I am a PhD student under the supervision of Prof. Stefan Roth. My research focuses on narrowing the gap between human and machine intelligence, especially in the context of visual perception. To this end, I study methods to incorporate human biases and reasoning into modern deep learning approaches.
For ReLU DNNs without bias-terms, we show that Input×Gradient is a closed-form solution of Integrated Gradients that is therefore significantly more efficient to compute than the usual numerical approximation.
Even though deep learning has caused a paradigm shift in most computer vision applications, traditional handcrafted methods are still used extensively for image registration. In this thesis, we show that deep learning also caused a paradigm shift in image registration...
This thesis addresses the problem of aligning connectome data. The advancing scale in connectome reconstruction introduces new problems. One particular problem are spatial shifts and distortions in the image data that make it hard to come to a global alignment...
Software verification of concurrent programs is hampered by an exponentially growing state space due to non-deterministic process scheduling. Partial order reduction (POR)-based verification has proven to be a powerful technique to handle large state spaces...
GPA: 1.11 (with honors)
GPA: 4.0 / 4.0
GPA: 1.65
My research areas include robust and explainable deep learning methods. Further, I assist in computer science classes (e.g. Computer Vision II) and supervise multiple students (theses & projects).
In order to improve image alignment methods, I was working on a novel self-supervised end-to-end keypoint detection and description network. My implementation was released under an open source license allowing me to use it as the basis of my master’s thesis.
My internship was about classifying 3D scans of highly reflective surfaces. In order to classify the scans, I trained a PointNet neural network that is invariant to permutations in the input, and hence, capable of working with point clouds.
I worked freelance on different projects related to web and mobile applications.
Within the context of my bachelor’s thesis in the field of deep learning and computational neuroscience, I interned at the department of connectomics (Prof. Helmstaedter, MPI Brain). In extension of the PointNet neural network, which can be used for point sets, I have developed an autoencoder that extracts features from 3D models of nerve tissue.
I have investigated and implemented algorithms to obtain and reduce the state space of concurrent programs. Additionally, I was involved in a project where a new algorithm was proposed that requires less overhead during state space exploration and shows an improvement in the runtime performance compared to existing algorithms (Metzler et al. 2016).
Together with a team of fellow students, I programmed a robot arm using Robot Operating System (ROS) to automatically scan three-dimensional objects. The scans were then classified and stored to build a knowledge base. I was highly involved in the registration process of the different views and in the classification of the objects.