AI Research Group at the ELLIS Institute Tübingen
We use theoretical and empirical approaches to build machine vision systems that see and understand the world like humans
In the past few years, deep neural networks have surpassed human performance on a range of complex cognitive tasks. However, unlike humans, these models can be derailed by almost imperceptible perturbations, often fail to generalise beyond the training data and require large amounts of data to learn novel tasks. The core reason for this behaviour is shortcut learning, i.e. the tendency of neural networks to pick up statistical signatures sufficient to solve a given task instead of learning the underlying causal structures and mechanisms in the data. Our research ties together adversarial machine learning, disentanglement, interpretability, self-supervised learning, and theoretical frameworks like nonlinear Independent Component Analysis to develop theoretically grounded yet empirically successful visual representation learning techniques that can uncover the underlying structure of our visual world and close the gap between human and machine vision.
We open the black box of modern neural networks by studying how information is represented, transformed, and routed inside them. Our work evaluates where popular visualisation methods (e.g., saliency, feature visualisation) succeed or fail, and develop tools and benchmarks that tie internal mechanisms to model behaviour. The goal is explanations that are reliable, actionable, and predictive—supporting debugging, safety, and the design of models that make decisions we can understand and trust.
We study when and why learning systems generalize beyond their training data. Using identifiability theory, we characterize the conditions under which models can recover the underlying rules and structure from data—yielding representations and internal mechanisms that transfer to novel scenarios. Our goal is a rigorous foundation for trustworthy generalization, informing both theory and the design of models that reliably handle out-of-distribution cases.
We study how to align perception, cognition, and action so that AI systems can build and use internal world models for multi-modal reasoning. Our work connects signals across vision, language, and interaction, then tests whether the resulting representations support explanation, counterfactual inference, and planning. The goal is robust, human-aligned agents that can navigate complex environments, tasks, and decisions just like humans.
We investigate the similarities and differences between human and machine perception, learning, and reasoning. By comparing how humans and artificial systems see, understand, and make sense of the world, we aim to identify key gaps that limit machine intelligence and uncover principles that drive human cognition. Our work spans vision, language, and general reasoning, combining behavioral experiments, computational modeling, and machine learning. These insights guide the development of AI systems that are more adaptable, robust, and aligned with human capabilities. We collaborate with Felix Wichmann and Matthias Bethge.
AI is already reshaping how we work, learn, and make decisions. We feel a responsibility to guide and uncover how machine learning can foster economically feasible solutioons that best address long-term human needs. We spin off new projects like robotics for sustainable agriculture or individualised learning, engage in mentoring and foster an open-minded innovation- and impact-driven mindset.
Our Work Beyond Academia
The Bundeswettbewerb für Künstliche Intelligenz (BWKI) is a federal competition for artificial intelligence (AI) in Germany, Austria and recently Switzerland. It is organized by the Tübingen AI Center funded by the Carl-Zeiss-Stiftung and is aimed at promoting interest and talent in AI among young people. The competition is open to students from different age groups, and participants work in teams to develop innovative solutions for real-world problems using AI technologies. The BWKI also aims to encourage the development of AI skills and knowledge, which is becoming increasingly important in today’s digital age.
IT4Kids is a non-profit organization in Germany that provides programming education for primary and lower secondary school students. Their mission is to support the digital transformation of education in schools by teaching coding and programming skills to children in a fun and engaging way.
The Kinderuni (“children’s university”) is a project by the University of Tübingen, where university Professors (or their PhD students) give child-friendly lectures about their field. We have contributed to this initiative by offering lectures on “What is AI, and why can’t computers do my homework?”