 |
Prof. Sanda Martinčić-Ipšić
Sanda Martinčić-Ipšić obtained her B.Sc. degree in Computer Science from the University of Ljubljana Faculty of Computer Science and Informatics, and her M.Sc. degree in informatics from the University of Ljubljana, Faculty of Economy. In 2007 she obtained a Ph.D degree in Computer Science from the University of Zagreb, Faculty of Electrical Engineering and Computing. Dr. Martinčić-Ipšić currently works as full professor of Computer Science at the University of Rijeka, Faculty of Informatics and Digital Technologies. She is a leader of Language Networks research group, and the head of Laboratory for Natural Speech and Language processing at Center for Artificial Intelligence and Cybersecurity at University of Rijeka. Her research interests include natural language processing, complex networks, data science and data analytics. She has published more than 100 scientific papers and books. |
Talk title: From Text to Insight: Enhancing Relation Extraction in Climate Change Research
Abstract: Global warming and climate change have profound and far-reaching effects on global ecosystems, weather patterns, sea levels, and human societies, constituting a critical threat to the planet's biodiversity and the prospect of a sustainable future. Simultaneously, the volume of climate change data is rapidly increasing, particularly in published scientific publications. The automated processing of information from unstructured textual data is crucial, with a primary focus on the natural language processing task of information extraction and, more specifically, its subtask, relation extraction. Relation extraction involves identifying relationships between entities within sentences, paragraphs, or larger text units, aiming to automatically generate machine-interpretable data collections that capture entities, their relationships, and associated attributes. This talk addresses the challenge of extracting named entities and relations from scientific publications from renowned journals in the climate change domain. Firstly, the statistics of the collected dataset will be presented along the problems encountered in data preprocessing. Secondly, the domain adaptive pretraining of the SciBERT and Climate(Ro)BERT(a) models and from scratch training of CliReBERT and CliReRoBERTa models will be elaborated. The discussion will focus on the model architectures and training parameters used, highlighting the advantages and disadvantages of domain adaptive pretraining compared to training from scratch. Thirdly, the task of extracting relations and named entities in the climate change domain will be elaborated upon, presenting results on LLM-enabled relation annotation and discovery. These results will be used to train all pretrained models (i. e. BERT and RoBERTa) to supervised relation extraction downstream task. Finally, the plan for constructing a knowledge graph from extracted relations in the climate change domain will be discussed.
 |
Andrius Januta
Andrius Januta is a cybersecurity technical manager at Nord Security. His responsibilities encompass designing, implementing, and maintaining the company’s cybersecurity strategy, including deploying and managing advanced security tools to protect sensitive data. Over the past five years, he has been an active participant in Lithuania’s cybersecurity and defense exercise, Amber Mist, where he served as a core member of both the Red Team and the technical cyber range development team. Additionally, he has been a regular participant in various cybersecurity exercises organized by the NATO Cooperative Cyber Defence Centre of Excellence (CCDCOE). |
Talk title: Adversarial Attacks on AI: Understanding and Securing Machine Learning in Cybersecurity
Abstract: Artificial Intelligence (AI) and Machine Learning (ML) have become essential tools in modern cybersecurity, but these models themselves are vulnerable to a wide range of attacks. One of the most serious threats is adversarial attacks, where malicious actors manipulate the inputs of ML models to produce incorrect or harmful outputs. The presentation will explore the primary vulnerabilities of AI and ML models that pose a threat to cybersecurity. The main focus will be on understanding how adversarial attacks work, the consequences they can have for cybersecurity applications, and how these attacks can be used to facilitate malicious activities such as manipulating machine learning outcomes. We will also examine the threats of prompt hacking, where AI models become susceptible to deceptive queries.
 |
Assoc. Prof. Dr. Dmitri E. Kvasov
Associate Professor in Numerical Analysis, DIMES, University of Calabria, Rende (CS), Italy. Italian National Scientific Habilitation as Full Professor in Numerical Analysis (2018–2027) and in Operations Research (2021–2030). Education: Ph.D. in Operations Research (05/2006), Department of Statistics, University of Rome "La Sapienza", Italy. Candidate (Ph.D.) of Physico-Mathematical Sciences (12/2016), "Lobachevsky" University of Nizhny Novgorod, Russia. Graduated, with honours, in Information Systems (06/2001), Faculty of Computational Mathematics and Cybernetics, "Lobachevsky" University of Nizhny Novgorod, Russia. Graduated, with honours, in Computer Systems Engineering (04/2001), Engineering Faculty, University of Calabria, Italy. Research interests: Numerical analysis; Continuous global optimization and applications; High-performance and Infinity computing. List of papers includes more than 130 items (among them: 2 research books). Research Interests: Continuous global optimization and applications; high-performance and infinity computing |
Talk title: Advanced Global Optimization Techniques and Their Applications
Abstract: In many simulation-based applications of optimization, the objective function can be multi-extremal and non-differentiable, thus precluding the use of descending schemes with derivatives. Moreover, the function is often given as a black-box and, therefore, each function evaluation is an expensive operation with respect to the available computational resources. Derivative-free methods can be particularly suitable to address these challenging problems studied in the framework of global optimization and can be deterministic or stochastic in nature. A numerical comparison of these two groups of methods is interesting for several reasons and has a notable practical importance. In the presentation, the methods of these two groups are considered and their applications (including the field of machine learning) are briefly examined.