models of decreasing quality. This analysis allows researchers to optimize the 3D scanning process by matching the size of the photographic dataset to their needs for 3D model quality, which prevents unnecessary data collection, processing and storage. 10:45 - 11:00 Pauza   11:00 - 12:45 Computer Vision Predsjedatelj: Marina Ivašić-Kos   1. D. Gržinić, M. Ivašić-Kos (Faculty of Informatics and Digital Technologies, University of Rijeka, Rijeka, Croatia) Detection of Sprites in Night Sky Images Using the YOLO Model   Sprites are brief transient luminous events occurring above active thunderstorms, lasting up to 300 ms and extending tens of kilometers in the upper atmosphere. Their short duration and high altitude make systematic observation difficult, and manual identification in Global Meteor Network (GMN) imagery slow and labor-intensive. This paper presents an automated sprite detection system based on the YOLO family of deep neural networks, designed for integration into the GMN processing pipeline. We investigated preprocessing strategies including image scaling tailored to the elongated sprite morphology and heuristic filtering to suppress false detections from noise and camera artefacts. Experiments compared YOLOv5, YOLOv8, and YOLOv11 architectures trained from scratch and with transfer learning. The YOLOv5m model fine-tuned on a sprite dataset achieved the best performance with mAP of 0.922, while the lightweight YOLOv5s variant reached 0.919 mAP with reduced complexity. These results demonstrate that YOLO-based detectors offer a reliable, scalable solution for automated sprite detection, enabling large-scale surveys and improved statistical studies of transient luminous events. 2. D. Kučak, B. Skračić, D. Pećarina (University Algebra Bernays, Zagreb, Croatia) On-Device Computer Vision and Monte Carlo Simulation for Board Game Unit Recognition and Battle Outcome Estimation  This paper presents a fully on-device mobile system that integrates computer vision and probabilistic simulation to support decision-making in complex board game combat scenarios. The proposed solution targets the board game Twilight Imperium and addresses the challenge of automating unit recognition and battle outcome estimation without manual input. A YOLOv8-based object detection model was trained on a custom dataset of game miniatures, optimized, and deployed using TensorFlow Lite to enable real-time inference on Android devices without network dependency. Detected units are classified, assigned to players through color analysis, and mapped to a battle simulation module based on Monte Carlo methods. The system performs thousands of simulation iterations per evaluation to estimate win, loss, and draw probabilities. Experimental results demonstrate that accurate unit detection and low-latency inference can be achieved on modern mobile hardware, enabling responsive interaction and practical in-game use. The presented approach confirms the feasibility of combining on-device computer vision and stochastic simulation into a lightweight, fully local assistant for complex analog board games. 3. V. Dimitrievska (Silicon Austria Labs, Villach, Austria), N. Ackovska, I. Ivanoska (Faculty of Computer Science and Engineering, Skopje, Macedonia) Graph Neural Networks for Explainable Computer Vision in Robotics: An Overview  Robotic systems depend on multiple computer vision tasks to achieve effective environmental perception, including object and human detection, scene understanding, and human activity recognition. Effective perception in robotics requires not only accurate recognition but also an understanding of the relationships and interactions among entities within dynamic environments. This paper presents an overview of the application of Graph Neural Networks (GNN) to these vision tasks, emphasizing their ability to provide explainable representations. Graph neural networks are particularly well suited for these applications because they can effectively model complex relational structures inherent in visual data, including scene graphs and human activity graphs, enabling structured and context-aware reasoning. By reviewing recent advances and categorizing existing methods, this survey highlights how GNNs contribute to the development of safer, more transparent, and more trustworthy robotic perception systems. 4. A. Bošnjak, P. Pejić, R. Cupec, J. Job, E. Nyarko (Faculty of Electrical Engineering, Computer Science and Information Technology Osijek, Osijek, Croatia), B. Lukić (Faculty of Agrobiotechnical Sciences Osijek, Osijek, Croatia) Computer Vision for Automated Cattle Monitoring: A Review of Detection, Tracking, and Re-Identification  Computer vision has become a practical, non-invasive alternative to wearable sensors for continuous cattle monitoring in barns and other farm environments. This paper reviews recent deep learning pipelines for automated cattle monitoring, focusing on three core components: detection (often combined with behavior recognition), multi-object tracking, and individual identification or re-identification. In the literature, YOLO-family detectors are prevalent due to their efficiency and accuracy, and are often enhanced with lightweight architectural modules (such as attention mechanisms, modified downsampling, and multi-scale feature aggregation) tailored to cattle-specific imagery. Tracking is typically addressed using tracking-by-detection frameworks (such as DeepSORT, BoT-SORT, and ByteTrack) with association refinements to reduce fragmentation and identity switches. For dorsal-view identification, most approaches use coat-pattern cues through either template or descriptor matching or learned embedding spaces, with growing attention to open-set recognition and cross-domain generalization. Representative methods and reported metrics are summarized, and highlight persistent limitations such as occlusions, domain shift, and the lack of standardized multi-camera benchmarks with identity-level ground truth. Finally, directions for more robust, unified evaluation and deployment-ready systems are outlined. 5. M. Stipičević, M. Ivanković, S. Stipetić, T. Radišić (University of Zagreb, Faculty of Transport and Traffic Sciences, Zagreb, Croatia) Cross-Modal Weak Supervision Training of UAV Detection Model for Degraded Environments  The proliferation of Unmanned Aerial Vehicles (UAVs) necessitates robust detection systems capable of operating in degraded visual environments (DVE) such as night, fog, or smoke. While visual spectrum (RGB) detection models have reached maturity, their performance collapses in low-light conditions. Thermal infrared (TIR) imaging offers a viable solution but is constrained by the scarcity of annotated datasets. This paper proposes a Cross-Modal Weak Supervision pipeline to automate the generation of thermal training data. We utilize a high-performance YOLO model trained on RGB imagery to generate pseudo-labels for synchronized thermal recordings, bridging the modality gap via a geometric coordinate transformation. We evaluate the efficacy of this pipeline by training a specific thermal detection model (YOLO11x) on the auto-generated dataset. Our results demonstrate that the thermal model achieves a mean Average Precision (mAP50) of 98.5%, effectively transferring semantic knowledge from the visual to the thermal domain. Furthermore, we provide a comparative analysis of visual versus thermal detection capabilities, validating the system’s robustness in conditions where visual models fail. 6. B. Baftijari (University of Tetovo, Faculty of Natural Sciences and Mathematics, Tetovo, Macedonia), S. Koceski, N. Koceska (Goce Delcev University, Stip, Macedonia) Object Recognition in Household