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 MAJ 2026.

Sreda, 06.05.2026. u 19:00, Online
Momir Adžemović, Matematički fakultet, Univerzitet u Beogradu
PUT KA PRAĆENJU OBJEKATA ZASNOVANOM NA DETEKCIJI BEZ HEURISTIKA
Praćenje više objekata (multi-object tracking, MOT) najčešće se rešava kroz praćenje po detekciju paradigmi, gde se objekti prvo detektuju u svakom frejmu, a zatim povezuju kroz vreme kako bi se formirale trajektorije. Iako je ovaj pristup efikasan i modularan, on se u praksi i dalje uglavnom oslanja na linearne modele kretanja, poput Kalmanovog filtera, i ručno definisane heuristike za asocijaciju, koje postaju nepouzdane u scenarijima sa nelinearnim kretanjem, sličnim objektima i šumom u detekcijama.
Prvi deo istraživanja bavi se modelima kretanja i pokazuje da se klasični filteri mogu zameniti naučenim alternativama zasnovanim na dubokom učenju. U tom kontekstu razmatramo i Bajesovske i end-to-end filtere, koji uče dinamiku kretanja objekata direktno iz podataka i time bolje modeluju složeno kretanje i šum u detekcijama. Na ovaj način model kretanja prestaje da bude ručno podešen heuristički modul i postaje komponenta koja se uči iz podataka u okviru procesa praćenja.
Drugi deo istraživanja bavi se problemom asocijacije i pokazuje da se i sam proces povezivanja može naučiti iz podataka. Umesto ručno definisanih funkcija cene asocijacije, asocijacija se formuliše kao problem predviđanja veza u bipartitnom grafu između trajektorija objekata i njihovih detekcija u trenutnom frejmu. Model tada uči da na osnovu geometrije, dinamike kretanja i izgleda objekata direktno predviđa verovatnoće veza, čime asocijacija prestaje da bude zasnovana na heuristikama.
Rezultati na referentnim skupovima podataka pokazuju da oba pravca donose konzistentna poboljšanja u scenarijima sa kompleksnim kretanjem i sličnim izgledom objekata. Kroz ovu seriju istraživanja pokazujemo da se ključne komponente praćenje po detekciji pristupa — prvo modelovanje kretanja, a zatim i asocijacija — mogu sistematski zameniti modulima koji uče iz podataka, što vodi ka robusnijim, fleksibilnijim i na sistemima koji se manje oslanjaju na heuristike.

Sreda, 13.05.2026. u 19:00, Online
Tatjana Jakšić Krüger, Matematički institut SANU
PARALLEL VARIABLE NEIGHBORHOOD SEARCH: FROM STRUCTURED SEARCH TO MULTI-AGENT SYSTEMS
In an era dominated by data-driven approaches, it is often assumed that effective search requires learning from data. We revisit Variable Neighborhood Search (VNS), a classical trajectory-based metaheuristic and argue that its performance relies on a combination of problem-informed design and systematic neighborhood exploration guided by a fixed neighborhood-change rule. On one hand, VNS incorporates structural knowledge through the definition of neighborhood systems and solution representations, reflecting the expertise of the algorithm designer. On the other hand, during execution, the neighborhood-change rule systematically redirects the search trajectory in response to local search outcomes -- without modifying the rule itself. Focusing on parallel VNS, we adopt the established High-Level Taxonomy of Parallel Metaheuristics (HTPM) and the complementary Implementation Taxonomy (ITPM) to analyze different parallelization strategies. These frameworks characterize algorithms along dimensions such as decomposition (independent vs. cooperative search), control (centralized vs. decentralized), communication (synchronous, asynchronous, or knowledge-based), solution diversity, granularity, and computing architecture. Within this perspective, parallel VNS is not merely a speed-up of the sequential algorithm, but a family of multi-search methods whose behavior depends fundamentally on the selected HTPM/ITPM configuration. Independent and cooperative strategies enable qualitatively different modes of parallel search, affecting both the exploration-exploitation balance and the robustness of the overall method. We show that these taxonomic dimensions naturally correspond to established AI concepts -- including inductive bias, exploration-exploitation tradeoff, and multi-agent coordination -- providing a unified perspective on parallel VNS as an AI method.
Zajednički sastanak sa Seminarom za računarstvo i primenjenu matematiku MISANU.

Sreda, 20.05.2026. u 19:00, Online
İlker Özçelik, Eskisehir Osmangazi University, Türkiye
RESILIENCE BY DESIGN: NAVIGATING THE INTERSECTION OF AI AND CYBERSECURITY
The advancement of distributed cyber-physical systems requires an integrated approach that combines sophisticated network defense and secure artificial intelligence. The Intelligent Systems Security Lab (ISSLab) addresses these challenges through a dual-pillar research approach focused on the intersection of machine learning and network security.
The first pillar, AI for network security, utilizes deep learning and Graph Neural Networks to detect large-scale anomalies such as botnets and distributed denial-of-service attacks within network environments. The second pillar investigates the security of artificial intelligence systems by analyzing vulnerabilities to adversarial attacks and developing robust mitigation frameworks to ensure system integrity. To reinforce these pillars, we utilize distributed edge authentication systems and blockchain-integrated protocols that ensure identity verification and immutable provenance across decentralized nodes. A significant dimension of our work investigates the identification and mitigation of deception, moving from the study of attacks performed to deceive network intrusion detection systems to current efforts in neutralizing adversarial attacks against artificial intelligence models. Building on these findings, we are also exploring how such technical manipulations and the broader digital life influence social dynamics.

Četvrtak, 27.05.2026. u 12:00, Online
Raka Jovanović, Qatar Environment and Energy Research Institute
BEYOND THE BLACK BOX: BRIDGING THEORETICAL PROMISE AND PRACTICAL REALITY IN APPLIED NEURAL NETWORKS
This lecture bridges the gap between the theoretical promise of neural networks and their practical performance in real-world forecasting tasks. Drawing on case studies in air quality forecasting, graph combinatorial optimization, crop import price prediction, food delivery demand forecasting, and wildfire spread modeling, we show that architectural complexity alone is rarely the key to success. Across these domains, simple feature engineering—such as value-adaptive smoothing, rolling statistics, and lagged variables—often improves accuracy more than switching from LightGBM to LSTMs or Transformers. Tree-based and robust regression models consistently match or outperform deep networks, especially when data is noisy, sparse, or contains outliers. Moreover, standard loss functions frequently misalign with operational goals, making asymmetric or robust losses essential for practical deployment. The talk offers a pragmatic framework for choosing models based on data constraints, computational resources, and domain-specific objectives, rather than theoretical elegance alone.
Zajednički sastanak sa Seminarom za računarstvo i primenjenu matematiku MISANU.

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