ARTIFICIAL INTELLIGENCE Seminar
PROGRAM
Predavanja možete pratiti i online putem MITEAM stranice Seminara iz veštačke inteligencije:
https://miteam.mi.sanu.ac.rs/asset/DPP9i2jhvYzp8dmRe
Plan rada Seminara iz veštačke inteligencije za DECEMBAR 2025.
Sreda, 03.12.2025. u 19:00, Online
Viktor Pristaš, Pavol Jozef Šafárik University in Košice, Faculty of Science, Institute of Computer Science
INTEGRATING MACHINE LEARNING AND SIGNAL PROCESSING FOR ELECTRICAL POWER ANALYSIS
Machine learning provides powerful techniques for pattern recognition, signal classification, and clustering. In this contribution, we explore the integration of machine learning and signal processing to extract knowledge from diverse signal types across various domains. A particular focus is placed on the power industry, where the proposed methods classify or cluster levels of reactive power generation and consumption. The rising reactive power in modern electrical grids creates significant challenges, including energy losses, reduced efficiency, and potential equipment failure. The developed approaches aim to enhance monitoring and management of these effects, supporting more stable and efficient power systems.
Sreda, 10.12.2025. u 19:00, Online
Ivana Štajner-Papuga, Prirodno-matematički fakultet, Univerzitet u Novom Sadu
HORIZONTALNA FAZI RELACIJA U STATISTIČKOM TESTIRANJU
Fazi skupovi, uvedeni 1965. godine, predstavljaju snažan okvir za modelovanje neizvesnosti, neporeciznih podataka i lingvističkih informacija, te su tokom decenija postali osnova brojnih savremenih tehnologija. Njihova prirodna prilagodljivost ljudskom rezonovanju čini ih posebno pogodnim za savremene probleme odlučivanja, gde je neizvesnost gotovo neizbežna. Razvoj fazi statistike se nadovezuje na potrebu za proširenjem klasičnih statističkih procedura na fazi-vrednosne podatke, što je pravac koji potiče iz Zadehovog pristupa fazifikaciji verovatnoće i statistike. Današnja istraživanja obično se grupišu u tri šire kategorije: primenu fazi metoda na crisp podatke, proširenje klasičnih statističkih koncepata na fazi veličine preko α-preseka i kombinovanje oba pristupa u „fuzzy-to-fuzzy" smeru. Ovde se razmatra pristup koji pripada trećoj kategoriji i zasniva se na horizontalnoj fazi relaciji za poređenje fazi brojeva bez upotrebe α-preseka. Sada su i registrovana vrednost test statistike i kritična vrednost u formi fazi broja, što omogućava definisanje fazi stepena prihvatljivosti hipoteza. Takva formulacija prirodno objašnjava situacije u kojima nijedna hipoteza ne može biti u potpunosti prihvaćena ili odbačena. Metod je razmatran u univarijantnom i multivarijantnom okruženju, pri čemu horizontalno fazi poređenje pokazuje potencijal za preciznije modelovanje složenih realnih procesa. Konačan cilj je doprinos razvoju fleksibilnijih statističkih alata za analizu podataka koji usled određenog stepena neodređenosti i neizvesnosti bivaju predstavljeni fazi vrednostima.
Sreda, 17.12.2025. u 19:00, Online
-
Irena Vodenska, Administrative Sciences Department, Metropolitan College, Boston University, USA; Faculty of Computer Science and Engineering, Ss. Cyril and Methodius University, Skopje, N. Macedonia
WHEN AI GOES WRONG: TAXONOMY AND FINANCIAL IMPACT OF SYSTEMIC FAILURES IN GENERATIVE MODELS
The rapid deployment of large language models (LLMs) across sectors has amplified concerns about systemic failures that generate not only technical errors but also legal, ethical, and financial harm. This paper analyzes real-world incidents involving LLMs and other generative AI systems, with a particular focus on cases that result in measurable financial consequences. Drawing on a curated news-based dataset of more than fifty AI incidents, we develop a structured framework that links concrete failure modes, such as hallucinations, reasoning errors, bias and discrimination, privacy and security failures, operational misuse, and insecure code generation, to ten categories of financial loss, including legal liability, regulatory fines, market loss, and reputational damage. We further classify incidents by industry domain (e.g., law, healthcare, news, academia, social impact, AI services), revealing distinct sectoral risk patterns and pathways from technical flaws to downstream harm. Building on this analysis, we propose a taxonomy of systemic failures and an associated catalog of mitigation strategies that span the full AI lifecycle, including data collection and preprocessing, model training and alignment, deployment-time guardrails, post-hoc auditing, and governance and compliance mechanisms. Finally, we outline the design of an AI agent for continuous systemic-failure monitoring, intended to support practitioners, regulators, and system architects in proactively managing AI risks and building more trustworthy, financially resilient AI ecosystems.
Sreda, 24.12.2025. u 19:00, Online
Palash Dutta, Department of Mathematics, Adjunct Faculty, Centre for Atmospheric Studies, Dibrugarh University, Assam, India
THE EVOLUTION OF HYPERBOLIC FUZZY SET AND ITS APPLICATION IN CRIMINAL LINKAGE ANALYSIS
Criminal investigations are inherently complex, often plagued by ambiguous, incomplete, and conflicting information that challenges traditional analytical methods. Crime linkage, the process of identifying crimes committed by the same offender, is a critical task that requires sophisticated tools to handle such profound uncertainty. While fuzzy sets and their extensions like intuitionistic fuzzy sets have been applied to model uncertainty in crime data, they can struggle with the simultaneous presence of multiple, independent types of uncertainty in evidential attributes. This paper explores the evolution of fuzzy set theory into the domain of Hyperbolic Fuzzy Sets (HyFSs), which provide a more nuanced mathematical structure for capturing and computing with complex uncertainty by assigning membership and non-membership degrees that are not necessarily complementary. We propose a novel similarity measure specifically designed for HyFSs, capable of effectively quantifying the resemblance between multiple crime profiles. The superiority of this measure over existing models is demonstrated through comparative numerical examples. Furthermore, its practical efficacy is validated through a detailed case study on criminal linkage, where it is used to identify connected series of offenses based on modus operandi and behavioral patterns. A systematic methodology for integrating HyFSs into the psychological profiling of offenders is also presented. The proposed framework offers investigators a powerful and mathematically robust tool for solving complex decision-making problems in criminology and forensic science.
Ovaj onlajn seminar nastao je kao nastavak sastanka "Serbian AI Meeting" i zamišljen je da na njemu istraživači iz Srbije i iz dijaspore, kao i istraživači sa univerzteta, naučnih instituta i iz prakse predstavljaju naučne teme i rezultate iz oblasti veštačke inteligencije.
Link za svako pojedinačno predavanje biće dostavljen dan pre održavanja predavanja.
Andreja Tepavčević
Rukovodilac seminara
Biljana Stojanović
Sekretar seminara
