Lunch Talk: Deep Generative Models for Inverse Problems

Abstract: Inverse problems have far-reaching applications in fields like medical imaging (EEG, CT), seismology, and image deblurring. These problems aim to uncover hidden properties from indirect measurements.  While the relationship between the properties and measurements (the forward model) is often easy to compute, determining the properties from the measurements (the inverse problem) is notoriously difficult due to its ill-posed nature. Deep learning offers a powerful way to improve inverse problem solutions by learning data-driven priors. In this talk, we'll explore the challenges of source reconstruction from EEG data, a task plagued by low signal quality and severe ill-posedness. By starting from pseudo-inverse, we enable the use of 3D Convolutional U-Nets. These robust deep learning architectures achieve superior noise tolerance and reconstruction performance compared to classical methods.

Bio: Marco Morik is a PhD Student in the Machine Learning Group at Technische Universität Berlin / BIFOLD working on Uncertainty Estimation and Generative Models for Inverse Problems. He received his master degree in Computer Science from TU Berlin in 2019 with a focus on discrete optimization and Machine Learning. Since 2022 he is working in the Research Group Probabilistic Modeling and Inference with Dr. Shinichi Nakajima.

To plan the small catering accordingly, I would like to ask you to register beforehand – either through the Outlook reply function or by sending an e-mail to: laura.wollenweber@tu-berlin.de