Jekyll2025-10-21T09:53:58+00:00/feed.xmlTim RoithTim RoithTim Roith tim{dot}roith{at}desy{dot}deBridging AI and Theory: Workshop on the Mathematics of Transformers at DESY2025-10-15T00:00:00+00:002025-10-15T00:00:00+00:00/2025/10/15/TransformerAt a recent workshop hosted by DESY, around 30 experts from across Europe came together to explore the theoretical foundations of transformers, the core architecture behind modern large language models. Organized by researchers from DESY, Helmholtz Imaging, and the University of Oxford, the event featured inspiring talks and a lively world café discussion. The positive response shows growing momentum in this young field. Stay tuned for a potential second edition!

🔗 https://bit.ly/mathematics-of-transformers

The workshop was jointly organized by Martin Burger, Samira Kabri, Tim Roith, Lukas Weigand (Computational Imaging at DESY and Helmholtz Imaging) and Konstantin Riedl (University of Oxford). It was a satellite event to the Conference on Mathematics of Machine Learning, also co-organized by Martin Burger.

]]>Tim Roith tim{dot}roith{at}desy{dot}dePosition-Blind Ptychography: Viability of image reconstruction via data-driven variational inference2025-09-28T00:00:00+00:002025-09-28T00:00:00+00:00/paper/2025/09/28/PtychoWe put a preprint on the arXiv dealing with position-blind pytychography. Compared to the standard setup, here we ask ourselves how well the reconstruction problem can be solved, when the scan position are completely unknown. Furthermore, we explore how to leverage a powerful data-driven prior with a variational inference approach. This is joint work Simon Welker, Lorenz Kuger Berthy Feng, Martin Burger, Timo Gerkmann, Henry Chapman. Part of this project was done during my DAAD funded research stay at Caltech, in Oktober 2024.

The preprint is available on the arXiv: Position-Blind Ptychography: Viability of image reconstruction via data-driven variational inference.

]]>Tim Roith tim{dot}roith{at}desy{dot}deIntroduction to Regularization and Learning Methods for Inverse Problems2025-08-27T00:00:00+00:002025-08-27T00:00:00+00:00/paper/2025/08/27/LIPTogether with Danielle Bednarski we uploaded our lecture notes Introduction to Regularization and Learning Methods for Inverse Problems to arXiv. The notes are based on a course we taught at the University of Hamburg in the winter term 2024/2025.

]]>Tim Roith tim{dot}roith{at}desy{dot}deConsensus-based optimization for closed-box adversarial attacks and a connection to evolution strategies2025-06-30T00:00:00+00:002025-06-30T00:00:00+00:00/paper/2025/06/30/AdvCBOTogether with Leon Bungert and Philipp Wacker we uploaded a paper to arXiv, which shows how CBO can be used to create closed-box adversarial attacks. Beyond that we establish an interesting connection between CBO and evolution strategies. The paper can be found here.

]]>Tim Roith tim{dot}roith{at}desy{dot}deAnalysis of mean-field models arising from self-attention dynamics in transformer architectures with layer normalization2025-06-05T00:00:00+00:002025-06-05T00:00:00+00:00/paper/2025/06/05/TransformersI’m happy to share that our paper Analysis of mean-field models arising from self-attention dynamics in transformer architectures with layer normalization has been published in the Philosophical Transactions of the Royal Society A.

]]>Tim Roith tim{dot}roith{at}desy{dot}deDr.-Klaus-Körper Prize2025-04-07T00:00:00+00:002025-04-07T00:00:00+00:00/2025/04/07/KKPI am very happy to share that I was awarded the Dr.-Klaus-Körper Prize for my PhD thesis today. The prize is awarded by the GAMM at the annual meeting in Poznań.

]]>Tim Roith tim{dot}roith{at}desy{dot}deMirrorCBO: A consensus-based optimization method in the spirit of mirror descent2025-01-21T00:00:00+00:002025-01-21T00:00:00+00:00/paper/2025/01/21/MirrorCBOTogether with Leon Bungert, Franca Hoffmann and Doh Yeon Kim we poste a new preprint on MirrorCBO: A consensus-based optimization method in the spirit of mirror descent.

]]>Tim Roith tim{dot}roith{at}desy{dot}deAnalysis of mean-field models arising from self-attention dynamics in transformer architectures with layer normalization2025-01-06T00:00:00+00:002025-01-06T00:00:00+00:00/paper/2025/01/06/TransformersTogether with Martin Burger, Samira Kabri, Yury Korolev and Lukas Weigand we posted a new preprint called Analysis of mean-field models arising from self-attention dynamics in transformer architectures with layer normalization.

]]>Tim Roith tim{dot}roith{at}desy{dot}deCOMFORT2024-10-18T00:00:00+00:002024-10-18T00:00:00+00:00/2024/10/18/COMFORTOur project COMFORT: Compression Methods for Robustness and Transferability is being funded by the BMBF!

]]>Tim Roith tim{dot}roith{at}desy{dot}deCMX Lunch Seminar2024-10-15T00:00:00+00:002024-10-15T00:00:00+00:00/presentation/2024/10/15/CMXI’m very excited to give a talk in the CMX Lunch seminar at Caltech today. I will present my work on the mathematics of adversarial robustness, you can find the abstract here.

]]>Tim Roith tim{dot}roith{at}desy{dot}de