Conference Agenda

Over the past decade, AI has catalyzed transformative advances across the sciences, evidenced by recent Nobel-recognized breakthroughs in artificial neural networks and computational protein design. In medicine, AI has transitioned from inflated expectations to robust clinical utility, particularly in radiology, which currently accounts for 76% of FDA-authorized AI medical devices. This presentation examines the trajectory of AI toward enabling personalized diagnosis, disease stratification, and precise prognostic evaluation. Key developments include AI-driven image reconstruction techniques that accelerate MR image acquisition and promise to democratize global access to medical imaging.

Furthermore, the integration of multimodal data architectures is unlocking latent diagnostic value in highly accessible modalities. Novel contrastive learning frameworks successfully align cost-effective electrocardiograms with high-fidelity cardiac MR imaging, significantly enhancing the prediction of cardiovascular disease. At the population level, large-scale analysis demonstrates that combining whole-body MRI, radiomics, and comprehensive clinical records substantially improves risk assessment for a diverse set of diseases.

The emergence of Large Language Models (LLMs) presents a paradigm shift. However, while state-of-the-art LLMs achieve passing scores on standardized medical examinations, they exhibit critical limitations in real-world clinical decision-making, including poor adherence to clinical guidelines, inadequate laboratory interpretation, and sensitivity to information ordering. To bridge this gap, the development of specialized Vision-Language Models (VLMs), such as RetinaVLM, demonstrates superior clinical utility in biomarker recognition and disease staging.

Realizing the full potential of clinical AI necessitates addressing profound technical and societal challenges. Standard federated learning paradigms remain vulnerable to adversarial interference, requiring advanced privacy-preserving frameworks that integrate differential privacy and secure aggregation. Beyond technical data security, the integration of AI into healthcare must navigate the fundamental limits of clinical predictability, characterized by inherent inter-observer variability. Finally, strategic focus must be directed toward mitigating the risk of clinician deskilling, establishing robust regulatory compliance, and ensuring that AI deployment fosters equitable care.

In many applications, data is observed in multiple modalities. In this talk, we explore methods and understanding for multi-modality in new settings. For instance, it has been shown that alignment with a learned representation can help guide diffusion models for better representations and hence generation quality. It turns out they can profit from guidance even from different modalities, at different levels. Second, while alignment assumes paired data, i.e., "views" of the same data point in different modalities, even more data is available in unpaired form. Is it possible to leverage unpaired multimodal data to enhance a model in the target modality? By viewing different modalities as projections of a shared underlying reality, it is possible to indeed obtain gains in practice and theory.

In many areas of science, graph models are used to represent complex systems with many interdependent elements. Examples include gene regulatory networks operating within cells, graphs representing molecular structures, social networks that interconnect human actors, or character networks in the study of literary texts. While deep graph learning techniques like graph neural networks (GNNs) have advanced our ability to model such data, we increasingly have access to temporal graph data that not only tells us who is connected to whom, but also when and in which temporal order those connections occur.

In this talk, I will introduce challenges that arise in the generalization of deep graph learning methods to temporal graphs. I specifically focus on issues that arise due to the arrow of time, i.e. the fact that the temporal order of connections shapes which nodes can possibly causally influence each other over time. This leads to non-trivial patterns in the *causal topology* of temporal graphs that must be accounted for in temporal graph learning. Addressing this issue, I will showcase a causality-aware temporal GNN architecture and report on recent advances to better understand the expressive power of GNNs in time series data on biological, social, and technical systems.

In this talk, I will discuss challenges in modeling uncertainty and logical reasoning in natural language processing, using question answering in complex technical domains and solving logic puzzles as exemplary tasks. Generative language models still often fail when it comes to analyzing imprecise statements, modality, hedging, conditions, and mathematical deductions.
To this end, I will present recent work on neurosymbolic models for Bayesian reasoning scenarios expressed in natural language text. Our results demonstrate that large language models mostly assume causal relations but fail on evidential and explaining‑away reasoning scenarios, while our neurosymbolic approach is robust across different types of reasoning.
I will also present recent results on solving logic puzzles by teaching LLMs to write code for answer set programming.