The first session of our weekly seminar (doctoral colloquium) is scheduled for 3:15 pm. The event will - according to the current situation - take place virtually via zoom.us.

We will use the first session for (1) a formal organisation and assignment of all relevant tasks and obligations within the SSE Research School and for a (2) general discussion.

HPI Research Symposium 2021

Conrad Lempert will give his introductory talk and present his Master's project ‘Assembler3’ at HCI Lab under the supervision of Thijs Roumen.

Assembler3 is a software tool that allows users to perform 3D parametric manipulations on 2D laser cutting plans. Assembler3 achieves this by semi-automatically converting 2D laser cutting plans to 3D, where users modify their models using available 3D tools (kyub), before converting them back to 2D. This workflow is 10x faster than using the traditional approach of editing 2D cutting plans directly. Assembler3 converts models to 3D using a 5-step algoritm. Once reconstructed, we expect users to store and share their models in 3D, which can simplify collaboration and thereby empower the laser cutting community to create models of higher complexity.

Tolga Buz will present the E-School. Tolga is PhD Candidate at the Chair ‚Artificial Intelligence and Intelligent Systems‘ since October 2020 as well as Startup Coach and Program Manager at the School of Entrepreneurship at HPI.

The HPI School of Entrepreneurship here for you when it comes to founding a business.

We inspire, motivate and support you. Every day. From finding suitable business ideas, through the first tentative thoughts of a start-up, to successful company growth.

We offer an inspiring ecosystem and a lot of formats that individually help entrepreneurs to bring together all the necessary ingredients: Mentoring and coaching, methods around ideation and business ideas testing, business know-how and tool sets, finding co-founders, funding, network access, collaboration partners and much more. We do everything with the goal of getting people excited about starting up and making a difference together.

https://hpi.de/en/entrepreneurship/hpi-e.html

Felix Grzelka will give his introductory talk and present the results of his master's thesis:

Training artificial neural networks on GPUs is time and energy-intensive. Modern state-of-the-art networks in 2019 had hundreds of millions of parameters, while most networks in 2020 had multiple billions of parameters. It is estimated that training such networks costs around 1$ per 1000 parameters. Reducing the energy for training by just ten percent could potentially save millions of dollars and reduce the environmental impact.

The focus of my thesis is to reduce the energy consumption during the training of deep neural networks. The necessary tooling was built to collect GPU energy data. Models, which can predict the impact of power limits on energy consumption, runtime, and energy delay product (EDP), have been developed. Five different benchmarks — four deep neural networks and one scientific simulation — have been run on three different Nvidia GPUs: V100, T4, and K80. The data collected from these experiments was used to build and validate models which can predict energy and EDP based on a given power limit. Model and data show that the optimal power limit to reduce the energy consumption is usually the lowest available power limit (150 W V100), (60 W T4), (125 W K80). At best, a 30 % energy reduction is possible on the V100.

To execute tests in the railway infrastructure domain, it's often not enough to only connect devices and test their compatibility and interoperability. It's also necessary to test these devices in real-world setups and configurations. Besides several standard setups, also a lot of special cases need to be considered. Manual creation of all these setups and test cases is way too complex and not flexible enough for future challenges.

A better approach is to generate the test cases from given railway infrastructures. To deal with these infrastructure setups, we are using common industry tools like the PlanPro-Werkzeugkoffer. With this data basis, we are currently developing a generic interlocking, to create the different possible setups, configurations, and operations on these tracks. Thereby we are facing several challenges regarding the complexity of the German railway system and its regulations. In the end, our goal is to have a simulation for an interlocking communicating with other infrastructure elements to handle traffic on a track system.

Laser cutting is a fast fabrication technique. 3D modeling for laser cutting such as with kyub, enables users to quickly produce models for laser cutting. The remaining bottleneck is putting the laser-cut plates together into a 3D model, a search problem of quadratic complexity for users. We developed Roadkill, a software tool that converts 3D models to 2D cutting plans for laser cutting—such that the resulting layouts allow for fast assembly. Roadkill achieves this by putting all relevant information into the cutting plan: (1) Thumbnails indicate which area of the model a set of parts belongs to. (2) Parts with exposed finger joints are easy to access, thereby suggesting to start assembly here. (3) Openings in the sheet act as jigs, affording assembly within the sheet. (4) Users continue assembly by inserting what has already been assembled into parts that are immediately adjacent or are pointed to by arrows. Roadkill maximizes the number of joints rendered in immediate adjacency by breaking down models into “subassemblies.” (5) Users complete subassemblies according to their labels 1, 2, 3…, following 1  > 1 links to insert subassemblies into other subassemblies, until all parts come together.

After streamlining the search process with roadkill, we found that the physical assembly of plates becomes the next crucial bottleneck to further speed up fabrication. In a follow-up project, we achieve another speed-up of a factor of three by switching to foamcore instead of plywood as material. By cutting through 2 layers of this three-layered material, we allow users to fold the plates together instead of assembling them in place. Unlike traditional cardboard/paper folding, foamcore retains its structural integrity because our software places hidden finger joints in the two layers we cut away.

