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 JUN 2025.

Sreda, 04.06.2025. u 18:00, Online
Tome Eftimov, Computer Systems Department, Jožef Stefan Institute, Ljubljana
LEVERAGING BENCHMARKING DATA FOR AUTOMATED OPTIMIZATION
At the start of 2022, the evolutionary computation community published a call for action highlighting significant issues with metaphor-based metaheuristics in black-box optimization (BBO): useless metaphors, limited novelty, and biased experimental validation. This talk presents recent benchmarking advances for robust and reliable results and meta-learning approaches for algorithm selection. We focus on two methods: i) selecting representative data instances to generalize study findings, and ii) using algorithm footprints to identify easy or challenging problem instances based on landscape characteristics. Ultimately, the goal is a paradigm shift toward reducing resource waste and duplicated efforts, accelerating progress, and enabling effective automated algorithm configuration and selection through transferable insights.

Utorak, 04.06.2025. u 19:00, svečana sala Ogranka SANU u Novom Sadu, Nikole Pašića 6 i Online
Miloš Savić, Departman za matematiku i informatiku, Prirodno-matematički fakultet, Univerzitet u Novom Sadu
HUBNESS AND LOCAL INTRINSIC DIMENSIONALITY FOR DYNAMIC GRAPH EMBEDDINGS BASED ON RANDOM WALKS
Large-scale complex systems present in nature and society are typically dynamic. Graph-structured data describing complex dynamical systems can be used to construct various types of useful predictive models based on machine learning (ML) techniques. There are three different approaches to build such models: (1) by using graph-native ML algorithms, (2) by applying ML algorithms designed for tabular data to graph embeddings, and (3) by graph neural networks. The focus of this talk will be on dynamic graph embedding methods. In contrast to graph neural networks that provide task-specific predictive models for feature-rich attributed graphs, graph embedding algorithms provide a task-agnostic approach to learn general-purpose graph representations that can be exploited in a variety of ML tasks. We will present novel dynamic graph embedding methods based on random walks developed within the TIGRA project ("Graphs in Space and Time: Graph Embeddings for Machine Learning in Complex Dynamical Systems"; supported by the Science Fund of the Republic of Serbia). In contrast to the existing approaches, the novel TIGRA methods take into account two important structural properties of nodes: hubness and local intrinsic dimensionality (LID). After empirically showing that hubness and LID measures tend to correlate with intrinsic embedding qualities, we have designed random walk sampling mechanisms with personalized hyper-parameters and transition probabilities strategically influenced by hubness and LID. The experimental evaluation of dynamic graph embedding methods based on such biased random walk sampling shows that they produce comparable or noticeably better embeddings than the state-of-the-art methods. At the end of the talk, we will discuss research activities planned for the second half of the TIGRA project.

Sreda, 11.06.2025. u 19:00, Online
Dragiša Mišković i Nikola Milošević, Istraživačko-razvojni institut za veštačku inteligenciju Srbije
GRUPA ZA HCI I PROJEKAT VERIFAI
Prvi deo predavanja biće posvećen aktivnostima HCI grupe na Institutu. U drugom delu predavanja pričaćemo detaljnije o projektu VerifAI. To je je otvoreni sistem za odgovaranje na pitanja zasnovan na dokumentima, razvijen kroz saradnju Istraživačko-razvojnog instituta za veštačku inteligenciju Srbije i kompanije Bayer, sa fokusom na biomedicinsku oblast. Posebno je usmeren na detektovanje halucinacija - situacija kada veštačka inteligencija generiše uverljive, ali netačne ili izmišljene odgovore predstavljene kao činjenice - i proveru tačnosti odgovora. Koristeći napredne tehnike pretrage i generisanja, VerifAI omogućava korisnicima pristup tačnim i proverljivim informacijama iz relevantne naučne literature, prvenstveno iz biomedicinskih izvora poput PubMed-a. Sistem automatski pronalazi i navodi izvore na koje se odgovori oslanjaju, a modeli za verifikaciju tvrdnji otkrivaju i označavaju potencijalne netačnosti ili dezinformacije. Na ovaj način, VerifAI doprinosi većem poverenju u primenu veštačke inteligencije u biomedicinskim istraživanjima.

Sreda, 18.06.2025. u 19:00, Online
Ranka Stanković i Mihailo Škorić, Rudarsko-geološki fakultet, Univerzitet u Beogradu
PROJEKAT TESLA (Text Embeddings - Serbian Language Applications): REZULTATI PRVE GODINE RAZVOJA JEZIČKIH MODELA ZA SRPSKI JEZIK
Tokom prve godine realizacije projekta TESLA, koji finansira Fond za nauku Republike Srbije, uspešno su postavljeni temelji za razvoj naprednih jezičkih modela za srpski jezik. Izvršena je detaljna analiza savremenih pristupa i definisana je inicijalna metodologija, nakon čega je usledilo sistematsko prikupljanje, klasifikacija, čišćenje i pretprocesiranje obimnih jezičkih podataka za srpski jezik. Predstavićemo ključne skupove podataka: neobeležene, automatski obeležene i ručno korigovane kolekcije. Na prezentaciji će biti reči i o razvoju predobučenih, statičkih i dinamičkih jezičkih modela, uključujući odabir i istraživanje arhitektura dubokog učenja prilagođenih srpskom jeziku. Značajne aktivnosti bile su usmerene na diseminaciju, kreiran je veb portal, ostvareno aktivno prisustvo na društvenim i naučnim mrežama, podaci i modeli su učinjeni dostupnim na relevantnim platformama (HuggingFace, Github, Zenodo), organizovane su dve radionice i objavljeno više naučnih radova.

Sreda, 25.06.2025. u 19:00, Online
Smiljana Antonijević Ubois, Illinois Institute of Technology
ADDRESSING HISTORICAL AND CONTEMPORARY INEQUALITIES IN LOW-RESOURCE LANGUAGES
The development of AI-based language models has predominantly focused on high-resource languages, reinforcing historical inequalities and marginalizing low-resource languages (LRLs) such as Serbian. While English, Mandarin, and other widely spoken languages have sizeable datasets, computational resources, and AI development funding, LRLs face significant marginalization in the age of AI. LRLs struggle with underrepresentation in AI applications, impacting their digital sustainability, accessibility, and global visibility (Joshi et al., 2020).
This paper presents a case study of the Serbian language, an LRL characterized by its rich dialectical diversity and dual alphabetic systems. The paper examines historical and contemporary barriers to development of AI tools and resources for Serbian, and proposes some possible solutions based on preliminary results of a study that included semi-structured, in-depth interviews with 10 scholars-- computational linguists, philosophers of language, AI developers, and digital humanists--involved in developing computational tools and resources for the Serbian language.

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