30.10.2025
4:15 PM - 5:15 PM | Mathematical Colloquium
Symbolic neural network learning
Prof. Dr. Helmut Bölcskei (ETH Zurich)
A central challenge for AI systems is symbolic reasoning. This is illustrated by the Abstraction and Reasoning Corpus (ARC) benchmark, with humans achieving 84% performance compared to 10% by GPT-5. Recent successes in AI reasoning, such as e.g. DeepMind’s AlphaEvolve or AlphaGeometry as well as IMO gold-medal level performance achieving systems rely on hybrid methods that combine neural networks with symbolic or geometric reasoning modules. In this talk, we consider the problem of learning the transition rules of cellular automata (CA) from observed evolution traces, a symbolic learning challenge that is even more demanding than ARC. While it has long been known that binary CA are essentially machines realizing operations in Boolean logic, we show that, in fact, all CA are logical machines, specifically, in Łukasiewicz propositional many-valued logic. This is accomplished by interpolating CA transition functions to continuous piecewise-linear maps and invoking the McNaughton theorem. Since deep ReLU networks realize continuous piecewise-linear functions, they are naturally suited to extract these logical rules from CA evolution traces. We show that all CA can be learned by recurrent neural networks. Moreover, the formula in many-valued logic characterizing the CA transition function can be extracted from the learned recurrent neural network. The talk builds a bridge between symbolic logic, cellular automata, and deep learning, pointing toward a path for endowing neural networks with genuine symbolic reasoning capability. | Petra Schwer, Johannes Walcher | | |
| NCT Data Science/ELLIS Life/Heidelberg.ai
Test-Time Training Agents to Solve Challenging Problems
Jonas Huebotter, ETH Zurich
The standard paradigm of machine learning separates training and testing. Training aims to learn a model by extracting general rules from data, and testing applies this model to new, unseen data. We study an alternative paradigm where the model is trained at test-time specifically for the given task. We investigate why such test-time training can effectively specialize a model to individual tasks. Further, we demonstrate that such test-time training enables models to continually improve and eventually solve challenging tasks, which are out of reach for the initial model. | | Zoom
Registration:
https://www.meetup.com/heidelberg-artificial-intelligence-meetup/events/311563770/ | |
| Dr. Annika Reinke, Helmholtz Imaging German Cancer Research Center, Heidelberg
"When AI Performance Misleads: From Success in Papers to Failure in Practice"? | Institute for Mathematics | Mathematikon conference room
(INF 205, 5th floor, 5/104) | |
| Great expectations: leveraging omics for understanding sepsis pathobiology and AI for the early recognition of sepsis
PD Dr. Holger A. Lindner, Translational Research in Anesthesiology and Critical Care, Universitätsmedizin Mannheim
Sepsis is a dysregulated host response to an infection that causes life-threatening organ dysfunction. It is a leading cause of hospital mortality. Timely source control, antimicrobial therapy, and hemodynamic stabilization are life-saving. Yet, in the absence of a gold standard diagnostic test, sepsis remains a clinical diagnosis. Since the early 2000s, the advent of omics technologies and the deployment of electronic health records, respectively, have fueled expectations of sepsis biomarker discovery and of algorithmic real-time surveillance on the hospital wards. I will briefly review our current understanding of sepsis pathobiology and how much closer we have come to new biomarkers of sepsis, considering our work on the innate immune response. For this as well as for AI-driven prediction and diagnosis of sepsis, I will discuss the importance of defining sepsis cases and selecting controls. Clinician feedback from our recent workshop revealed that clinicians' expectations of AI consist in the support of competencies in all domains of medical practice. For this, the CanMEDS roles represent an already established framework. I will argue that this framework can guide development of AI to steward patient care and improve sepsis outcomes. | | INF 205 (Mathematikon), SR11 | |
| Workshop on "Advances in Algorithmic Optimization"
This workshop brings together leading experts to discuss recent advances in algorithmic optimization, spanning theory, methods, and applications. Topics include optimization with partial differential equation constraints, numerical methods for large-scale systems, model reduction, optimal control, and inverse problems, as well as emerging approaches in uncertainty quantification and machine learning–driven optimization. By highlighting both methodological innovations and practical applications, the workshop reflects the broad and deep impact of algorithmic optimization in modern science and engineering. The event is organized on the occasion of Volker Schulz’s birthday, honoring his pioneering contributions to the field.
Please find further information on the workshop on the webpage. | | Mathematikon Conference Room* (INF 2025, 5. Floor, 5/104) | |
| Patient trajectories with EHR integration
Prof. Dr. Martin Dugas, Direktor (Institut für Medizinische Informatik, Universitätsklinikum Heidelberg)
Quality of life (QoL) is a key outcome in clinical research. New electronic tools enable high-frequent collection of QoL.
To make QoL data available for the treatment process, integration into Electronic Health Record (EHR) systems is needed. Transferring QoL data from a patient smartphone on the Internet into the protected EHR system is challenging from an information security perspective. Preliminary data for this novel approach are presented and discussed. | | INF 205 (Mathematikon), SR11 | |
