I recently had the privilege of attending the 3rd European Summer School on Artificial Intelligence (ESSAI 2025), hosted at the Slovak University of Technology in Bratislava. I chose to participate in this event because of its strong alignment with my research interests and the unique opportunity it offered to deepen my knowledge while engaging with leading researchers in the field.
The summer school spanned over five days and was structured around intensive lecture series complemented by tutorials and social activities. Each participant could follow several in-depth courses over the week, while the tutorials provided a broader perspective on the state of the art in AI research. This combination of depth and breadth proved invaluable, thereby allowing me to explore specific topics closely tied to my doctoral work while also keeping track of wider developments in related fields.
One of the most inspiring courses for me was Self-Governing Systems, taught by Jeremy Pitt and Asimina Mertzani from Imperial College, London. The course explored how future socio-technical systems can be designed to combine human and machine intelligence in ways that balance cooperation, agency, and accountability. For my own research, where I study how large language models (LLMs) can collaborate through discourse on complex socio-economic issues, this was particularly relevant since it offered me new ways of thinking about how to design reward signals and evaluation criteria that encourage fairer and higher-quality interactions between these models. I left with concrete ideas for how to improve my current work on debiasing LLMs in political discussions.
Another highlight was the course on Uncertainty in Machine Learning, led by Matias A. Valdenegro Toro and Marco Zullich from the University of Groningen. The lectures addressed how models can be extended to not only make predictions but also quantify their uncertainty, since typically, most machine learning models are poorly calibrated for both classification and regression tasks. The course included both lectures and hands-on scripts for us to implement and debug which was extremely useful. I was able to apply these insights almost immediately in my own work on sentiment analysis for South Slavic languages, including Slovene. By introducing uncertainty modelling into my pipeline, rather than relying solely on softmax scores, I achieved better calibrated results and, in fact, reached a new state-of-the-art score on this benchmark for Slovene on our dataset. This concrete outcome alone made the course extremely rewarding.
Other courses and tutorials, such as those on machine unlearning, ethics in trustworthy AI, multimodal vision models, and large-scale training of LLMs, were also valuable in shaping my understanding of privacy, fairness, and scalability in AI. These broader perspectives are particularly relevant as I look for unsupervised approaches to extend my work.
Beyond the lectures, the summer school was also an important networking experience. I had the chance to exchange ideas with professors and peers, some of whom I hope to collaborate with in the near future. The informal setting of social events, whether discussing research over dinner or even going indoor bouldering together, helped build meaningful connections. Experiencing the city of Bratislava with fellow participants added a personal dimension that made the week even more memorable.
I am deeply grateful to the SDJT association for supporting my participation through a scholarship. ESSAI 2025 was not only a rich academic experience but also a chance for professional and personal growth. The knowledge and connections I gained there will undoubtedly continue to influence my research and future collaborations.
Nishan Chatterjee