Environments Using YOLOv11  Object detection has become a key component in many computer vision applications, particularly in areas such as smart homes, health care, robotics and assistive technologies. Detecting daily household items accurately in real-world environments remains challenging due to variations in object appearance, lighting conditions and background cluttering. In this study, we investigate the use of YOLOv11 for recognizing common household items such as keys, remote controllers, eye glasses and pills. A dataset consisting of the daily home items was collected and manually annotated using Roboflow in YOLO format. The dataset was divided into training, validation and testing sets, and preprocessing and data augmentation techniques were applied to improve model generalization. The YOLOv11 model was trained on the dataset using an RTX4060 GPU and evaluated using standard evaluation metrics such as precision, recall and mean Average Precision (mAP). Experimental results demonstrate that the trained model achieves strong detection performance across most classes, highlighting the effectiveness of YOLOv11 for custom, smallscale object detection tasks. The results indicate that the YOLOv11 can be effectively applied to real-world household object detection scenarios and provide a practical foundation for future smart home and automation applications. 7. B. Jakupović, D. Đekić, K. Host (Faculty of Informatics and Digital Technologies, University of Rijeka, Rijeka, Croatia), M. Ivašić-Kos (Faculty of Informatics and Digital Technologies, University of Rijeka and Centre for Artificial Inte, Rijeka, Croatia) Pedestrian Tracking in Dynamic Urban Scenes Using YOLO11x and OC-SORT with Camera-Motion Compensation  Reliable automatic detection and tracking of pedestrians is important for traffic safety and autonomous driving, yet remains challenging in dynamic urban scenes. Clustered backgrounds, occlusions, pedestrians moving in different directions at varying distances, and continuous ego-motion of the camera often lead to fragmented trajectories and identity switches, which degrade the system’s ability to anticipate pedestrian intent. This paper investigates and compares several tracking strategies, ranging from Kalman-filter-based single-object trackers to modern multiple-object tracking approaches such as Norfair, DeepSORT, and OC-SORT, and proposes enhancements that compensate for camera motion and improve tracking through occlusions. The proposed modifications reduce identity switches and extend the duration of consistent pedestrian tracks. Experiments on the Multiple Object Tracking 2017 benchmark demonstrate that the best performance, with Multiple Object Tracking Accuracy (MOTA) of 40.4% and Identity F1 score (IDF1) of 51.7%, is achieved by a configuration that combines a YOLO11x detector with the OC-SORT tracker augmented by ByteTrack association logic and a Camera Motion Compensation module. These results indicate that reliable perception for autonomous driving requires coupling robust object detection with explicit camera-motion modelling and advanced data-association techniques within the MOT pipeline. četvrtak, 28.5.2026 15:00 - 19:00, Camelia 2, Grand hotel Adriatic, Opatija 15:00 - 16:45 Natural Language Processing - I Predsjedatelj: Darko Huljenić   Pozvano predavanje  Davor Runje (Synthpop AI, Zagreb, Croatia) Communicating AI Agents  Radovi  1. L. Hobor, M. Brčić, M. Kovač, K. Poje (University of Zagreb Faculty of Electrical Engineering and Computing, Zagreb, Croatia) Bayesian Elicitation with LLMs: Model Size Helps, Extra "Reasoning" Doesn’t Always  Large language models (LLMs) have been proposed as alternatives to human experts for estimating unknown quantities with associated uncertainty, a process known as Bayesian elicitation. We test this by asking eleven LLMs to estimate population statistics, such as health prevalence rates, personality trait distributions, and labor market figures, and to express their uncertainty as 95% credible intervals. We vary each model’s reasoning effort (low, medium, high) to test whether more "thinking" improves results. Our findings reveal three key results. First, larger, more capable models produce more accurate estimates, but increasing reasoning effort provides no consistent benefit. Second, all models are severely overconfident: their 95% intervals contain the true value only 9–44% of the time, far below the expected 95%. Third, a statistical recalibration technique called conformal prediction can correct this overconfidence, expanding the intervals to achieve the intended coverage. In a preliminary experiment, giving models web search access degraded predictions for already-accurate models, while modestly improving predictions for weaker ones. Models performed well on commonly discussed topics but struggled with specialized health data. These results indicate that LLM uncertainty estimates require statistical correction before they can be used in decision-making. 2. R. Peran, L. Hobor, M. Kovac, M. Brcic (University of Zagreb Faculty of Electrical Engineering and Computing, Zagreb, Croatia) Large Language Models as Optimizers: A Survey of Direct vs. Tool-Augmented Approaches and Their Performance Frontiers  Large Language Models (LLMs) are increasingly involved in complex mathematical optimization, even if the pragmatic user who triggers them is unaware of it. After all, many real-world problems reduce to the search for better or the best solutions. The field of LLM-as-optimizer has three paradigms: direct optimization, tool-augmented optimization, and tool-creating optimization. Direct optimization uses iterative prompting and heuristic generation to navigate solution spaces. Tool-augmented optimization translates natural language problems into formal specifications and orchestrates external solvers. Tool-creating optimization goes further, using LLMs to discover reusable algorithms or heuristics that can be deployed at zero marginal LLM cost. We describe current performance frontiers based on the benchmarks from the literature. We identify the critical reasoning gap in current architectures and argue for tradeoffs between the future potential of direct optimization and the auditability of tool-augmented optimization. Even future, more powerful models might opt for tool-making to improve operational efficiency for repetitive families of problems. 3. M. Jakimoski, M. Toshevska, S. Gievska (Faculty of Computer Science and Engineering, Skopje, Macedonia) From Syntax Trees to Embeddings: A Comparative Study of AI-Generated Code Detection  Large language models(LLMs) are increasingly used to generate program code, which complicates code authorship attribution in settings where originality is required (e.g., academic assessment). This paper studies binary classification of source code as either AI-generated or human-written using the CoDeT-M4 dataset (500,552 samples across Java, Python, and C++). We compare three families of approaches: (i) traditional machine learning on engineered features extracted from concrete syntax trees, (ii) embedding-based deep models built on pretrained CodeBERT representations, and (iii) graph-based models operating on concrete syntax tree-derived graphs. Our results show that embedding-based architectures achieve the highest performance, while graph-based methods remain competitive and highlight the value of structurally grounded representations, including simple comment-indicator nodes, for authorship detection. 