Branislav Pecher

Branislav has expertise in artificial intelligence, machine learning and natural language processing, with a specific focus on learning with limited labelled data and low-resource settings. In this domain, during his doctoral studies, he mainly focused on a rigorous investigation how sensitive the approaches for learning with limited labelled data are to non-deterministic decisions, such as random seeds or sample selection that are often overlooked but may have significant effects. Currently, his focus has expanded in two directions. The first is evaluation and understanding of vulnerabilities and capabilities of large language models – through mechanistic understanding of what is actually happening inside these models. The second direction is auditing of social media platforms for potentially harmful behaviour – such as spreading misinformation, enclosing users in filter bubbles, or negatively affecting minors.

He has co-authored more than 20 papers that were accepted to top journals (ACM Computing surveys) or top-tier conferences in AI and NLP (including, EMNLP, ACL, SIGIR or RecSys). Together, these publications have more that 440 citations, resulting in a h-index of 11 (according to Google Scholar, as of 2026/04). He has also been involved in multiple projects, including AI-Auditolog, VeraAI, or TAILOR. This included international research collaborations, with universities such as Bar Ilan University or Eindhoven University of Technology.

Supervising team: Mária Bieliková (KInIT), Ivan Srba (KInIT)

Learning with limited labelled data, including but not limited to concepts such as zero/few-shot learning, transfer learning or meta-learning, aims to effectively train a model while using only a small amount of labelled samples. Although such approaches were successfully applied in many domains, they are especially sensitive to the effects of uncontrolled randomness caused by non-determinism in the training process. The non-determinism negatively affects the stability of the trained models and leads to large variance in results across training runs. Such uncontrolled randomness can unintentionally, but unfortunately also intentionally, create an imaginary perception of research progress. When disregarded, the randomness can prohibit objective comparison of new methods, leading to overestimation, or in some cases underestimation, of their performance. Even though a growing number of studies focus on addressing the negative effects of randomness on stability, the research is still limited in its extent, facing many problems and inconsistencies. 

Our aim in this work is to deal with the drawbacks when addressing the effects of randomness in learning with limited labelled data. Based on a comprehensive survey, we identify the most significant open problems, specifically the focus on investigation only on a small fragment of randomness, disregard for the interactions between sources of randomness and an absence of effective mitigation strategies. As a part of preliminary experiments, we design a general methodology for investigating effects of randomness. We apply this investigation methodology to determine the importance of multiple sources of randomness in meta-learning and language model fine-tuning. The preliminary results revealed a significant impact of randomness originating from splitting of data and selecting what data is labelled, which are mostly ignored in current research.

Reviewer for journals:

Reviewer at multiple conferences, including: