Research and Teaching

  • Domain Adaptation for Machine Translation and Speech Processing

    Facebook AI Research, Menlo Park, CA since 10 /2018

    I am working on neural machine translation and speech processing with Michael Auli.

  • Domain Adaptation in Brains and Machines

    Bethge Lab, Max Planck Institute for Intelligent Systems 04/2018 - 10 /2018

    I worked on domain adaptation algorithms and their connections to adaptation mechanisms in biological learning systems. Results will become gradually available at

  • Deep Learning for Molecular Imaging

    HelmholtzZentrum München 05/2017 - 05/2018

    Working in Gil Westmeyer’s group on novel image processing techniques. I work on real-time deep learning based image processing techniques for temporal 2D and 3D imaging data, with the goal of enabling closed-loop cell-circuit-control with spatiotemporal precision and deep tissue penetration in zebrafish.

  • Deep Learning for Transfer Learning and Time-Series Analysis

    School of Computing, University of Kent 03/2017 - 05/2017

    I am currently on a research visit in the labs of Caroline Li and Prof. Yi-Ke Guo and work on unsupervised learning methods for analysis of time-series data such as EEG.Implementation of our approaches is realized in TensorFlow and TensorLayer.

  • Deep Learning for Medical Computer Vision

    Institute of Imaging and Computer Vision, Aachen 05/2016 - 02/2018

    Within the ILUMINATE project, I am working on deep learning algorithms for semi-supervised dense classification of histopathological images used in cancer research. Apart from deployment of networks in our software system, I worked on a novel approaches to apply deep learning in contexts with little available labeled training data. Used software packages are mainly Theano and TensorFlow.

  • Teaching Assistant

    RWTH Aachen University 09/2014 - 06/2015

    Winter Term 2014: Mathematical Methods in Electrical Engineering, Prof. Merhof, Institute of Imaging and Computer Vision, Aachen, Summer Term 2015: Fundamentals of Electrical Engineering II, Prof. DeDoncker, ISEA, Aachen

  • Computer Vision for Robotics, Software Intern

    Institute for Real-Time Learning Systems, University Siegen Summer 2013

    Development of a software system for automated calibration of 3D camera systems as a preparation for sensor fusion algorithms, using C++, the Point Cloud Library and ROS.


  • Student Engineer, Perception and Sensor Group

    TUfast e. V. Driverless Racing Team 2016 - 2017

    At TUfast, we are developing an autonomous version of a Formula Student Racecar to participate in the Formula Student Driverless competition in Hockenheim. I work in the Sensors and Perception group on deep learning approaches for processing of sensor inputs.

  • Student Engineer, Control Systems

    Formula Student Team RWTH Aachen e. V. 2014 - 2016

    I worked on the hardware and software design of data acquisition devices and the battery management system in the Formula Student racecars eace04 and eace05.

  • Founder and Project Manager

    IT4Kids, Enactus Aachen e. V. since 2013

    To provide children in primary school with courses in computer science, I founded IT4Kids in 2013 and build up the student initiative that is still active in Aachen as of now. With our classes, we have reached hundreds of pupils. The project was awarded the 3rd place at the Enactus National Competition 2015 and a winning project in the Google Impact Challenge (awarded 10.000 €). We also started the development of a flexible programming environment for pupils, combining advantages of Scratch and Python.

Publications and Talks

Transfer learning for adaptation to new tasks is usually performed by either fine-tuning all model parameters or parameters in the final layers. We show that good target performance can also be achieved on typical domain adaptation tasks by adapting only the normalization statistics and affine transformations of feature maps throughout the network. We apply this adaptation scheme to supervised domain adaptation on common digit datasets and study robustness properties under perturbation by noise. Our results indicate that (1) adaptation to noise exceeds the difficulty of widely used digit benchmarks in domain adaptation, (2) the similarity of the optimal adaptation parameters for different domains is strongly predictive of generalization performance, and (3) generalization performance is highest with training on a rich environment or high noise levels
NIPS 2018 Workshop on Continual Learning

We introduce salad, an open source toolbox that provides a unified implementation of state-of-the-art methods for transfer learning, semi-supervised learning and domain adaptation. In the first release, we provide a framework for reproducing, extending and combining research results of the past years, including model architectures, loss functions and training algorithms. The toolbox along with first benchmark results and further resources is accessible at
NIPS 2018 Workshop on Machine Learning Open Source Software

This study presents a novel data-driven approach to detect desynchronization among biosignals from two modalities. We propose to train a deep neural network to learn synchronized patterns between biosignals from two modalities by transcribing signals from one modality into their expected, simultaneous or synchronized signal in another modality. Thus, instead of measuring the degree of synchrony between signals from different modalities using traditional linear and non-linear measures, we simplify this problem into the problem of measuring the degree of synchrony between the real and the synthesized signals from the same modality using the traditional measures. Desynchronization detection is then achieved by applying a threshold function to the estimated degree of synchrony. We demonstrate the approach with the detection of eye-movement artifacts in a public sleep dataset and compare the detection performance with traditional approaches.
NIPS Time Series Workshop 2017

A long-standing objective in neuroscience has been to image distributed neuronal activity in freely behaving animals. Here we introduce NeuBtracker, a tracking microscope for simultaneous imaging of neuronal activity and behavior of freely swimming fluorescent reporter fish. We showcase the value of NeuBtracker for screening neurostimulants with respect to their combined neuronal and behavioral effects and for determining spontaneous and stimulus-induced spatiotemporal patterns of neuronal activation during naturalistic behavior.
Published in Nature Methods (October 2017)

While human observers are able to cope with variations in color and appearance of histological stains, digital pathology algorithms commonly require a well-normalized setting to achieve peak performance, especially when a limited amount of labeled data is available. This work provides a fully automated, end-to-end learning-based setup for normalizing histological stains, which considers the texture context of the tissue. We introduce Feature Aware Normalization, which extends the framework of batch normalization in combination with gating elements from Long Short-Term Memory units for normalization among different spatial regions of interest. By incorporating a pretrained deep neural network as a feature extractor steering a pixelwise processing pipeline, we achieve excellent normalization results and ensure a consistent representation of color and texture. The evaluation comprises a comparison of color histogram deviations, structural similarity and measures the color volume obtained by the different methods.
Published at the 3rd Workshop on Deep Learning in Medical Image Analysis (DLMIA) at MICCAI 2017

Electrophysiological recordings are usually prone to artifacts that are commonly removed by visual inspection and signal decomposition techniques such as ICA in clinical and research contexts. In this work, we propose a generative deep learning model for EEG signals inspired by the WaveNet architecture conditioned on biosignals from other modalities (EOG, EMG and ECG) that are expected to capture relevant context information on artifacts that are simultaneously recorded in the EEG signal. As our main contributions, we show that it is possible to capture non-trivial correlations between time-series from multiple modalities in a fully-convolutional neural network. By dampening the context signals during inference, it is possible to remove the caused artifacts in the EEG signal by (1) generating a possibly uncorrelated EEG sequence from a dampened version of the context signal (2) estimating the context signal’s contribution to a particular kind of artifact and removing this effect from the original EEG sequence, preserving the signal information. In this poster, we present preliminary results on the openly available MASS cohort-3 sleep database, comprising recordings from various patients recorded during sleep.
Poster Presentation at the 13th Bernstein Conference 2017, Göttingen, Germany

Deep Learning has transformed the ways in which we approach data science and machine learning in a variety of fields and disciplines substantially. Problems in computer vision, speech recognition and synthesis, protein folding prediction and many more fields can now more or less easily be tackled. In this talk, I will give a comprehensive introduction to the recent advances in the field and their relevance to concrete problems in medical and biological applications. As supervised learning in these domains usually falls short due to a lack of data, techniques for transfer learning and domain adaptation between datasets as well as the quantification of uncertainty in deep neural networks.
Poster Presentation and Talk at the 4th HBP Summer School in Obergurgl, Austria

Compressed sensing has proven to be an important technique in signal acquisition, especially in contexts in which sensor quality or the maximum possible duration of the measurement is limited. In this report, deep learning techniques are used to improve compressive sensing in the context of image acquisition. In a previous approach, stacked denoising autoencoders capable of reconstructing images considerably faster than conventional iterative methods were deployed. Apart from reviewing this approach, a possible extension using convolutional autoencoders inspired by the popular VGGnet architecture is discussed. Instead of learning models from scratch, a simple yet effective way for adapting available filters used in ImageNet classification is presented. By reformulation of the autoencoder structure in terms of a fully convolutional network, the previous approach can be adapted to arbritrarly large images for efficient learning of the measurement matrix and sparsity basis. Suggestions on the real implementation of such as system conclude the report.
Seminar Paper (Unpublished)

In this thesis, various methods for the design of deep architectures for tissue classification are presented. By using transfer learning and unsupervised feature learning, it is shown that powerful state of the art models with millions of parameters can be finetuned to outperform previous approaches despite the lack of sufficient labeled training examples. Several models such as the 16-layer VGGnet, the GoogLeNet model with some exten- sion, convolutional restrict boltzman machines and convolutional denoising autoencoders were trained on the ILUMINATE-9 dataset. Along with evidence provided on how the training policy for the networks should be designed, a whole model zoo, trained on the ILUMINATE-9 dataset, is provided along with this thesis.
Bachelor Thesis at RWTH Aachen University (Unpublished)


KI macht Schule

KI macht Schule provides classes in AI & Machine Learning for German high school students

TUFast Driverless nb017

Development of a fully autonomous racecar for the FSG Driverless competition in Hockenheim.


M.Sc. Neuroengineering Student Blog with latest information about our study program and events.

Ecurie Aix eace04/05

Contributions include the design of electric control units


ILUMINATE develops a novel platform for integrated analysis of in-vivo models for preclinical evaluation of new compounds in oncology, including innovative therapeutic approaches in oncoimmunology.


IT4Kids provides computer science classes to elementary school pupils - Providing software, teaching materials and easy communication between schools and teaching students