4. J. Blanco-Guzmán, J. Higuera Reina, P. Bornes Roldán (University of Seville, Sevilla, Spain) From Search Rankings to Answer Rankings: Generative Search and the Transformation of Online Visibility  For decades, online visibility has been structured around ranking-based logics, where search engines ordered information according to relatively stable and transparent criteria. This model has shaped how websites, brands, and organizations compete for attention in digital environments. The recent integration of large language models (LLMs) into search systems represents a qualitative shift in this logic. Instead of ranking links, generative search engines increasingly deliver personalized, synthesized answers designed to directly satisfy user intent and reduce navigation effort. This transformation marks a turning point in the evolution of online positioning. Visibility is no longer determined solely by predefined optimization criteria, but by opaque, adaptive, and context-dependent generative processes. As a result, digital competition is reconfigured: some sources are selectively incorporated into answers, while others become effectively invisible, regardless of their previous positioning. This study explores the implications of generative search for online visibility and digital competitiveness. By examining how generative systems reshape access to information, the paper highlights emerging risks for websites and firms that are unable to adapt to this new competitive environment, and outlines the need for systematic approaches to understanding visibility in answer-based search ecosystems. 5. A. Nedbaylo, D. Hristovski (Faculty of Public Administration, University of Ljubljana, Ljubljana, Slovenia) Large Language Model-Driven Literature-Based Discovery: A Retrieval-Augmented Generation Framework for Hypothesis Generation and Evidence-Based Screening  Literature-based discovery (LBD) has traditionally relied on structured biomedical terminologies to identify "undiscovered public knowledge," a constraint that limits its applicability in heterogeneous domains such as the public sector. Advancing from initial manual prompting techniques, this study presents a transition toward an automated Retrieval-Augmented Generation (RAG) framework. The developed LBD-RAG framework couples the generative reasoning of Large Language Models with empirical grounding against external bibliographic databases including Scopus and PubMed. To assist in evaluating generated hypotheses and ensure empirical grounding, the system integrates the Retrieval-Augmented Generation Assessment (RAGAS) framework. Through a three-pass evaluation (A→B, B→C, A→C), the system measures faithfulness, answer relevancy, and context precision to verify that model responses are supported by retrieved literature rather than internal parametric knowledge. By expanding the application of the classic ABC discovery model beyond biomedicine, this research demonstrates the utility of the framework in identifying structural solutions for public sector challenges. Initial trials indicate that this hybrid architecture facilitates domain-agnostic discovery and provides a transparent methodology for the standardized evaluation of AI-generated insights, offering a scalable approach for identifying novel research directions in established scientific literature. 16:45 - 17:00 Pauza   17:00 - 18:45 Natural Language Processing - II Predsjedatelj: Marko Horvat   1. J. Rybicki (Forschungszentrum Juelich GmbH, Juelich, Germany) How to Chat with a Supercomputer  Historically, accessing supercomputers has been rather tedious and often required substantial technical knowledge, limiting their user base to a small group of highly skilled individuals. However, as demand for computation grows, more scientific users are seeking access to these resources. In this paper, we leverage large language models (LLMs) to provide a novel approach for accessing supercomputers. We describe the technical details of this integration, followed by an initial evaluation demonstrating the solution’s ability to handle standard scientific use cases and offering insights into system performance. Our goal is to broaden the user base for supercomputing facilities, while the solution also holds interest from an AI perspective: it extends the language-based information retrieval interface of LLM agents with powerful processing capabilities, potentially accelerating scientific discovery 2. J. Scheffler, S. Naumann, F. Mohr, L. Begic Fazlic (Trier University of Applied Sciences, Environmental Campus Birkenfeld, Birkenfeld, Germany) Characterizing Prefill and Decode Regimes for Edge LLM Inference  Large language models (LLMs) are increasingly deployed on edge devices to reduce latency, preserve privacy, and enable offline operation. However, edge platforms impose strict constraints on memory capacity, power consumption, and thermal dissipation, which shape inference behavior. Modern LLM inference consists of distinct execution regimes—prompt evaluation and autoregressive decoding—with different scaling and stability characteristics, yet are often assessed using aggregate performance metrics. In this work, we experimentally characterize LLM inference on representative edge hardware by isolating prefill and decode behavior under input length scaling and sustained execution. We evaluate a CPU-only Raspberry Pi 5 and a GPU-enabled NVIDIA Jetson Orin Nano using a common inference runtime and quantized LLaMA models. Our results show that increasing input length degrades both prefill and decode performance, with decode throughput decreasing as context length grows and thermal effects amplifying this degradation on CPU-only platforms. Under sustained workloads, the Jetson Orin Nano maintains stable performance and favorable energy efficiency, while the Raspberry Pi 5 exhibits throughput and efficiency degradation due to passive cooling and thermal throttling. These results underscore the importance of regime-aware evaluation and thermal considerations for practical LLM deployment on edge devices. 3. I. Kostadinova, M. Garvanova, I. Garvanov (University of Library Studies and Information Technologies, Sofia, Bulgaria), N. Kerimbayev ( Al-Farabi Kazakh National University, Almaty, Kazakhstan) Linguistic Sovereignty in Аlgorithmic Recommendation Systems  This paper introduces the Diversifying Algorithm for Recommendations (DAR), a conceptual technical framework designed to mitigate systemic algorithmic bias against small linguistic markets within the European Union’s digital ecosystem. By embedding an Ethnolinguistic Cultural Factor (ECF) into the collaborative filtering scoring function, the proposed framework aims to operationalize the “prominence” requirements mandated by Directive (EU) 2018/1808 (AVMSD). The research draws on Value Sensitive Design (VSD) and User-Centered Design (UCD) methodologies and adopts a Dual-Layer Algorithmic Fairness approach to balance individual user preferences with group-level linguistic equity. A critical analysis of existing recommendation system literature identifies three key gaps: the absence of linguistic dimensions in diversity-aware methods, the lack of regulatory grounding, and the absence of quantifiable prominence metrics. DAR addresses these gaps by proposing a mathematically formalized scoring mechanism that is auditable, scalable, and aligned with the European Union (EU) regulatory objectives. The framework’s empirical validation through simulation studies and real-world deployment is identified as the primary direction for future research. 