Machine learning has arrived in the post-dataset era with models being used on ever increasing data and tasks with dynamically evolving evaluation criteria including safety, domain contamination and computing costs in addition to performance. Therefore, developing new concepts and tools for lifelong/infinite benchmarking and the ability to efficiently democratize evaluation are increasingly vital for transparent model assessment. At the same time this opens many opportunities for using machine learning beyond prediction towards understanding in continual scientific model building.

Language model agents — AI systems capable of autonomous thinking, communication, and reasoning — enable rich, natural human-machine interactions and collaboration on complex tasks. We aim to develop assistants for theorem proving, automating scientific discovery, and aggregating information from the web to make reliable, near-future predictions in uncertain scenarios.

Lifelong learning requires making experiences in the past reusable for the future. Common regularization methods for preventing catastrophic forgetting in lifelong learning do not scale. We hypothesize that compositional learning is key to scalable lifelong learning in humans. The object-centric nature of human perception is a strong indication for an inherently compositional representation of the world. We combine conceptual research on compositionality and object-centric perception with scalable, practically relevant lifelong learning methods and benchmarks, until we figure out how to merge them.

We develop machine learning models for neural data analysis to understand how populations of biological neurons perform inference and learning in the brain. We are particularly interested in understanding the principles that govern distributed processing in populations of neurons. To this end, we build and benchmark digital twins and detail-on-demand models of certain brain areas (primarily mammalian retina and visual cortex), and develop machine learning tools for interpreting, comparing and ultimately understanding the representations and computations in these neural networks.

Human Attention facilitates active perception and inference and can serve as a signature of lifelong cognitive navigation. We want to understand how humans benefit from this mechanism and how it may help us to improve attention mechanisms in machine learning. We build and benchmark models of human attention in various modalities (image and video saliency, scanpath prediction, eye movements in VR) and aim at including them as building blocks in downstream models of computer vision tasks and human behavior. We collaborate with Felix Wichmann, Alexander Mathis, Ralf Engbert, and Christoph Teufel.

Machine learning is rapidly expanding the range of skills that can be used for new solutions to relevant problems in the world either by being more scalable or more precise then human labor. We seek to develop a better understanding of how we can develop economically feasible solutions that best address long-term human needs. We spin off and collaborate with startups such as Maddox AI or Black Forest Labs.