Tobias Uelwer

I am a postdoctoral researcher in the Artificial Intelligence Group at the Technical University of Dortmund. I did my PhD degree under supervision of Prof. Dr. Stefan Harmeling in the same group. My research interests include machine learning and its applications in different areas.

I obtained my Bachelor's and Master's degree in computer science with a minor in mathematics from the Heinrich Heine University Düsseldorf.

Email / Google Scholar / DBLP / Twitter / Github

Research

My research focus on solving inverse problems, e.g., phase retrieval or CT, using deep neural networks and I also work on adversarial examples for deep neural networks. I am enthusiastic about everything that involves machine learning.

Selected Publications

	Optimizing Intermediate Representations of Generative Models for Phase Retrieval Tobias Uelwer, Sebastian Konietzny, Stefan Harmeling (equal contribution) Transactions on Machine Learning Research (TMLR)* bibtex / paper/ review Refined intermediate layer optimization of generative models and new learned initialization schemes perform exceptionally well at the Fourier and Gaussian phase retrieval problem. This work was also presented at the NeurIPS 2022 AI4Science Workshop.
	Evaluating Robust Perceptual Losses for Image Reconstruction Tobias Uelwer, Felix Michels, Oliver De Candido NeurIPS 2022 ICBINB Workshop - Understanding Deep Learning Through Empirical Falsification bibtex / paper We show that robust perceptual losses do not perform better than non-robust ones for image super-resolution.
	[Re] Solving Phase Retrieval With a Learned Reference Nick Rucks, Tobias Uelwer, Stefan Harmeling ReScience C (ML Reproducibility Challenge 2021) bibtex / report We reproduce and extend the experimental results from the ECCV 2020 paper "Solving Phase Retrieval With a Learned Reference" by Hyder et al. This work was presented at the NeurIPS 2022 Journal Showcase Track.
	A Closer Look at Reference Learning for Fourier Phase Retrieval Tobias Uelwer, Nick Rucks, Stefan Harmeling NeurIPS 2021 Workshop on Deep Learning and Inverse Problems bibtex / paper / code We show how the Gerchberg-Saxton algorithm can be unrolled to learn a reference for Fourier phase retrieval and we analyze the performance gain of learned references over references that were not learned.
	Spirometry‐based reconstruction of real‐time cardiac MRI: Motion control and quantification of heart-lung interactions Lena Maria Röwer, Tobias Uelwer, Janina Hußmann, Halima Malik, Monika Eichinger, Dirk Voit, Mark Oliver Wielpütz, Jens Frahm, Stefan Harmeling, Dirk Klee, Frank Pillekamp Magnetic Resonance in Medicine, 2021 bibtex / paper / supplementary material Real-time MRI combined with MR-compatible spirometry and retrospective binning can be used for image stabilization and offers opportunities to analyze heart-lung interactions.
	Non-Iterative Phase Retrieval With Cascaded Neural Networks Tobias Uelwer, Tobias Hoffmann, Stefan Harmeling International Conference on Artificial Neural Networks (ICANN 2021) bibtex / paper / poster The reconstruction performance of end-to-end learned methods for Fourier phase retrieval can be improved by using a cascade of neural networks.
	Learning to Plan via a Multi-Step Policy Regression Method Stefan Wagner, Michael Janschek, Tobias Uelwer, Stefan Harmeling International Conference on Artificial Neural Networks (ICANN 2021) bibtex / paper / poster Our proposed policy horizon regression method learns a policy vector using policy distillation to predict multiple sequential actions for a single observation.
	Learning to Detect Adversarial Examples Based on Class Scores Tobias Uelwer, Felix Michels, Oliver De Candido German Conference on Artificial Intelligence (KI 2021) bibtex / paper / slides Adversarial examples can be detected surprisingly well by training a support vector classifier on the class scores.
	Smaller World Models for Reinforcement Learning Jan Robine, Tobias Uelwer, Stefan Harmeling arXiv preprint bibtex / paper World models for Atari can be shrunk by using a vector quantized-variational autoencoder (VQ- VAE) to encode observations.
	Phase Retrieval Using Conditional Generative Adversarial Networks Tobias Uelwer, Alexander Oberstraß, Stefan Harmeling International Conference on Pattern Recognition (ICPR 2020) bibtex / paper / poster Conditional generative adversarial networks can be used to achieve state-of-the-art reconstructions for compressive and Fourier phase retrieval.
	Fast Multi-Level Foreground Estimation Thomas Germer, Tobias Uelwer, Stefan Conrad, Stefan Harmeling International Conference on Pattern Recognition (ICPR 2020) bibtex / paper / poster We propose a novel method for foreground estimation of an image given its alpha matte.
	PyMatting: A Python Library for Alpha Matting Thomas Germer, Tobias Uelwer, Stefan Conrad, Stefan Harmeling Journal of Open Source Software (JOSS), 2020 bibtex / paper / repository / docs PyMatting toolbox for Python which implements various approaches to solve the alpha matting problem. Our library is also able to extract the foreground of an image given the alpha matte.
	On the Vulnerability of Capsule Networks to Adversarial Attacks Felix Michels, Tobias Uelwer, Eric Upschulte, Stefan Harmeling (equal contribution) ICML 2019 Workshop on Security and Privacy of Machine Learning bibtex / paper / poster This paper extensively evaluates the vulnerability of capsule networks to different adversarial attacks. Our experiments show that capsule networks can be fooled as easily as convolutional neural networks.

