The group researches mathematical foundations of AI, machine learning, and signal/image processing—covering neural network generalization, robustness, and trustworthiness, plus reliable/sustainable AI methods. It also studies high-dimensional probability, optimization, harmonic analysis, and compressed sensing, linking core math theory to real-world AI challenges.
The research group investigates the mathematical foundations of modern methods in machine learning, artificial intelligence, and signal and image processing. The goal is to develop a theoretical understanding of complex data-driven processes and to derive powerful, reliable, and efficient methods from this understanding.
A key focus is on the mathematical foundations of AI systems and deep learning. Among other topics, the group investigates generalization, robustness, explainability, and trustworthiness of neural networks, as well as issues related to the security and transparency of AI. In addition, new approaches for reliable and sustainable AI are being developed, including novel hardware-software concepts and methods for meeting regulatory requirements, such as those outlined in the EU AI Act. The research combines mathematics, computer science, and statistics and also encompasses applications in medical imaging, robotics, and scientific and technical problems.
Another area of focus is the mathematical theory of machine learning, signal processing, and compressed sensing. In particular, the research centers on high-dimensional probability theory, optimization, harmonic analysis, and inverse problems. Among other topics, the group examines the convergence properties of optimization methods, the theory of neural networks, and methods for the efficient reconstruction and processing of high-dimensional data.
The combination of these research areas creates a coherent framework that links fundamental mathematical research with current challenges in artificial intelligence and data-driven sciences.
The project "gAIn - Next Generation AI Computing" is pioneering a theory-driven framework to revolutionize AI through disruptive hardware-software solutions, addressing critical barriers in energy efficiency, reliability, regulatory compliance, and computability.
In July 2024, the European legislative initiative to regulate artificial intelligence in the form of a regulation [“AI Act”] was adopted. The AI Act came into force on August 1, 2024, with a general transition period, meaning that the European AI ecosystem must now adapt to these legal requirements in a legally compliant manner. To ensure the protection of health, safety, and human rights, providers and users of high-risk AI systems must meet numerous legal requirements.The implications of the AI Act immediately call for an active policy strategy and appropriate measures to address these multifaceted challenges and continue to capitalize on the opportunities arising from the use of trustworthy AI systems. The AI Act raises important questions about the future of AI development and the role of government in shaping it.Relevant and applied research is essential to support the development and deployment of trustworthy AI in the Bavarian ecosystem in an evidence-based manner, especially as a basis for political and internal organizational measures.
The GeniusRobot project aims to improve robotic gripping and manipulation by using generative AI to predict tactile information from visual data and vice versa. By integrating camera and tactile sensor inputs through interpretable multimodal AI models, robots will be able to adapt flexibly and safely to changing environments and objects. This approach opens new possibilities for applications in manufacturing, medicine, and human-machine interaction.
AI-based methods for medical image analysis have garnered significant attention due to their potential to revolutionize healthcare. However, challenges persist, including the scarcity of labeled medical data and the inherent complexity of medical knowledge. Furthermore, the successful integration of AI into clinical practice requires decision-making processes that are both interpretable and reliable for healthcare professionals.
In this project, we are collaborating with medical and industry partners to develop AI models specifically tailored for biopsy image analysis. Our initial focus is on assisting in the identification of Gleason patterns in prostate cancer, with the aim of expediting the diagnostic process. Ultimately, our goal is to create a transparent and trustworthy AI-powered assistant that enhances clinical decision-making.
We study Spiking Neural Networks (SNNs) as biologically-inspired models of computation, focusing on their expressivity, dynamics, learning algorithms, and energy efficiency. Our goal is to understand how SNNs encode and process information robustly and efficiently, including their implementation on emerging neuromorphic hardware.
The vision of the 'Konrad Zuse School of Excellence in Reliable AI' (relAI) is to train future generations of AI experts, who for the first time combine technical brilliance with awareness of the importance of AI’s reliability. Our novel, highly innovative AI program will educate top international candidates in the end-to-end development of reliable AI systems (including scientific knowledge, business expertise, and industrial exposure), both for industry and academia, and perform cutting-edge research to make AI ready for deployment in critical application domains.
The AI-HUB@LMU brings together Artificial Intelligence and Data Science at LMU Munich to foster cutting-edge, interdisciplinary research, from theoretical foundations to real-world applications and societal impact. As a central platform at LMU, it supports collaboration, education, and public engagement to shape the future of trustworthy, data-driven technologies.
As artificial intelligence systems are increasingly deployed in high-stakes domains—such as healthcare, finance, law, and autonomous driving—the need for reliability has become a central research challenge. Reliability concerns whether AI systems consistently behave in predictable, safe, and trustworthy ways under a wide range of conditions. However, current models often fail when exposed to small adversarial changes, shifted data distributions, or rare corner cases. Furthermore, issues of fairness and accountability arise when AI systems treat certain groups differently, with specific biases. Due to the safety-critical nature, having not only empirical evidence but also theoretical foundations and guarantees for robustness, fairness, stability, and fail-safe behavior is especially important.
Developing AI methods that are energy-efficient, resource-conscious, and environmentally friendly, including through the use of neuromorphic and other energy-efficient hardware.