The first session of our weekly seminar (doctoral colloquium) is scheduled at 15:15. The event will take place - according to the current status - virtually via zoom.us.

We will use the first session for a brief introduction of new members (1), for a general discussion (2) and for the formal organisation of all relevant tasks and obligations within the SSE Research School (3).

Business Process Management has proved to make a difference in organizational performance and competitiveness to the industry at large. However, health care providers, such as Hospitals, are barely aware of process management and lack a holistic view of their patient treatment. The aim of my research is to explore how BPM principles can help to achieve better health care management and identify the application of BPM principles within health care. This presentation gives an introduction to Digital Health, my projects conducted in health care, and my research topic.

Synchronization – the spontaneous adjustment of rhythms among coupled oscillatory systems – is a fascinating dynamical phenomenon found in many biological, social, and technical systems. Biological examples include networks of pacemaker cells, congregations of synchronously flashing fireflies and many more. This presentation gives an introduction to synchronization and possible research topics for my PhD.

Towards an envisioned interaction paradigm where computing is more seamlessly integrated with the users' everyday mobility, voice-based interaction is likely to play a pivotal role as speaking is a natural form of human communication and also, offers an untethered and device-independent channel of communication with a computing interface. Additionally, the interaction is hands- and eyes-free, leaving the users free to engage in other tasks. Yet, to lower the interaction burden when the interactions are embedded into the users' everyday mobility and activities, the traditional (existing) interaction vocabulary for on-the-go (mobile) interactions needs to be redesigned under the new paradigm. To this end, the talk presents voice-based and multimodal interaction techniques for text input/editing as an everyday mobile computing task for both eyes-free interfaces and heads-up computing-based interfaces like Augmented Reality Smart Glasses (ARSG), that better support flexible and natural interactions consistent with the users' mobility needs.

Automated and Fine-Grained Memory Tiering for In-Memory Databases

By moving less frequently accessed data from DRAM to lower memory and storage tiers, such as non-volatile memory and SSD, an in-memory DBMS can reduce its DRAM footprint. Current approaches either statically move pre-defined parts of the data or depend on the DBA's selection of suitable data. By enabling the DBMS to make automatic tiering decisions, more efficient configurations can be found in cases where static or manual approaches fail. As a research and evaluation platform, we have (1) built a new DBMS whose TPC-H results are on par with state-of-the art research systems. We (2) contribute low-overhead, iterator-based access counters that track all data accesses. Not only can we identify hot, warm, and cold data, but can also distinguish between different access patterns. We further (3) introduce a lightweight memory abstraction layer to allocate data on DRAM, NVM, or secondary storage. This way, we can move table and temporary data from one memory tier to another without having to pause query processing. Our (4) Dynamic Tiering Plugin uses the tracking information and automatically improves the memory utilization.

How would today's hardware start-ups manufacture their products? The known way is to build hardware and video prototypes and mass produce them with a venture capital fund. However, there are several problems with this method. For example, a lot of money has already been wasted before improving the product based on user feedback. An alternative is bootstrapping, i.e., gradually increasing production volume and profits without the support of a large funding source. However, in this case, the founders have to choose their manufacturing method. This decision-making process is time-consuming and difficult for people with no manufacturing experience. We will create a system to support the founder's choice of a manufacturing method in the bootstrapping type of small-batch production. In this talk, I will present the motivation, system I'm thinking, and related work.

The best immersion into virtual reality is achieved through a one-to-one mapping from physical to virtual user movement. However, such mapping only works with sufficient physical space and is generally not feasible when users are in confined spaces such as during public transportation. We present embrI/O, a system that detects the onset of its user's movement (e.g., hips or shoulders), and extrapolates this onset to the respective complete gesture the user intents to perform, such as punches. The user does not need to perform the rest of the gesture. Instead our system controls the animation of the virtual avatar accordingly. Our setup prevents the user from reaching out of the their confined space by restricting their physical movement through wearable sensing straps. In this presentation, we show our current hardware prototype and software architecture of a real-time gesture recognition and animation reconstruction system.

Block-based programming interfaces support tool integration, as well as domain-specific interfaces, and present a good way for programming novices to start learning. However, current block-based programming interfaces are designed for programming novices, and lack efficiency for repetitive interactions, hindering adoption from programming professionals.
In this talk, we present a short comparison of different programming interfaces. We then present our prototypical block-based editor interface that supports efficient editing. We further demonstrate tools that have been built using the interface's framework to argue why we believe block-based programming could play an important role in modern integrated development environments.