4. M. Ribarić, A. Poleksić, S. Martinčić Ipšić (Faculty of Informatics and Digital Technology, University of Rijeka, Center for Artificial Intellige, Rijeka, Croatia) An Empirical Comparison of LoRA and DoRA for Parameter-Efficient Fine-Tuning for Climate Change  The paper compares the performance of Parameter-Efficient Fine-Tuning (PEFT) methods LoRA and DoRA for adapting the SciBERT model to the climate change domain. The paper focuses on evaluating model quality changes with altering data size and adapter configuration. The climate-change corpus of approximately 400,000 sentences is extracted from scientific papers and then partitioned into nested subsets, where each smaller subset is fully contained within the larger ones. Separate PEFT adapters are trained for SciBERT with LoRA and DoRA on each subset, enabling systematic analysis of performance scaling with data size. All LoRA and DoRA models are trained on scientific climate-change texts using SciBERT as the base model and evaluated on the SciDCC and Climate-FEVER datasets from ClimaBench using macro F1 score. Results are compared against the fully domain-adapted CliSciBERT model, which was pretrained on a larger climate-change corpus, while only a small portion of that corpus is reused here for PEFT finetuning. Experiments identify the optimal data amount for LoRA/DoRA models to approach or match CliSciBERT performance, while containing far fewer trainable parameters. 5. C. Martínez-Araneda, M. Gutiérrez Valenzuela, P. Gómez-Meneses (Universidad Católica de la Santísima Concepción, Concepción, Chile), D. Maldonado Montiel (Universidad Católica del Maule, Talca, Chile), A. Segura Navarrete, C. Vidal-Castro (Universidad del Bio Bio, Concepción, Chile), E. Pérez Álvarez (Universidad Católica de la Santísima Concepción, Concepción, Chile) A Curated Spanish-language Dataset for Detecting LGBTQIA-phobic Language in Digital Environments  This article presents the LGBTQIAphobia dataset (augmented and balanced), a curated Spanish-language corpus designed to support research on the automated detection of discriminatory and hateful language targeting LGBTQIA+ communities in digital environments. The dataset comprises 1,000 short textual phrases collected from publicly available content on the web, X/Twitter, Instagram, TikTok, and YouTube comments. Each instance was manually annotated by three independent raters using a binary labeling scheme (1 = LGBTQIAphobic; 0 = non-LGBTQIAphobic). To improve robustness and reproducibility, text augmentation and class-balancing via undersampling were applied. The final dataset is evenly distributed across classes and released under a Creative Commons Attribution 4.0 International license through Zenodo. This resource aims to bridge the gap in Spanish-language datasets for hate speech detection, aiding future research in content moderation, bias analysis, and the creation of inclusive artificial intelligence systems. 6. M. Krajči, S. Ljubić, I. Wolf, I. Štajduhar (Faculty of Engineering - University of Rijeka, Rijeka, Croatia) Multilingual Housekeeping Automatic Speech Recognition and Intent Classification  This paper presents a two stage, voice driven pipeline for hotel housekeeping that converts multilingual speech into English text and an intent label. We introduce an in-house dataset of 400 short housekeeping utterances, translated into Croatian, English, and Italian, and annotated with three intent classes and a detailed maintenance taxonomy. Whisper Large model is evaluated for both transcription in the source language and speech-to-English translation, with the translation setting further adapted using parameter-efficient fine-tuning with LoRA. Output quality is measured using standard error and overlap metrics, and the resulting texts are used for intent classification. For classification, we compare BERT and ELECTRA models using grid search and early stopping, reporting accuracy, F1 variants, top-k accuracy, and confusion matrices. Results indicate improved translation performance after LoRA adaptation and stronger intent classification performance for ELECTRA compared to BERT. 7. J. Dobruna, Z. Limani Fazliu, E. Hamiti, E. Mucolli (University of Prishtina, Prishtina, Kosovo), M. Volk (University of Ljubljana, Ljubljana, Slovenia) AI-Driven Prediction of Electric Field and Magnetic Flux Density in Electric, Hybrid, and Mild Hybrid Vehicles  The general public is increasingly exposed to low-frequency electromagnetic fields (LF-EMF) generated by a wide range of technologies, including household electrical appliances, power transmission systems, and renewable energy installations. In recent years, this exposure has further expanded with the rapid development of electric transportation technologies. Electric vehicles (EVs), hybrid electric vehicles (HEVs), and mild hybrid electric vehicles (MHEVs) incorporate high-voltage electrical systems, power electronics, and electric motors, which represent potential sources of electric fields and magnetic flux density for vehicle occupants. Use of electric vehicles has also steadily increased due to their lower environmental impact, as part of a general push towards greener transportation. This study investigates an artificial intelligence (AI)-driven approach for predicting electric field intensity and magnetic flux density in electric, hybrid, and mild hybrid vehicles. Supervised learning techniques, including regression models and Random Forest algorithms, are trained using measurement data collected under real-world operating scenarios. The performance analysis results showed that the proposed model can produce accurate electric field intensity and magnetic flux density estimation results for different types of vehicles. The proposed methodology offers a scalable and complementary approach to conventional measurement campaigns, supporting exposure assessment and safety evaluation in modern electric and hybrid transportation systems. petak, 29.5.2026 9:00 - 13:00, Camelia 2, Grand hotel Adriatic, Opatija 9:15 - 10:45 AI in Embedded and Edge Systems Predsjedatelj: Marko Horvat   1. D. Leko (University of Zagreb Faculty of Electrical Engineering and Computing, Zagreb, Croatia), F. Golubić, N. Čerkez (Rimac Technology d.o.o., Zagreb, Croatia), M. Vašak (University of Zagreb Faculty of Electrical Engineering and Computing, Zagreb, Croatia) Implementation of a Protective Set on a Microcontroller with a Floating-point Unit Support  This paper proposes a real-time implementation of a protective set that facilitates nonlinear system online control computations with a polytopic sub-approximation of the feasible set of system states and future control inputs. The proposed approach is particularly relevant for systems where permissible wear and aging limits are characterized by machine learning models processing large datasets, as these models introduce complex, non-linear constraints to the control problem. By pre-calculating the protective set of admissible states and feasible control sequences offline in the form of linear inequalities, the online burden for characterizing feasible future control inputs reduces to a non-iterative sequence of multiplication and summation operations that implements a matrix-vector product. This formulation is specifically tailored for microcontrollers equipped with a floating-point unit (FPU), as it replaces iterative nonlinear programming (NLP) solvers with deterministic matrix-vector multiplication and addition. 2. J. Ledenčan, J. Zidar, I. Vidović, T. Matić (Faculty of Electrical Engineering, Computer Science and Information Technology Osijek, Osijek, Croatia) Application of FPGA-Based AI Systems for Smart Agriculture: A Review Across Field Crops, Fruits, and Vegetables  With the recent development of smart agriculture, there is increased demand for efficient embedded computing platforms capable of operating under strict cost, energy, and latency constraints. Field-Programmable Gate Arrays (FPGAs) have emerged as a promising solution for embedded AI applications due to their reconfigurability, parallel processing capabilities, and energy efficiency. This review paper presents an overview of FPGA-based embedded AI systems applied in smart and precision agriculture. The overview includes recent studies on the implementation of AI models on FPGAs for agricultural applications, with a focus on specific crop types. We analyze and compare used FPGA boards, AI models, and tools, and assess their performance to identify current challenges and research directions for future FPGA-based AI systems across field crops, fruits, and vegetables. 