In this CRC colloquium we're happy to welcome:
Illia Horenko (TU Kaiserslautern)
Mathematics as a key to the accountable and robust AI?
The ubiquitous and boosting rise of AI technologies like AlphaGo, AlphaFold and GPT-4 heralds a new historic era. Multiple areas of human activity will be affected by these rapid and - in many aspects, alarming and potentially dangerous - developments. A brief review of some key mathematical challenges omnipresent in this rise of AI will be discussed in the talk, highlighting how mathematics can potentially help in mitigating some of the emergent problems and risks. lt will be discussed how math may potentially help in reducing AI vulnerability to adversarial attacks, as well as in decreasing its sensitivity to data amount, lack of the necessary hardware power and exploding AI costs. lt will also be briefly discussed how mathematics can potentially help increasing accountability, explainability and computational scalability of the "black box" AI, with some examples from mathematics, climate research, economics and biomedical sciences.
Gitta Kutyniok (LMU München)
Reliable AI: Successes, Challenges, and Limitations
Artificial intelligence is currently leading to one breakthrough after the other, both in public life with, for instance, autonomous driving and speech recognition, and in the sciences in areas such as medical imaging or molecular dynamics. However, one current major drawback is the lack of reliability of such methodologies.
In this lecture we will take a mathematical viewpoint towards this problem, showing the power of such approaches to reliability. We will first provide an introduction into this vibrant research area, focussing specifically on deep neural networks. We will then survey recent advances, in particular, concerning generalization guarantees and explainability, touching also the application of reliable AI methods to solving imaging problems and PDEs using suitable combinations of AI and model based approaches. Finally, we will discuss fundamental limitations of deep neural networks and related approaches in terms of computability, which seriously affects their reliability.
Karsten Reuter (Fritz-Haber-Institut)
Data-Enhanced Multiscale Modelling of Operando Energy Conversion Systems
Emerging operando spectroscopies and microscopies reveal a highly dynamic behavior of interfaces in energy conversion systems. Insufficient insight and the concomitant inability to control or exploit the corresponding strong structural and compositional modifications centrally limits the development of performance catalysts, electrolyzers or batteries. Traditional predictive-quality modeling and simulation is essentially unable to address the substantial, complex and continuous morphological transitions at such working interfaces. I will review this context from the perspective of first-principles based multiscale modeling, highlighting that in particular machine-learning surrogate models are key to help bridge the scales and thereby tackle the true complexity of the working systems. Central challenges concern active learning loops and a seamless coupling with physical models to reach the required maximum computational and data efficiency.