Reviewing

Conferences/Workshops: NeurIPS 2023, AABI Symposium@ICML 2023, ICANN 2023, ICML 2023, NeurIPS 2022 ICBINB Workshop, NeurIPS 2022 AI4Science Workshop, ICML 2022, ML Reproducibility Challenge 2021 Fall Edition, ICLR 2022, NeurIPS 2021 ICBINB Workshop, NeurIPS 2021, ICANN 2021, ICLR 2021, ML Reproducibility Challenge 2020, NeurIPS 2019 Reproducibility Challenge, NeurIPS 2019
Journals: Inverse Problems & Imaging, Optics Express, IEEE Transactions on Neural Networks and Learning Systems, Journal of Open Source Software

Teaching (as TA)

At Department of Computer Science at the Technical University of Dortmund:

Undergraduate Project: Reproducibility in Machine Learning Research: summer 2023
Probabilistic Reasoning and Machine Learning: winter 2022
Data Structures, Algorithms and Programming II: summer 2022

At the Department of Computer Science at the Heinrich Heine University:

Programming: winter 2020
Reinforcement Learning: summer 2020
Deep Learning: summer 2019, summer 2020, summer 2021
Machine Learning: winter 2017, winter 2018, winter 2019, winter 2020, winter 2021
Causality: summer 2018, summer 2019, summer 2021
Algorithms and Data Structures: winter 2015, winter 2018
Database Systems: winter 2015, winter 2017
C Programming: summer 2016
Theoretical Foundations of Computer Science: summer 2016

News

August 2023: I successfully defended my PhD thesis and am now a doctor!
July 2023: I am co-organizer of this year's ICBINB workshop at NeurIPS 2023.
May 2023: I have received an outstanding Reviewer Award from the ML Reproducibility Challenge 2022.
November 2022: Tom, Jan, Sebastian, Stefan and I have won the Helsinki Tomography Challenge 2022.
April 2022: Our group is joining the Department of Computer Science at the Technical University of Dortmund.
May 2022: I have received an outstanding Reviewer Award from the ML Reproducibility Challenge 2021.
April 2020: I am volunteering for ICLR 2020.
March 2020: I am spending two weeks at the Max Planck Institute for Intelligent Systems in Tübingen.
August 2019: I am participating at the Machine Learning Summer School at the Skolkovo Institute of Science and Technology in Moscow.

This website is based on the template of Jon Barron's website. Used with permission.