3. D. Gookyi, E. Nansuuri, J. Awotwi, M. Wilson, R. Ahiadormey (CSIR-INSTI, Accra, Ghana), L. Chabala (University of Zambia, Lusaka, Zambia), O. Damba (University for Development Studies, Tamale, Ghana) A Low-Cost Hardware–Mobile Platform for Comprehensive Soil Monitoring, Crop Phenotype, and Image Dataset Acquisition  Affordable field tools often measure soil variables but fail to capture the structured context required for reusable, machine learning (ML)-ready datasets. This paper presents a low-cost, integrated hardware–mobile platform designed specifically for the acquisition of standardized multimodal datasets. The system utilizes an ESP32 sensing node interfaced with an RS485 multi-parameter soil probe, paired with a mobile application for live visualization, session-based logging, and contextual data entry. At each sampling point, seven physicochemical soil metrics, including moisture, Electrical Conductivity (EC), pH, temperature, Nitrogen (N), Phosphorus (P), and Potassium (K), are synchronously linked to crop observations (phenological traits), soil descriptors, and in-situ images. Field trials demonstrate the system's capability to generate cohesive, exportable datasets (CSV/PDF), validating its utility as a foundational tool for developing cost-effective agricultural ML models. 4. D. Gookyi, R. Tetteh (CSIR-INSTI, Accra, Ghana), R. Gyaang, C. Tengan (Bolgatanga Technical University, Bolgatanga, Ghana), P. Danquah, S. Bekoe (CSIR-INSTI, Accra, Ghana), S. Danso (GCTU, Accra, Ghana) Survey and Design Framework: Comparative Evaluation of Edge Computing Platforms for LowCost AI-Enabled Real-Time Soil Monitoring, Dataset Collection, Crop Recommendation, and Fertilizer Optimization in African Settings  This study offers a side-by-side analysis of affordable edge-computing hardware options designed for Artificial Intelligence (AI) powered soil monitoring systems in African farming contexts. It evaluates microcontroller units (MCUs), single-board computers (SBCs), and powerful AI accelerators across key aspects of digital soil management, including live soil sensing, data gathering and transmission, Tiny Machine Learning (TinyML) based crop suggestions, and deep learning for fertilizer recommendations. The review covers specifications that include Central Processing Unit (CPU) types, memory sizes, connection options, speed/latency, power use, and AI support for devices such as Arduino, ESP32, RP2040, Raspberry Pi, NVIDIA Jetson, and Arduino Portenta. Key findings indicate that no single device handles everything perfectly; a layered setup with multiple hardware components works best. Cheap MCUs shine for instant sensing, SBCs handle data crunching well, and Graphic Processing Unit/Neural Processing Unit (GPU/NPU) equipped hardware power up fertilizer tweaks and nutrient analysis from images. This work includes a practical guide that maps hardware to the needs of budget-limited African farms, as well as ideas for future work on TinyML improvements, field-ready setups, and efficient next-generation edge AI technology. 5. E. Guberović, I. Čavrak, I. Biuk (Faculty of Electrical Engineering and Computing, Zagreb, Croatia) FEDCODE: A Modular Framework for Evaluating Communication Compression in Federated Learning  Federated learning is a technique that facilitates distributed global model training without requiring raw data exchange. The training process runs on end devices, which subsequently transmit their locally trained model weights to a central aggregation server. In practical applications, the distributed configuration of these devices, along with the substantial size of contemporary model updates, creates a significant communication bottleneck. This article introduces FEDCODE, an extensible open-source framework designed for simulating and evaluating various statistical data compression strategies on the communication channel during the federated learning process. The framework is modular, allowing straightforward integration of new methods and datasets, and includes a collection of established statistical coders and standard dataset benchmarks. We demonstrate the use of the FEDCODE framework to showcase the possible reduction in total communication cost on the MNIST and CIFAR-10 datasets achievable with the provided statistical and adaptive coders. 6. M. Vukšić, J. Ćelić ( Faculty of Maritime Studies, University of Rijeka, Rijeka, Croatia), T. Bronzin, B. Prole, D. Adamec, A. Stipić (CITUS d.o.o., Zagreb, Croatia), D. Katović (Faculty of Kinesiology, University of Zagreb, Zagreb, Croatia) Hybrid Digital Twin and Deep Learning Framework for Monitoring Maritime Refrigerated Containers  Maritime refrigerated containers are a critical component of global cold-chain logistics; however, their onboard monitoring is still predominantly based on simple threshold alarms that react only after temperature deviations have occurred. Such reactive strategies provide limited situational awareness and insufficient time for effective intervention when cooling system degradation develops. This paper introduces a hybrid artificial intelligence framework for monitoring maritime refrigerated containers that combines physics-based modelling with data-driven anomaly detection. The framework follows a three-layer system architecture consisting of an IoT-enabled physical layer, a digital twin and AI analytics layer, and an operator interface supporting human-in-the-loop supervision. A lightweight thermodynamic digital twin represents expected thermal behaviour under varying operating conditions, while a deep learning autoencoder captures patterns of nominal operation and identifies deviations through reconstruction error analysis. Physics-based residuals and learning-based anomaly indicators are fused to improve detection robustness while limiting false alerts. System behaviour is examined using controlled fault scenarios generated in simulation and complemented by a small-scale practical data acquisition setup using REFCON connectivity. The results indicate that the hybrid approach can identify abnormal operating conditions prior to critical temperature excursions while remaining interpretable and suitable for edge deployment. 10:45 - 11:00 Pauza   11:00 - 13:00 Artificial Intelligence Applications and Other Topics - I  Predsjedatelj: Alan Jović   1. M. Burić (Hrvatska elektroprivreda d.d., Rijeka, Croatia) Identifying Structural Patterns in Short-Term Electricity Load Forecasting Using Explainable AI  Short-term electricity load forecasting is an important factor in reliable power system operation. Although deep learning models can achieve good predictive performance, their internal conduct is often difficult to interpret. This can limit practical use in real operational environments. This research paper explores whether explainable artificial intelligence (XAI) can be used to identify stable patterns in electricity consumption, rather than focusing solely on explaining individual model predictions. Publicly available hourly load data for the Croatian power system are used to train and evaluate three forecasting approaches, including a statistical base model and two neural network models. Input features are gained from past load values and basic temporal information. Model performance is evaluated using standard error metrics. Explainability methods are subsequently applied to analyze the contribution of individual features to the forecasting process. The analysis is performed across different models and multiple temporal perspectives, such as workdays and weekends. The results indicate that a specific features remain consistently important across models and contexts. This suggests the presence of structural consistencies in electricity consumption that are not bound to a specific model architecture. The proposed approach supports more transparent forecasting and provides interpretable insights into power system behavior. 