Growing data volumes and velocities in fields such as Industry 4.0 or the Internet of Things have led to the increased popularity of data stream processing systems. Enterprises can leverage these developments by enriching their core business data and analyses with up-to-date streaming data. Comparing streaming architectures for these complex use cases is challenging, as existing benchmarks do not cover them. ESPBench is a new enterprise stream processing benchmark that fills this gap. We present its architecture, the benchmarking process, and the query workload. We employ ESPBench on three state-of-the-art stream processing systems, Apache Spark, Apache Flink, and Hazelcast Jet, using provided query implementations developed with Apache Beam. Our results highlight the need for the provided ESPBench toolkit that supports benchmark execution, as it enables query result validation and objective latency measures.

The goal of my research is to allow virtual reality experiences to be run in arbitrary tracking volumes and with arbitrary physical objects. VR experiences today are designed with a specific tracking volume and  objects in mind, such as “square 5x5m space with a rubber sword”. This prevents experiences from running with different objects or in tracking volumes of smaller size or different shape, making it impossible to share experiences, especially with home users. I  address this by creating an abstraction between VR applications and the  space and physical objects they are using. Instead of accessing space and physical objects directly, in my system  applications express their needs in an abstract way, which my systems  then maps to the actual available physical space and physical objects. This allows VR applications to run on a wide range of installations. Solving this problem would have substantial commercial  impact, as the proliferation of real-walking VR is currently hindered by  developers’ reluctance to require users to have space and objects. My work is inspired by operating systems research. Before opening systems, application programs were written for a specific  machine. Operating systems allow applications to run on arbitrary  computers and architectures by creating an abstraction of the physical  hardware, an API, that allows applications from accessing the hardware directly.

In this last talk in the research school I want to share an updated draft of my defense talk, for you to give feedback and input on. 

The digitalization of the railway infrastructure proceeds day by day. Together with new standards developed by the Eulynx initiative, this does not stop at the German border anymore and the market of infrastructure elements, like switches or signals, will be opened for many new manufactures. This opens new challenges for the infrastructure operators like the DB Netz: Every new infrastructure element must be certified for several quality characteristics like interoperability, dependability, and availability in several different environments. Currently, this is a task for a test executer, who runs serval tests manually. But in the future, the goal is to have an automatic system, which automatically certifies new infrastructure elements. This presentation gives a short overview of this vision, challenges in this vision, and the current state of my work and research in that area.

Since last decade it is observed the rapid growth in utilizing machine learning models in medicine. More and more organizations aim to use patients' vital data to improve quality of not only doctors' work but also patients' care. In my presentation I would like to picture two projects I am involved into. I will present theirs progress, limitations and challenges. The first one is a part of Telemed5000 project and is made in collaboration with Charité, Berlin. The aim is to devise a Decision Support Model for predicting the daily per-patient risk of a critical condition for a patients with cardiovascular diseases (CVDs). It would help doctors to prioritize patients, who are concerned to need intervention. The second project is conducting in collaboration with Getemed AG and is called WiseMat. The target is to devise a model, which will be capable of predicting an occurrence of bedsores (Dekubitus) based on pressure in the mat. It would decrease need of trained nursing staff to look after patients who need constant healthcare.

Laser cutting is a fast fabrication technology that enables rapid prototyping. The typical workflow for fabrication using lasercutting has three steps; modelling the object, lasercutting and then assemble the 3D object from 2D parts. Traditionally, modelling/modifying the objects was done in 2D which resulted in this step being the most time consuming and cumbersome part. However, in recent years researchers have proposed systems that allow users to model in 3D which has sped up the modelling process.

This has resulted in assembly, which requires users to search for parts with matching edges and join them together, replacing modelling as the most time consuming part of the workflow. The reason is that these parts are placed randomly and packed together for space efficiency resulting in an O(n^2) search. As a solution we present roadkill, a system that creates fast to assemble layouts for laser-cut objects. Roadkill achieves this by preserving the adjacency relationship between plates effectively reducing the search to a linear (O(n)) process. In our pilot evaluation we found that models are assembled twice as fast using layouts made with Roadkill.

A virtual spatial model can be described as a complex amalgamation of various data and objects, which are made visible via the spatial component. The result is a multidimensional information medium in the major application fields of economy, society and environment, serving as an interface and communicator of various data, objects and issues. An object in a geographical context consists - usually - of the three components spatial reference (1), geometric representation (2) and semantic expression (3). The spatial reference forms the link between different data and objects and is directly or indirectly considered in almost every application. If it is missing due to insufficient base data, only a limited amount of reliable knowledge is attainable.

