Learning new movements individually without professional assistance can be difficult. Therefore, a system was implemented that uses inertial measurement units (IMUs) and different visual pose estimation techniques in order to generate raw IMU accelerations from computer vision and 3D trajectories from IMU data respectively. Additionally, the system allows to compare IMU data and poses in the domains of acceleration as well as position, which enables a multi-level comparison and generates deep insights into the performed movements. Finally, a dashboard was implemented in order to visualize the different data sources and calculate insights about Key Performance Indicators interactively.

Self-supervised methods learn representations from unlabeled data, to be reused in subsequent downstream tasks. To use a deep neural network in a downstream task of interest, the network is pre-trained with a self-supervised method on a larger unlabeled dataset. The pre-training dataset can be the same one used in the downstream task, or a similar one that is larger with favourable features to learn. Afterwards, we use transfer learning to apply the pre-trained model on the target annotated dataset. Our experimental results on Kaggle retinopathy classification dataset (2D) and Decathlon pancreas tumor segmentation dataset (3D), illustrate the benefits of these methods on multiple downstream tasks.

In current laser cutting practice, users share 2D files that describe the cutting path for the laser to execute. The drawback is that it is highly specific to the actual process executed by the laser. Reliable reproduction of a model requires describing the result, not the process of making it. We propose to use 3D models which make reference to the 2D plates the machine ultimately cuts. We developed a software tool to convert conventional models to a portable format by reconstructing their 3D assembly. The resulting files are worth sharing; other users can reliably reproduce them on their hardware.

Nina Kiwit, Jonas Cremerius, Margaux Gatrio, Melanie Hackl (Students)

Suparno Datta, Ariane Sasso, Dr. Girish Nadkarni, Prof. Dr. Erwin Böttinger (Supervisors)

This research project deals with the prediction of the onset of essential hypertension in patients based on electronic health records (EHR) using advanced machine learning methods. Hypertension is one of the most prevalent chronic diseases worldwide and is a predisposing factor in a range of diseases. As hypertension is often unnoticed by the patients, a prediction model to forecast the onset of hypertension could play a vital role. Longitudinal EHR data from the Mount Sinai hospital system was used for this project. We were able to achieve AUCs upto 87%.

Andrea Eoli, Mirko Krause, Sam Matthews, Hao Nguyen (Students)

Dr. Hanna Drimalla, Alex Charney, Ben Glicksberg, Prof. Dr. Erwin Böttinger (Supervisors)

Clinical notes of psychiatric patients are a rich resource for gaining a better understand of mental disorders and developing better phenotypes. In this master project, we use natural language processing on clinical notes of electronic health records (EHR) from Mount Sinai hospital system to develop meaningful language-based representations of patients with depression. With unsupervised machine learning, we aim to find categories that are closer to underlying mechanisms as well as subcategories that could inform treatment decisions.

Antonia Winne, Dennis Kipping, Julian Sass, Stephan Krumm (Students)

Jan Philipp Sachs, PhD Riccardo Miotto, Prof. Dr. Erwin Böttinger (Supervisors)

Non-specific back pain is the most frequent cause for absence from work and disability worldwide and holds a huge potential for data-driven knowledge discovery approaches. In this master project we wanted to leverage that potential with Electronic Health Record (EHR) data. We defined a robust phenotyping algorithm, applied descriptive statistics and Natural Language Processing on the resulting cohort and its data, and we applied clustering techniques to identify potentially clinically relevant, yet undiscovered subgroups.

The Travelling Salesman Problem (TSP) asks for a fastest round-trip through a set of cities. Good approximations for TSP usually require that travelling from city A to city B takes as long as travelling from B to A.
This may not seem like a restriction at first, but in reality, blocked roads, one way streets or traffic may violate the required symmetry. In this project we therefore specifically studied the effects of asymmetry on approximate solutions to the TSP. We developed generalizations of two classical approximation strategies for symmetric TSP to also efficiently compute provably good solutions for instances with a moderate amount of asymmetry.

Programming is a challenging mental task, because programmers have to simulate program behaviour in their head during development. Example-based live programming helps to bridge this mental gap by providing live feedback about program execution during development. Existing implementations for example-based live programming, however, depend on specific features of individual programming environments and languages.
To ease the adoption of live programming for other languages, we developed an approach to support language-agnostic live programming in widespread code editors. Our prototypical implementation uses GraalVM as a multi-language execution environment and builds on top of the Language Server Protocol (LSP) to provide Babylonian Programming in a language-agnostic way in code editors such as Visual Studio Code.

When and how we use a certain word changes over time. For example, while kids nowadays think of the internet when they hear "cloud", their grandparents probably think of the weather. In our Master’s project, we developed an approach to detect words in our language that gained a new sense over time. Our graph based language model enables us to not only look at word relationships, but also to visualize them. We had our algorithm read hundreds of books and magazines published within the last 200 years.

Energy consumption is one of the major cost factors when operating a data center. Hence, monitoring the energy consumption of a data center is crucial for operation. This requires a comprehensive overview of the energy distribution among servers and running jobs. We propose a power monitoring system for heterogenous data centers. The system collects resource and power consumption metrics on data center-, node-, and process-level. It supports data center operators in monitoring the total energy consumption, detecting jobs with high energy profiles, and planning the allocation of new resources. We deploy and evaluate the system on the HPI Future Soc Lab, an on-site data center providing computational power to the research community.

We present HDES, a dynamic stream processing engine (SPE). Our engine focuses on enabling ad-hoc queries to allow end-users to dynamically add and remove queries. While approaches like AStream and AJoin introduced ad-hoc functionality on top of common open-source SPEs HDES aims to make dynamicity a first-class citizen. This change in perspective allows us to optimize this use-case upfront and elevates the performance of ad-hoc queries. Moreover, we include resource sharing optimizations to increase the performance of expensive operations such as joins.
 

In the context of business process management, resources are often reduced to active entities, in most cases humans, that execute tasks. However, this perspective neglects the importance of the resources needed in order to perform certain tasks. Like finding the most suitable surgeon is a good start, but is of little use without an available operation room. In this work, we propose a more holistic approach on resources in BPM by introducing a platform capable of managing resources and allocating them based on arbitrary algorithms during process execution. We also present a proof of concept implementation and evaluate it with a real world use case.