2. F. Aziz, M. Žnidaršič, A. Osojnik , B. Ženko (Jožef Stefan Institute, Ljubljana, Slovenia) Addressing Skewed Distributions in Electricity Consumption Forecasting  Accurate forecasting of electricity consumption is crucial for smart grids, enabling dynamic matching of supply and demand in both domestic and industrial use. This paper focuses on predicting electricity usage for individual households over a two-day horizon, a scenario that could facilitate flexible consumption adjustments based on grid conditions showcasing methods that can also be applied in industrial applications. A significant challenge in this context is the presence of numerous peaks in consumption patterns, which are difficult to predict in terms of timing and magnitude. These peaks resemble outliers and introduce skewness into the target value distribution, ultimately degrading the performance of forecasting models. To address this issue, we empirically evaluate two approaches: data weighting and square root transformation. Our goal is to assess the effectiveness and limitations of these methods in mitigating the impact of skewed distributions on forecasting performance. 3. S. Baressi Šegota (Juraj Dobrila University of Pula, Pula, Croatia), N. Anđelić, V. Mrzljak (Faculty of Engineering - University of Rijeka, Rijeka, Croatia), M. Bobanović Dasko, I. Lorencin (Juraj Dobrila University of Pula, Pula, Croatia) Understanding the Energy Expenditure Predictors of Collaborative Robots Using Statistical and Machine Learning Methods  The use of collaborative robots is ever-growing, which increases the issue of energy use during prolonged operation. The goal of the presented research is the evaluation of the importance of various factors that may have an influence—such as different loads, movement types, and temperatures of the robot system. The information is contained within a publicly available dataset, "Dataset for Energy Assessment of Collaborative Robots" which contains data on the UR3e cobot from various experiments. In the presented research, the authors combine data from multiple experiments to obtain a single dataset containing 226,472 observations, allowing for cross-comparison of different experimental setups. Statistical methods including three-way factorial ANOVA with interaction effects, effect size analysis (eta-squared), post hoc comparisons (Tukey’s HSD), and Random Forest feature importance analysis are applied to determine the influence of selected factors on energy consumption. The main goal of this research is to determine which factors have the largest influence on the energy use of the cobot, which may serve as a starting point for a knowledge base of guidelines and best practices for the general setup, as well as the programming and control of collaborative robots in terms of lowering their energy use. Results revealed that temperature is the dominant predictive factor (80.84% feature importance), followed by movement type (18.91%), while load showed negligible influence (0.26%). 4. M. Boban (University of Split Faculty of Law, Split, Croatia), M. Gombar (Ministry of Defense of the Republic of Croatia, Zagreb, Croatia), M. Pejić Bach (Faculty of Economics and Business, University of Zagreb; Faculty of Commercial and , Zagreb, Croatia) From Susceptibility to Resilience: Validating User-Level Audit Support Measures for AI-Mediated Platforms  AI-mediated platforms increasingly shape users’ content exposure, choices, and participation, yet their evaluation often remains limited to model-centric metrics and system-level logs. This paper introduces two user-level measures intended to support the evaluation of trustworthy AImediated environments: the Algorithmic Resilience Scale (ARS; 12 items) and the Algorithmic Susceptibility Index (ASI; 12 items). A cross-sectional survey was conducted on a sample of 250 adult users using a 5-point Likert scale. Exploratory factor analysis with parallel analysis supported a two-factor structure corresponding to resilience and susceptibility. Both measures demonstrated high reliability, with Cronbach’s alpha coefficients of .91 for ARS and .87 for ASI. The two constructs were strongly and inversely related (r = -.64), supporting their conceptual distinction. Regression analyses further showed that ARS and ASI were meaningfully associated with trust-related outcomes, including perceived control, transparency salience, and response to AI intervention. The findings suggest that these measures can contribute to user-level evaluation in contexts such as explainable AI and recommender systems, particularly in pre- and post-intervention assessments and in the analysis of interface conditions related to control, transparency, and exposure diversity. Practically, the proposed measures provide a way to operationalize user resilience and susceptibility as measurable self-reported dimensions of human–AI interaction, while also identifying design features relevant to resilience-oriented platform evaluation, including data provenance, controllable personalization, exposure diversity, and salient error signaling. 5. R. Orlić (Croatian Academic and Research Network – CARNET, Zagreb, Croatia), S. Baressi Šegota (Juraj Dobrila University of Pula, Faculty of Informaticste, Pula, Croatia), A. Novak (Croatian Academic and Research Network – CARNET, Zagreb, Croatia), I. Lorencin (Juraj Dobrila University of Pula, Faculty of Informatics, Pula, Croatia) Croatian State Exam Benchmark: Evaluating Large Language Models on High School Croatian Language Proficiency  The rapid advancement of Large Language Models (LLMs) has transformed natural language processing, yet comprehensive evaluation of LLM capabilities in Croatian, a moderately resourced language, remains limited. Existing multilingual benchmarks often rely on machine-translated datasets that may not adequately capture Croatian’s linguistic complexity and cultural specificity. To address this, we propose a novel benchmark based on authentic Croatian state exams (high school final exams), providing rigorous assessment grounded in standardized educational materials. Our benchmark comprises multiple-choice questions spanning four categories: reading comprehension, literature, Croatian language (grammar and linguistics), and sentence completion. We investigate two methodological considerations: (1) whether all question types contribute equally to language proficiency assessment or if certain subsets provide sufficient evaluation coverage, and (2) the impact of prompt engineering, specifically comparing performance with Chain-of-Thought reasoning versus minimal prompting (simple answer selection). We evaluate multiple state-of-the-art LLMs, including both proprietary and open-source models, to establish performance metrics and identify current capabilities and limitations in Croatian language understanding. By providing this benchmark, we aim to advance Croatian language technology development, facilitate fair model comparisons across different model architectures and accessibility paradigms, and contribute to creating inclusive AI systems that serve diverse linguistic communities. 