In this context, the example of the agglomeration area of East Jerusalem is used to demonstrate how, in the absence of reliable reference data, functionally rich virtual multidimensional city models can be generated, derived, visualised and made usable in an application-oriented manner. Sub-complexes (data acquisition, processing, visualisation, evaluation and analysis) are prototypically implemented in a generic, service-oriented comprehensive process. Within the framework of an ongoing R+D project, the objective is to provide civil society and decision-makers with a high-resolution city model with dedicated spatial data and system functions that allow the fulfilment of civilian societal tasks.

Software programmers can spend up to 40 percent of their time searching for the causes of software failures. To alleviate that, programmers use various debugging techniques to reduce the search space for software faults (bugs). However, debugging techniques are hindered by a phenomenon known as "imperfect fault understanding”. This phenomenon consists of programmers generating incorrect or incomplete explanations about how a putative problem in the source code is causing a failure. This phenomenon has severe impacts as it wastes programmer's time with invalid bug fixes and, in the long run, makes teams lose confidence in their debugging tools and techniques.

I study how to mitigate the "imperfect fault understanding" phenomenon by predicting the accuracy of the programmers' understanding of a software fault during a source code inspection task. For that, I built causal and sequential decision models from data that I obtained from three large experiments that I planned and executed with more than 1000 programmers who performed more than 6000 tasks. These tasks consisted of inspecting the source code to identify the cause of failure in real bugs from popular open source software projects. The causal models allowed to make predictions and determine how to intervene in the inspection process to maximize its accuracy and efficiency. I evaluated these causal models in a realistic setting, for which I instantiate different sequential decision mechanisms based on contextual multi-armed bandit algorithms.

My results are promising in a sense that I could uncover a set of prediction factors and determine task interventions that allow to correctly locate all the real bugs with more than 90% precision and with less than 20% of the total tasks available for code inspection. I close my talk with a brief discussion on how these results could have implications to the processes and tools supporting on-boarding, bug triage, and software debugging.

To optimize the throughput of FPGA accelerator resources and facilitate their applicability in realistic applications featuring multiple concurrent tasks, multi-tenancy is a desirable feature.
Based on our currently single-tenant research prototype MetalFS, I am adapting the data stream infrastructure to handle data flows from multiple different task contexts at the same time.
The resulting architecture must provide mechanisms to ensure isolation between independent task contexts.
Furthermore there is a large design space to choose different tradeoffs between overall throughput, single-task performance, hardware complexity and development overhead for users of the MetalFS framework.

In this talk, I will give a short progress report of my work on MetalFS. Then I will dive into a discussion of the expected performance and alternatives to the current hardware design, based on simulated results.

Intrinsic image decomposition refers to the problem of decomposing an image into intrinsic scene characteristics such as albedo, shading, and specularity. The inverse problem setting is ill-constrained considering multiple unknowns with a single known pixel-value. On the other hand, Neural Style Transfer (NST) refers to a technique wherein feature distribution from a "style" image is transferred to a "content" image using neural networks. NST is quite promising when it comes to image-based artistic rendering and in the past few years, multiple variants of this approach have been proposed. However, NST suffers from challenges such as structure preservation, interactive controls, relighting, etc. In this talk, I will present my initial findings of how we can mitigate the problems w.r.t. NST using intrinsic images.

Process mining is widely embraced in many organizations to gain a deep knowledge about the data and the running processes. This can be achieved by applying different process mining techniques like discovery, conformance checking, performance analyzing etc. All these techniques are purely applied on event logs, which are before extracted from the organization’s database. This implies the need of having access to the database and deep knowledge of the schema. However, this might not always be possible and we argue that another kind of log is more readily available—Redo log. Such log is used to bring the system into consistent state in case of a system failure. This paper proposes a semi-automatic approach to extract an event log from redo logs in the absence of a database. The feasibility of the proposed approach is evaluated on two synthetic redo logs.

In large-scale microservice architecture, reliability issues can be caused by a variety of reasons, such as restricted host resources, unavailable hardware or software, unstable networks, etc. To keep the system running reliably, cloud operators often deploy large-scale performance checkpoints and metric collection mechanisms. These monitoring mechanisms can help us initially verify that applications and services are running healthily. However, since the applications run on a complex architecture consisting of various microservices, it is a challenging task to reveal the mechanism of anomaly propagation from massive monitoring metrics, and to pinpoint the root cause of the failure. In conclusion, traditional anomaly diagnosis methods are usually based on key performance indicator (KPI) thresholds. System administrators set the KPI monitoring threshold manually according to their domain knowledge for early warning. However, due to the very large number of services in the MSA application and the complex dependencies between services, it is difficult for system administrators to detect anomalies by setting reasonable KPI monitoring thresholds, let alone diagnose root causes in fine granularity. So, Automated root cause analysis reduces the overall dependency on expert knowledge and provide solution to failure propagation in Microservice Architecture.