Sympozjum Polskiego Stowarzyszenia
Sztucznej Inteligencji

Artificial Intelligence (AI) has moved from being a niche research curiosity to becoming an indispensable component of modern science and a foundational element of digital infrastructure. In science, AI already drives discovery and innovation, accelerating progress in fields from environmental modelling to life sciences and materials research. In society, it increasingly operates as an enabling infrastructure—supporting research, innovation, and public services at national and international scales.
The lecture will explore two interrelated perspectives on AI’s role in the immediate future. First, I will give an overview of our work in developing AI methods specifically suited for scientific discovery. These include approaches for explainable machine learning—covering multi-target prediction, relational learning, and interpretable modelling of complex dynamical systems—that balance predictive accuracy with scientific insight. I will also discuss strategies for learning from limited data, including semi-supervised learning and foundation models, which extend AI’s reach to data-scarce domains. Finally, I will highlight the role of ontologies and semantic technologies in supporting open, reproducible science by formalising and sharing scientific knowledge, data, and computational artefacts.
Second, I will discuss AI as a national-scale infrastructure through the lens of the Slovenian AI Factory (SLAIF), which I coordinate technically. SLAIF will provide computing resources, tools, and expertise to a wide spectrum of users—from researchers to public institutions and industry—lowering barriers to AI adoption and fostering a vibrant, inclusive AI ecosystem. By integrating AI into the very fabric of scientific and societal operations, we are moving towards a future where AI is as essential and ubiquitous as electricity or the internet.
I will argue that the transition from AI as innovation to AI as infrastructure is not just a technological evolution. It is a societal transformation. As such, it requires thoughtful design, broad accessibility, and deep collaboration between science, policy, and practice.

Recent research confirms that acoustic features extracted from speech are valid markers for assessing the severity of manic and depressive symptoms. At the same time, while numerous applications arise to support the monitoring of individuals with affective disorders, these systems typically do not provide explanations for their reasoning or decisions.
To address this gap, we design an explainable AI-driven, voice-based system for mental health monitoring aiming at generating explanations regarding how a person's voice speaking has changed across different affective states. The co-design process has been conducted by an interdisciplinary team of experts and stakeholders comprising engineers, technical experts, psychiatrists, legal scholars, ethicists, philosophers, and other specialists. The trustworthiness of the proposed AI system is evaluated using the Z-Inspection® methodology. This methodology involves investigation of the relevant ethical standards in the development, implementation and use of artificial intelligence systems intelligence and also consideration of the socio-technical context in which the system will be used. During the presentation, we will share insights and lessons learned from the trustworthiness assessment and co-design process of the proposed system.
Next, we will address key challenges related to the construction of linguistic summaries as explanations in evolving environments, as well as their validation. The overall aim is to deliver human-consistent information granules that describe in natural language large datasets, including inhomogeneous time series. We consider how recent developments in the theory of evaluative linguistic expressions and fuzzy association rule mining can complement summarization. We will briefly recall related work on the evaluation criteria for the quality of both individual sentences as well as group of sentences. The assessment of the quality of the groups or sequences of summaries becomes even more complex for real-world scenarios in which additional data becomes available over time or the existing information is incomplete.
The talk will conclude with some practical examples from the remote smartphone-based mental health monitoring system, which aims to support doctors and patients in monitoring affective episodes and fuzzy information granules will be its core component.