6. A. Kitanovski, M. Toshevska, S. Gievska (Faculty of Computer Science and Engineering, Skopje, Macedonia) Knowledge-assisted Prompting for Sentiment Transfer  This work presents a comparative evaluation of zero-shot prompting strategies for text sentiment transfer using large language models (LLMs). We explore four prompting variants that incorporate different forms of lexical guidance and refinement: (1) a baseline zero-shot prompt, (2) SentiWordNet-assisted prompting, (3) Style Knowledge Graph-assisted prompting, and (4) a double-pass method for improving fluency and tone. All approaches are tested using the LLaMA 3.1, Mistral 7B and DeepSeek LLM 7B language models without fine-tuning, with a focus on maintaining fluency, preserving content, and successfully shifting sentiment polarity. Evaluation across multiple automated metrics (SacreBLEU, BERTScore, perplexity, and sentiment classification accuracy) reveals distinct trade-offs between style strength and content fidelity 7. S. Papić, K. Josić, B. Kostelac (Algebra Bernays University, Zagreb, Croatia) Trustworthiness of AI-Generated Network Configurations  This paper presents a qualitative evaluation of AI-generated network configurations using a complex capstone topology derived from an undergraduate routing and switching course. The task combines multiple routing protocols, static routing, redistribution, and HSRP-based redundancy, making correct end-to-end behavior dependent on the interaction of several networking mechanisms rather than isolated device commands. ChatGPT, Gemini, and Claude were tested using a single self-contained prompt, and their outputs were evaluated against a validated reference configuration and functional failover tests. The results show that all three tools were generally able to generate syntactically valid and deployable configurations, but recurring problems appeared in redundancy behavior, static routing decisions, and preservation of the intended routing logic, often remaining hidden under nominal connectivity and becoming visible only during failure-sensitive validation. These findings indicate that successful deployment alone is not a sufficient indicator of correctness in integrated networking tasks and that AI-generated configurations still require expert validation, particularly when semantic consistency and failover behavior are critical. petak, 29.5.2026 15:00 - 17:00, Camelia 2, Grand hotel Adriatic, Opatija 15:00 - 16:45 Artificial Intelligence Applications and Other Topics - II  Predsjedatelj: Alan Jović   1. C. Quintana Reyes, A. Segura-Navarrete (Universidad del Bio-Bío, Concepción, Chile), C. Martinez-Araneda (Universidad Católica de la Santísima Concepción, Concepción, Chile), C. Vidal-Castro (Universidad del Bio-Bío, Concepción, Chile), P. Gómez-Meneses (Universidad Católica de la SSMa Concepción, Concepción, Chile) SafeTune – a Deep Learning-Based Application for Detecting Gender-Based Violence against Women in Song Lyrics  Purpose: This work describes the development of SafeTune, a music player that integrates deep learning-based detection of gender-based violence (GBV) in song lyrics within a socially motivated application context. SafeTune communicates with Spotify™ through its API and can automatically detect and skip songs associated with gender-based violence when operating in restrictive mode. Method and tools: A BERT-based model (BETO) was employed for Spanish text classification. The software followed an incremental development comprising four iterations of analysis, design, implementation, and testing in a lab environment. The system architecture was designed as a microservices structure and implemented using Docker containers, enabling functional testing and deployment. Results: The resulting web application accesses Spotify’s API to retrieve song lyrics through scraping and to detect gender-based violence against women. In restrictive mode, the system prompts users to validate the classification results. Conclusions: A minimum viable prototype of SafeTune, the dataset, and the BERT-based model are currently available. This proposes a solution that can support women's protection and parental control, while promoting educational awareness of respectful language. A modular, container-based design enables rapid incorporation of new detection models. The BETO classifier can potentially be improved to detect more nuanced forms of bias in language, including micromachismo, microsexism, and gender stereotypes. 2. I. Kovačević, I. Aleksi, T. Matić (Faculty of Electrical Engineering, Computer Science and Information Technology Osijek, Osijek, Croatia) Application of AI Across the Wine Production Process  Artificial intelligence (AI) is having a growing impact on the winemaking process, affecting everything from vineyard management to evaluation of wine quality. This review analyzes the present applications of AI across various stages of production and evaluates the contributions of existing approaches to consistency, efficiency, and traceability. Recent research demonstrates an increasing interest in digital and data-driven approaches, although the scope and depth of their implementation differ significantly. Studies are frequently limited in terms of data quality, technology compatibility, and methodological comparability. These limitations influence not only the reliability of research conclusions but also the practical applicability of AI throughout the production process. Based on these findings, the analysis suggests several directions for future research, including the need for uniform evaluation standards, larger datasets, and better integration of digital technologies into existing production workflows. The paper compares existing applications of AI within the winemaking process, highlighting both advancements and current limitations. 3. P. Prenc, I. Lorencin (Juraj Dobrila University of Pula, Faculty of Informatics, Pula, Croatia) Utilization of Semantic Routing for the Development of an Intelligent Student Support Chatbot  This paper presents Sembot, a prototype student-support chatbot centered on semantic routing for FAQ-style question answering. The implemented system combines local sentence embeddings (all-MiniLM-L6-v2) with cosine-similarity routing in a FastAPI backend and a lightweight web frontend. We focus on the implemented retrieval pipeline, threshold behavior, and practical deployment constraints. Experimental validation is exploratory: the current test suite contains 30 queries and is used to characterize failure modes rather than to claim production readiness. Results show limited performance (10/30 successful responses under the defined pass criterion), with strong category dependence and clear coverage gaps in the knowledge base. We therefore position the current system as an early-stage prototype and report concrete methodological limitations and improvement directions, including larger relevance-annotated test sets, established retrieval metrics, and explicit baseline comparisons. 