Machine learning (ML) algorithms are increasingly permeating our lives. They make predictions, but their decisions often remain opaque – that is, inaccessible or incomprehensible. Regulators and experts are calling for greater transparency and explanations, as if ML algorithms were communication partners. We view them as physical systems that interact with their environment. This opens up new perspectives for our understanding: Machine learning is a form of adaptive behavior. Opacity arises from complexity, which may or may not be reduced depending on the level of abstraction. Understanding then arises not through greater transparency or explanations, nor through reduced complexity, but through a meaningful shift in the level of abstraction. The physical perspective also provides impetus for a new generation of AI: if ML algorithms can simulate physical systems, this process is also reversible, and thus physical systems can embody new types of artificial intelligence.

About the authors: Dr. Miriam Klopotek studied physics in Berlin and Tübingen and received her Ph.D. from the University of Tübingen in 2021. Since 2022, she has been a group leader at the Stuttgart Center for Simulation Sciences (SimTech Cluster of Excellence). Since 2023, she has been co-director (with Eric Raidl) of the WIN project “Complexity Reduction, Explainability, and Interpretability” at the Heidelberg Academy of Sciences and Humanities. She is interested in the interactions and analogies between artificial intelligence and physical dynamics, particularly those underlying condensed matter.

PD Dr. Eric Raidl studied philosophy, computer science, and mathematical logic in Berlin and Paris. He received his Ph.D. from the University of Paris Sorbonne in 2014 and completed his habilitation at the University of Konstanz in 2022. He has worked at the École Normale Supérieure in Paris, the University of Konstanz, and University College Freiburg. Since 2019, he has been at the University of Tübingen as a co-PI of the Philosophy and Ethics Lab within the Cluster of Excellence “Machine Learning for Science.” His research interests include epistemology, philosophy of science, logic, and AI.

About the lecture series: This public lecture series has been taking place for over 20 years now, featuring scholars from the Heidelberg Academy of Sciences and Humanities as well as from its seven sister academies. The lectures are aimed at a broad audience to provide insights into the research being conducted. Afterward, attendees have the opportunity to chat with the scholars over pretzels and wine in the Academy’s courtyard garden.

The series is held in cooperation with vhs Heidelberg.