School of Computation, Information and Technology

Artemis is a Learning Management System (LMS) that facilitates programming exercises using a Version Control System (VCS) and a Continuous Integration System (CIS). Since some VCSs include a Continuous Integration (CI) application, it is possible to reduce the effort required to maintain the servers and migrate the CIS from external applications to the integrated system.
In the existing implementation, programming exercises rely on a configuration stored in the CIS as part of build plans. These configurations are duplicated for every student participation, making it hard to update the configuration throughout the exercise lifecycle.
GitLab CI is the CIS provided by the GitLab VCS. This system supports provisioning build plans via the Artemis Application Programming Interface (API), which makes it possible to update the build plan retrospectively, as the build plan is fetched from the API on each run.
In this thesis, we show how Artemis can be extended to support an VCS with an in-built CIS. We analyze the underlying problem, formalize the requirements, and model the adapted system design. We implement the changes as part of the open-source Artemis system. We generalize the existing Jenkins Server Notification Plugin for usage with different CISs to notify Artemis. Artemis is now capable of conducting Java programming exercises with a single dependency on GitLab.

Deep Learning has received much attention in the past few decades and is recognized as being one of the most important methodologies for potential breakthroughs in the medical field. This work is a preliminary attempt to introduce a multifaceted view on static and dynamic learned MRI reconstruction. First, we investigate the problem of domain shift in the context of state-of-the-art Magnetic Resonance Imaging (MRI) reconstruction networks with respect to variations in training data. We provide visualization tools and support our findings with statistical analysis for the networks evaluated    on the 1.5T/3T fastMRI knee and neuro data. We observe that the signal-to-noise ratio of the examined sequences plays an essential role, and we statistically prove the hypothesis that both the type and amount of training data are less important for low acceleration factors. Finally, a visualization tool facilitating the examination of the networks’ performance on each individual subject of the fastMRI data is provided. In the second part of this research paper, we adapt existing methodologies and findings in MRI reconstruction to novel 7T MRI reconstruction of dynamic cardiac processes. Despite the less severe degradation of image quality at high acceleration factors for higher magnetic field strengths, 7T MRIs, especially those with an additional temporal dimension, have rarely been studied directly in the field of learning-based undersampled MRI reconstruction. In this context, we identify the substantial impact certain neural network architectures and configurations have on reconstruction quality. In summary, we conclude that findings on domain shift, reconstruction quality and factors impacting network performance in 1.5T and 3T MRIs can also be translated to other magnetic field strengths, facilitating the transition to ultra-high-field (7T) and low-field MRI.

We consider a distributed machine learning algorithm in the presence of stragglers. We focus on algorithms running approximate gradient decent. We compare the performance of two algorithms that aim at mitigating stragglers, which are both based on gradient cod-ing, namely ErasureHead and Stochastic Gradient Coding. Even though either approach gives theoretical guarantees for their respective performance, we compare the practical performance in the same environment using the same data and the same loss function. We run the simulations using python programming language on a single machine with parallelized worker threads. We find that ErasureHead is a faster solution for our pur-pose, as it decreases the loss function faster. However, Stochastic Gradient Coding has a smaller margin of error after the same number of iterations, the run time of which is much longer.

– Leopold Thomas, OvTG

In dieser wissenschaftlichen Arbeit wird das Phänomen der Lautheitsintegration bei Nutzern eines Cochlea-Implantats (CI) untersucht. Bereits vorliegende Untersuchungen zeigen, dass das menschliche Gehör akustische Signale von einer Zeitdauer bis zu 200 ms integriert. Dies hat zur Folge, dass längere Töne als lauter und kürzere Töne (kleiner als 200 ms) als leiser empfunden werden. Hierfür wurden zwölf Normalhörende und sechs CI-Nutzer untersucht. Für die Messungen dieser Studie befinden sich die Testpersonen in einer schallisolierten Kammer, in welcher mithilfe der Methode des Einregelns die Ruhehörschwelle der Testpersonen bei unterschiedlicher Stimulusdauer bestimmt wird. Die Ergebnisse dieser Experimente bestätigen, dass Lautheitsintegration auch bei Nutzern eines Cochlea Implantates auftritt. Der durchschnittliche Unterschied, der Ruhehörschwelle bei einer Zunahme der Tondauer von 10 ms auf 1s, war bei den CI-Nutzern deutlich geringer als der, der normalhörenden Testpersonen. Daher ist die Lautheitsintegration bei CI-Nutzern schwächer ausgeprägt. Dies führt im Alltag dazu, dass kurze Tonimpulse, wie das Klappern von Geschirr, für CI-Nutzer lauter und dadurch als unangenehm wahrgenommen werden können. Schlussendlich stellt sich damit die Frage, ob man diese veränderte Wahrnehmung durch eine Änderung in der Programmierung bzw. Einstellung des Cochlea-Implantates ausgleichen könnte. Dies würde zu einer natürlicheren Lautheitswahrnehmung im Alltag der CI-Nutzer führen.