MISANU

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 APRIL 2026.

Sreda, 01.04.2026. u 19:00, Online
Milica Gašić, Univerzitet Hajnrih Hajne, Dizeldorf, Nemačka
ANALOGIJE, BOL I GEOMETRIJA INTELIGENCIJE: TRI POGLEDA NA (NENADGLEDANO) UČENJE
Nedavni napredak u velikim jezičkim modelima i veštačkim agentima ponovo je otvorio osnovna pitanja o tome kako se inteligencija uči, strukturira i predstavlja. U ovom preglednom predavanju predstaviću tri komplementarne perspektive na ova pitanja. Prvo, govoriću o šumovitim oznakama (noisy labels), razmatrajući kako modeli mogu učiti i generalizovati uprkos nesavršenim ili nepreciznim podacima, koristeći se analogijama. Drugo, predstaviću kako inspiracija ljudskim bolom i averzivnim signalima može doprineti efikasnijem učenju putem potkrepljenja (reinforcement learning), posmatrajući bol kao mehanizam za oblikovanje dinamike učenja, a ne samo za optimizaciju nagrade. Treće, pokazaću šta geometrija prostora reprezentacije u velikim jezičkim modelima otkriva o preprilagođavanju (overfitting), konvergenciji i, na kraju, o „grokovanju" (dubokom razumevanju). Iako se razlikuju po fokusu, ove perspektive zajedno doprinose dubljem razumevanju prirode nenadgledanog učenja.

Utorak, 07.04.2026. u 14:15, Knez Mihailova 36, sala 301f i Online
Goce Trajčevski, Department of Electrical and Computer Engineering, Iowa State University, USA
MOBILITY AND AI IN MOLECULAR DYNAMICS
Evolving molecular systems are at the heart of multiple challenges across chemistry, biology, biophysics, materials science, etc. These systems consist of large numbers of atoms belonging to interacting molecules, whose collective dynamics govern critical phenomena such as drug discovery, protein folding, and drug–target interactions. Due to cost and safety constraints associated with physical experimentation, particularly in early-stage exploration, Molecular Dynamics Simulations (MDS) have become an essential computational tool – generating high-resolution spatio-temporal trajectories by simulating atomic motion in three-dimensional space over time. However, MDS outputs massive volumes of spatio-temporal trajectory data, placing severe pressure on storage, query processing, and downstream analysis pipelines. Complementary to these challenges, the majority of the analytics works targeting molecular interactions still treat the interacting molecules as static graphs, ignoring the atomic motions.
We postulate that mobility is an inherent component of molecular dynamics and that many of the queries and analytical quests of interest should properly consider it and, with proper adaptation, capitalize on algorithmic findings from the field of Mobile Data Management (MDM). Clearly, it is not a straightforward task to adapt the existing MDM results to MDS data and there are unique challenges to be addressed. This talk will present some recent results targeting the specific problem of Hydrogen Bond (HB) formation during chemical reactions. As it turns out, it is not only the formation of an HB bond but also its persistence that is of interest in many applications such as drug discovery. The talk will present two complementary perspectives: query processing and deep learning based prediction of HB formation. In addition, we will discuss the global issue of truthfulness in the context of AI use in Molecular Dynamics. The talk will also discuss further challenges and opportunities arising from elevating the mobility to the first-class property of chemical reactions.
Zajednički sastanak sa Seminarom za računarstvo i primenjenu matematiku MISANU.

Sreda, 08.04.2026. u 19:00, Online
Ivan Damnjanović, Elektronski fakultet, Univerzitet u Nišu
REINFORCEMENT LEARNING FOR GRAPH THEORY: A UNIFIED FRAMEWORK FOR EXTREMAL PROBLEMS
The opening part of the lecture will offer a brief overview of the most significant results of the LZWK project, funded by the Serbian Science Fund. The remainder of the lecture will be devoted to the presentation of a research paper on reinforcement learning that is currently being developed. Reinforcement learning (RL) is a branch of machine learning focused on developing models that learn effective decision-making strategies through interaction and experience. In a recent influential paper, Wagner demonstrated that the Deep Cross-Entropy method from RL can be applied to problems in extremal graph theory by reformulating them as combinatorial optimization tasks. This idea sparked growing interest in the area, leading to a variety of refinements and extensions of Wagner’s original framework, as well as the development of RL environments tailored specifically to graph-theoretic problems. As a result, several problems in extremal graph theory have already been successfully studied using RL-based methods.
This lecture will introduce Reinforcement Learning for Graph Theory (RLGT), a new framework that unifies and organizes these earlier efforts.
RLGT is designed to handle both undirected and directed graphs, with optional support for loops and an arbitrary number of edge colors. The framework offers efficient graph representations and is intended to support future research in RL-based extremal graph theory through improved computational performance and a modular, extensible architecture.

