MISANU

PROGRAM

Plan rada Seminara iz veštačke inteligencije za NOVEMBAR 2024.

Registraciona forma za učesće, i link na predavanje ako ste već registrovani:
https://miteam.mi.sanu.ac.rs/asset/CW5nJWDSEZDj7p32p
Ukoliko želite samo da gledate predavanje bez mogućnosti aktivnog učešća, prenos će biti dostupan na:
https://miteam.mi.sanu.ac.rs/asset/4LNW8WtML7rLKojoz
Na ovom linku se mogu pronaci kratka uputstva na srpskom i engleskom:
https://miteam.mi.sanu.ac.rs/asset/Kc7qJtEvoMFx9MFnz

Sreda, 05.11.2024. u 14:15, Online
Arutyun Avetisyan, Andrey Belevantsev, Yuri Markin, Institute for System Programming of the Russian Academy of Sciences (ISP RAS)
DEVELOPMENT AND VERIFICATION OF ARTIFICIAL INTELLIGENCE TECHNOLOGY
The report will primarily focus on ISP RAS research for trusted AI technology, including trusted frameworks for training neural networks, the theory and practice of creating models that can withstand attacks, preventing model aging, searching for machine learning vulnerabilities, and others.
The second part of the report will be devoted to the development of secure software. The report will present technologies developed at ISP RAS for secure software development (SDLC), which are necessary for creating efficient and secure software of any type, including artificial intelligence. Among them are approaches to program analysis, including static, dynamic analysis and fuzzing, secure compilation methods, and attack detection methods.The final part of the report will be devoted to research on digital watermarks, which ISP RAS conducts jointly with the Steklov Mathematical Institute of the Russian Academy of Sciences. The rapid development of AI poses new challenges in protecting training datasets, as well as trained neural network models, from anonymous theft.
Therefore, on the one hand, it is necessary to guarantee the possibility of establishing the fact of content synthesis, and on the other hand, to prevent the creation of deepfakes based on it.
Zajednički sastanak sa Seminarom za računarstvo i primenjenu matematiku.

Sreda, 06.11.2024. u 19:00, Online
Jovan Blanuša, IBM Research Europe, Zurich, Switzerland
EFFICIENT MACHINE LEARNING OVER RELATIONAL DATA
Various forms of real-world data, such as social, financial, and biological networks, can be represented using graphs. An efficient method of analysing this type of data is to extract subgraph patterns, such as cliques, cycles, and motifs, from graphs. For instance, finding cycles in financial graphs enables the detection of financial crimes such as money laundering and circular stock trading. In addition, extracting cliques from social network graphs enables the detection of communities and could help predict the spread of epidemics. However, extracting such patterns can be time-consuming, especially in larger graphs, because the number of patterns can grow exponentially with the graph size. Therefore, fast and scalable parallel algorithms are required to make the enumeration of these subgraph patterns tractable for real-world graphs.
In this talk, we will first talk about how to efficiently implement algorithms for mining cliques and cycles in graphs. Then, we introduce Graph Feature Preprocessor, which leverages the developed fast graph pattern mining algorithms to expand the feature set of financial transactions by enumerating well-known money laundering and financial fraud subgraph patterns. When used in combination with gradient-boosting-based machine learning models, the expanded feature set produced by the library enables significant improvements in prediction accuracy for the problems of money laundering and phishing detection.
Furthermore, the efficiency of the underlying graph pattern mining algorithms enables this library to operate in real time.

Sreda, 13.11.2024. u 19:00, Online
Valerijan Matvejev, Polytech Nantes, University of Nantes, France; School of Electrical Engineering, University of Belgrade; BioSense Institute, University of Novi Sad
CREATING HARMONY: ADVANCING AI AND ITS IMPACT ON RESEARCH
Visual attention is a cross-disciplinary topic in the field of computer vision that has been particularly interesting to researchers in the 21^st century. In this research, the focus was on exploring the complex relationship between human covert and overt attention and human memory on static images. Through a conducted scientific experiment, eye-tracking data was gathered from participants, which was then used to train a machine learning model, aimed at predicting human memory on a trial-by-trial basis with great accuracy. The research aimed to provide fresh insights into the intricate workings of the human mind and its visual system, while introducing an innovative approach to applying machine learning models in various practical contexts, such as predicting memory patterns, diagnosing attention and memory disorders, evaluating psychological readiness, and exploring many other yet-to-be-discovered areas.

Sreda, 20.11.2024. u 19:00, Online
Željko Tekić, Graduate School of Business, HSE University, Moscow
A FRAMEWORK FOR UNDERSTANDING AI TRANSFORMATION: COMPANY ARCHETYPES AND STRATEGIC CHALLENGES
As interest in AI grows within the business world, so does the complexity of concepts, factors, and challenges associated with AI transformation. In this talk, we'll introduce an effective framework to help companies navigate this dynamic landscape--a comprehensive typology of companies in the age of AI, including both AI-native and traditional firms. This typology is structured around three key dimensions: AI readiness of a company's business model, control over high-quality data, and expertise in AI algorithms.
In exploring this framework, five distinct company archetypes are identified: AI-born Companies, Digital Champions, Problem Owners, AI Tools Providers, and Digital Zombies. Each archetype faces unique challenges and needs specific strategies, which will be examined in depth, offering valuable insights to support decision-making, collaboration, and research in the rapidly evolving AI environment.

Sreda, 27.11.2024. u 19:00, Online
Đorđe Žikelić, Singapore Management University (SMU)
NEURALNI KONTROLNI SISTEMI I NJIHOVA FORMALNA VERIFIKACIJA
Učenje s podsticanjem je ostvarilo impresivne rezultate, što je podstaklo interesovanje za njegovu upotrebu u automatskom upravljanju i robotici. Međutim, sigurnosno-kritična priroda automatskih kontrolnih sistema kao što su autonomna vozila dovodi u pitanje sigurnost naučenih kontrolora i zahteva sigurnosne garancije. Ovo predavanje će predstaviti neuro-simbolički okvir za učenje i formalnu verifikaciju neuralnih kontrolora u kontrolnim sistemima. Za zadati kontrolni sistem i specifikaciju svojstva koje sistem treba da zadovolji, naš metod uči neuralni kontrolor zajedno sa neuralnim sertifikatom koji formalno dokazuje da je specifikacija zadovoljena. Metod se takođe može koristiti i za formalnu verifikaciju već naučenih neuralnih kontrolora u odnosu na zadatu specifikaciju. Metod je primenjiv na stohastičke kontrolne sisteme, i predstavlja prvi metod za formalnu verifikaciju neuralnih kontrolora u stohastičkim kontrolnim sistemima.

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