4. P. Kuhar, A. Mandić (University North, Koprivnica, Croatia) Artificial Intelligence in Public Relations: A Bibliometric Analysis of Research Trends and Thematic Structures  This research conducts a bibliometric analysis of scientific literature on the application of artificial intelligence tools in public relations. The aim is to map the structure, thematic orientations, and development of this research area. The methodological approach involves systematically collecting and compiling bibliographic data from the Web of Science database. A total of 236 papers were analyzed using the following keywords: "artificial intelligence," "public relations," "strategic communication," and "communication analytics." The analysis examines publishing trends, thematic clusters, co-authorship networks, citation patterns, and keyword co-occurrence using bibliometric tools VOSviewer and Bibliometrix. Conceptual mapping facilitates the identification of dominant and peripheral research topics, as well as insights into their interconnectedness and evolution over time. The results provide a structured overview of scientific production, including the geographical and institutional distribution of contributions and patterns of scientific collaboration within the field. Furthermore, the analysis identifies research gaps, underrepresented thematic areas, and potential directions for future research. These findings contribute to a better understanding of the development and dynamics of AI research in public relations and serve as a valuable reference for further theoretical and empirical work in both academic and professional contexts. 5. F. Avramovski, M. Stojcheva, I. Mishkovski (Faculty of Computer Science and Engineering, Ss. Cyril and Methodius , Skopje, Macedonia) Intelligent Thesis Committee Recommendation System Using Relational Graph Convolutional Networks  Selecting appropriate thesis committee members is a critical academic process that directly impacts evaluation quality and student development. Traditional approaches often fail to consider the complex relationships between faculty expertise, prior mentorships, and research collaboration patterns. This paper presents an intelligent recommendation system for thesis committee assignment using Relational Graph Convolutional Networks (R-GCN). We construct a heterogeneous academic graph containing 98 professors and 5,858 theses, connected through five relation types: mentorship, committee membership (C2, C3), research authorship, and faculty collaboration. The system employs a three-layer R-GCN architecture with multi-task learning, simultaneously predicting suitable candidates for mentor, second member, and third member roles through dedicated prediction heads. We introduce an edge masking strategy during training to ensure the model generalizes unseen thesis proposals rather than memorizing existing assignments. Experimental evaluation demonstrates strong performance with 97% Hits@3 for mentor prediction, 93.09% for second member, and 96.31% for third member recommendations. The system achieves MRR scores of 0.81, 0.70, and 0.77 respectively, with AUC values exceeding 0.74 across all tasks. Results indicate that GNN-based approaches effectively capture academic network semantics for committee recommendation. 6. Y. Januzaj, E. Beqiri (University "Haxhi Zeka", Peja, Kosovo), A. Luma (South East European University, Tetovo, Macedonia) An AI-Based Framework for Academic Staff Performance Evaluation in Public Universities  Performance evaluation of academic staff is a key component of quality assurance and strategic management in higher education institutions. Traditional evaluation systems often rely on manually defined indicators and subjective weighting schemes, which may limit transparency, consistency, and objectivity. With the increasing availability of institutional data, artificial intelligence offers new opportunities for more data-driven and adaptive performance assessment. This paper proposes an AI-based framework for academic staff performance evaluation that integrates multiple quantitative indicators into a unified analytical model. The framework combines research productivity metrics, including publication counts, citation indicators, h-index, and i10-index, with teaching and institutional engagement features to construct multidimensional performance profiles. Machine learning techniques, such as clustering and ensemble-based models, are employed to identify performance patterns, estimate feature importance, and generate objective performance scores. The framework is evaluated using a real-world dataset from public universities. The results demonstrate that the proposed model can effectively distinguish between different performance groups and provide interpretable insights for institutional decision-making. The findings suggest that AI-based evaluation enhances fairness, transparency, and scalability compared to traditional approaches, and offers a practical decision-support tool for university management and quality assurance processes. 7. D. Leljak, E. Maruševec, A. Najev, S. Vojvodić, K. Vidović (Ericsson Nikola Tesla d.d., Zagreb, Croatia) LTE Data–Based Traffic Flow Estimation Using Machine Learning: A Case Study for the City of Matulji  In an era characterized by high demand for optimized urban traffic transport infrastructure and, at the same time, immense progress in artificial intelligence (AI) solutions, public governance seeks to leverage the latter to address the former. This publication presents a case study of a machine learning-based traffic monitoring system implemented in the context of the town of Matulji (Croatia). The proposed system formulates traffic monitoring as a supervised regression problem, where the number of vehicles on selected road segments is estimated from 4G/LTE performance counter data. This data source offers inherent advantages such as being privacy-preserving, readily available, and easily scalable globally, while also being strongly correlated with real-time vehicle traffic infrastructure load. Ground truth data used for training the machine learning model originates from vehicle counters temporarily installed in more densely urbanized parts of the Matulji municipality. A blueprint of the operative process that enables continuous and concrete metric-based quality assurance is also presented, including a discussion on cost efficiency and considerations of advantages and disadvantages.

Osnovni podaci:

Prijava/Kotizacija:

Studentski popust se ne odnosi na studente doktorskog studija.

OBAVIJEST AUTORIMA: Uvjet za objavu rada je plaćanje najmanje jedne kotizacije po radu. Autorima 2 ili više radova, ukupna se kotizacija umanjuje za 10%.

Darko Huljenić
Fakultet elektrotehnike i računarstva
Unska 3
10000 Zagreb, Hrvatska

E-mail: huljenicdarko@gmail.com 

Alan Jović
Fakultet elektrotehnike i računarstva
Unska 3
10000 Zagreb, Hrvatska

Tel.: +385 1 612 9548
E-mail: alan.jovic@fer.hr
 

Najbolji radovi bit će nagrađeni.
Prihvaćeni radovi bit će objavljeni u zborniku radova s ISSN brojem. Radovi na engleskom jeziku prezentirani na skupu bit će poslani za uključenje u digitalnu bazu IEEE Xplore.

Opatija je vodeće je ljetovalište na istočnoj strani Jadrana i jedno od najpoznatijih na Mediteranu. Ovaj grad aristokratske arhitekture i stila već više od 180 godina privlači svjetski poznate umjetnike, političare, kraljeve, znanstvenike, sportaše, ali i poslovne ljude, bankare, menadžere i sve kojima Opatija nudi svoje brojne sadržaje. 

Opatija svojim gostima nudi brojne komforne hotele, odlične restorane, zabavne sadržaje, umjetničke festivale, vrhunske koncerte ozbiljne i zabavne glazbe, uređene plaže i brojne bazene i sve što je potrebno za ugodan boravak gostiju različitih afiniteta. 

U novije doba Opatija je jedan od najpoznatijih kongresnih gradova na Mediteranu, posebno prepoznatljiva po međunarodnim ICT skupovima MIPRO koji se u njoj održavaju od 1979. godine i koji redovito okupljaju preko tisuću sudionika iz četrdesetak zemalja. Ovi skupovi Opatiju promoviraju u nezaobilazan tehnološki, poslovni, obrazovni i znanstveni centar jugoistočne Europe i Europske unije općenito.

Detaljnije informacije se mogu potražiti na www.opatija.hr i www.visitopatija.com.