Četvrtak, 09.04.2026. u 12:00, svečana sala Ogranka SANU u Novom Sadu, Nikole Pašića 6 i Online
Goce Trajčevski, Department of Electrical and Computer Engineering, Iowa State University, USA
ENHANCING URBAN REGION REPRESENTATION VIA ADAPTIVE RISK-AWARE CONSENSUS LEARNING
In a typical urban environment, there are various kinds of regions with distinct characteristics like, for example, residential neighborhoods, business districts, industrial zones, entertainment zones, etc. Each of these regions exhibits significant differences in terms of spatial structure, functional distribution and occupation frequency with inhabitants. In recent years, learning region representation with (deep) neural networks has become crucial for improving urban planning and influencing the development of smart cities. Accurately depicting the characteristics of urban regions can be used to facilitate various downstream tasks – e.g., crime prediction, traffic flow forecasting, socio-economic aspects estimation, etc. However, despite the notable achievements, the existing practices still face certain challenges: (1) When multiple views contain distinct semantic information, ignoring reliability and possibly inadequate collection differences (e.g., data missingness) among those views may degrade the representation robustness. (2) Consensus semantics extracted from different views are often fused in a simplistic manner, without considering the uniformity of embeddings (quality variations) and the complementarity between views.
In this talk, we will present our recent results targeting such challenges – based on a novel Adaptive Risk-aware Consensus Learning (ARC) solution for urban region embeddings. Specifically, we design both local- and region-level masking within the inter-view representation, following the paradigm of masked autoencoders, to better handle uncertainty risks. More importantly, we introduce a self-weighted contrastive mechanism in consensus learning to achieve maximum alignment and mitigate degradation. To enhance the uniformity of embeddings, we employ entropy, ensuring the diversity and complementarity of information. Ultimately, we apply the learned embeddings to down-stream tasks, demonstrating remarkable improvements compared to several representative baselines.
Zajednički sastanak sa Novosadskim seminarom MISANU.

Utorak, 14.04.2026. u 14:15, Knez Mihailova 36, sala 301f i Online
Jesús Medina Moreno, Department of Mathematics, University of Cadiz, Spain
FORMAL CONCEPT ANALYSIS, MACHINE LEARNING AND LOGIC
In this talk, we will see how formal concept analysis (FCA) can be a mathematical tool to machine learning and logic.
FCA was introduced in the eighties by Rudolf Wille as an application of lattice theory in order to provide a mathematical tool to obtain information from data bases. Different theoretical developments have been investigated to increase its flexibility and applicability.
Nowadays, FCA is attracting the attention of numerous researchers around the world.
Zajednički sastanak sa Seminarom za računarstvo i primenjenu matematiku MISANU.

Sreda, 15.04.2026. u 19:00, Online
María Eugenia Cornejo Piñero, Department of Mathematics, University of Cadiz, Spain
MULTI-ADJOINT FORMAL CONCEPT ANALYSIS FOR KNOWLEDGE DISCOVERY
This presentation will provide a wide view on multi-adjoint formal concept analysis, which is a fuzzy mathematical tool for managing and extracting knowledge from relational databases. We will show some relevant theoretical definitions, properties, and important tasks where this theory has successfully been applied, such as renewable energies and digital forensics.

Sreda, 22.04.2026. u 19:00, Online
Marina Tropmann-Frick, Faculty of Computer Science and Digital Society, Hamburg University of Applied Sciences
RESPONSIBLE AI
Modern AI systems achieve strong performance, but accuracy alone does not guarantee reliability, safety, or trust. This talk examines Responsible AI as a core requirement in the design and deployment of machine learning systems, emphasizing the need to move beyond performance-centric evaluation.
We focus on four key properties: fairness, privacy, robustness, and explainability. The talk presents a structured approach to evaluating these properties using measurable criteria, illustrating how responsibility can be assessed alongside traditional metrics.

Sreda, 29.04.2026. u 19:00, Online
Muzafer Saračević, Univerzitet u Novom Pazaru
PRIMENA MAŠINSKOG UČENJA I VEŠTAČKE INTELIGENCIJE U EVALUACIJI KRIPTOGRAFSKIH I STEGANOGRAFSKIH METODA
Primena mašinskog učenja (ML) i veštačke inteligencije (AI) u analizi i proceni kriptografskih i steganografskih metoda pokazuje značajan potencijal u unapređenju njihove bezbednosti, robusnosti i performansi. Tradicionalni pristupi proceni kriptografskih algoritama i steganografskih tehnika uglavnom se zasnivaju na teorijskoj analizi, statističkim testovima i eksperimentalnim napadima, dok savremeni ML/AI modeli omogućavaju automatizovanu detekciju obrazaca, anomalija i potencijalnih ranjivosti u šifrovanim i skrivenim podacima.
U našim prethonim istraživanjima, posebna pažnja posvećena je upotrebi veštačke inteligencije u automatskom prepoznavanju anomalija i optimizaciji parametara kriptografskih i steganografskih sistema. Razmatraju se različiti modeli koji se primenjuju u kriptoanalizi i steganalizi radi procene stepena sigurnosti, entropije i otpornosti na napade. U oblasti kriptografije ispitivana je mogućnost da se, isključivo na osnovu šifroteksta (ciphertext-a), identifikuje korišćeni algoritam enkripcije primenom ML/AI metoda. Paralelno s tim, u oblasti steganografije ispitano je da li algoritmi mašinskog učenja mogu automatski detektovati da li je određena slika modifikovana, odnosno da li sadrži skrivenu poruku. Posebnu pažnju smo posvetili primeni kombinatornih problema i teoriji brojeva u efikasnom generisanju kriptografskih ključeva.
Rezultati ukazuju da integracija ML/AI tehnika značajno doprinosi unapređenju procesa evaluacije, omogućavajući bržu i precizniju identifikaciju sigurnosnih propusta, ali istovremeno otvaraju i nova pitanja u vezi sa ranjivostima sistema na inteligentne adaptivne napade. Zaključuje se da kombinacija tradicionalnih kriptografskih principa i savremenih AI pristupa predstavlja perspektivan pravac daljih istraživanja u oblasti informacione bezbednosti.

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