 |
Universität Augsburg
|
|
|
Universität Augsburg
|
|
Professor Dr. Martin Benning
University College London
spricht am
Mittwoch, 23. Juli 2025
um
16:00 Uhr
im
Raum 2004 (L1)
über das Thema:
| Abstract: |
| In this talk, we introduce a class of Fenchel functions and demonstrate how they can be utilised for the training and inversion of deep neural networks and the estimation of source condition elements in inverse problems. The latter are synonymous with the definition of function-minimising inverse problems solutions. In particular, we show how training of so-called proximal neural networks can be carried out without backpropagation and the need to differentiate activation functions, and present a first convergence result (to the best of our knowledge) for the regularised inversion of a perceptron that only assumes that the solution of the inverse problem is in the range of the regularisation operator. We conclude the talk with examples of using Fenchel functions to estimate source condition elements and show how estimating sparse source condition elements can be applied to learning optimal sampling patterns in Magnetic Resonance Imaging. This is joint work with Audrey Repetti & Xiaoyu Wang (Heriot-Watt), Andreas Mang (Univ Houston), Alexandra Valavanis & Danilo Riccio (QMUL), Azhir Mahmood (UCL), Tatiana Bubba (Univ Ferrara) and Luca Ratti (Univ Bologna). |
| Hierzu ergeht herzliche Einladung. |
| Prof. Dr. Jan-Frederik Pietschmann |
Kaffee, Tee und Gebäck eine halbe Stunde vor Vortragsbeginn im Raum 2006 (L1).