Understanding and improving business processes have become important success factors for organizations. Process mining has proven very successful with a variety of methods and techniques, including discovering process models based on event logs. Process mining has traditionally focused on control flow and timing aspects. However, getting insights about a process is not only based on activities and their orderings, but also on the data generated and manipulated during process executions. Today, almost every process activity generates data; these data do not play the role in process mining that it deserves.

This talk gives a brief introduction into process mining in healthcare and discusses our research dealing with a visualization technique for enhancing discovered process models with domain data, thereby allowing data-based exploration of processes. Data-enhanced process models aim at supporting domain experts to explore the process, where they can select attributes of interest and observe their influence on the process. The visualization technique is illustrated by the MIMIC-IV real-world data set on hospitalizations in the US.

Machine learning models have numbers of hyperparameters that have to be set before training the model. The predictive performance depends on the way how we chose them. Currently, some algorithms for selecting the best levels of hyperparameters have been developed, such as random search or Bayesian optimization. The second one is proven to have better performance in most cases but the limitation of it is no possibility to execute parallel computation on different processing units. The other disadvantage of modern optimization algorithms is the fact that they do not focus on defining important hyperparameters and interactions between them. It makes the data scientist could spend time tuning unimportant hyperparameters. Also, the understating of the above relationship could enlarge our knowledge about algorithms and how they work. We will be more aware of what we are tuning and why we select specific values.

My approach to solving mentioned limitation is including the Design of Experiments (DoE) in the optimization process. DoE are statistical techniques, which determine the influence of single variables and the interaction between them on the output. With it, we are also able to estimate the optimal values of variables to achieve the expected output. The main idea of this methodology is to receive the maximum information with a minimum number of runs. The other advantage is the fact we can conduct parallelization of computing. This methodology is successfully used in such areas as manufacturing, chemistry, and agriculture. In my research, I will focus on comparing my algorithm with the ready ones and define if I can receive good predictive performance and, at the same time, have the possibility to interpret hyperparameters' significance.

Block-based programming supports novice programmers to get started with programming. For professional programmers, it is commonly perceived as too slow and limited. Yet, the block metaphor with its reification of a program's AST carries potential as a platform for deeply integrated tools, especially in a professional context.

In our research, we investigate means to make block-based programming usable with general-purpose programming languages and with high input efficiency. We present the prototype of a block-based programming environment, discuss factors of usability, and present the design of a study for its validation. Finally, we give an outlook on our plans to quickly support porting new languages to block-based environments using existing language grammars.

Speech as an input modality has been used to interface with computers since as early as the 1950s and the 60s, when researchers successfully devised machines that could understand dictated digits and single words. 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. From its inception to where it stands today, how has the field of speech interaction evolved? What are the key opportunities and challenges underlying the field? As a speech interaction researcher, in this talk, I will present a broad overview of the field, introduce the audience to some of the latest research being explored in this field, and talk about where the field is headed towards. Finally, I’ll situate my research on speech-based text input/editing interactions within the context of the broader research space. 

At the current state of the art, practical FPGA designs leverage several abstraction levels and predefined IP-blocks, to specify the actual application logic against a convenient and relatively portable interface. In consequence, an end-to-end simulation of such a design involves the accurate modeling of design parts which the developer would have considered part of the system environment. Cycle accurate models for complex host-interfaces or local resources such as DRAM, HBM or network ports can add an overwhelming overhead over the simulation time of the application logic.
Even though these large simulation costs are necessary in the late development stages of the design, to validate the design with the maximum possible degree of realism, earlier stages can benefit significantly from faster development cycles.

In this talk I will present my approach and prototype for a high-level simulation environment, that disposes with the cycle-accurate simulation of the system-environment. Instead, the system-environment is modeled through less expensive behavioral specifications in software. Only the actual application design is simulated by a real hardware simulator, saving much simulation overhead and giving quicker results for (relatively) rapid prototyping.
 

Business process architectures are an important vehicle to organize business process models of an organization, since they provide an holistic view on the interrelationships between business process models. Current techniques to designing business process architectures derive them directly from a repository of process models, neglecting  processes execution information like event logs. In this paper, we propose an approach for automatically discovering a business process architecture from a set of event logs. The resulting architecture captures behavioral dependencies between two ore more business process models, using information stored in event logs. We extended an existing business process architecture representation to accommodate such behavior. The application of the approach on two real-life event logs shows its effectiveness in discovering the intricate dependencies between business processes, based on data.

Steady progress in machine learning-based predictive models has propelled companies to incorporate these models in systems for online recommendations, diagnostics, forecasts, and cyber-physical control. However, recent surveys show that from 55% to 72% of companies are not being able to successfully deploy these systems. The putative reason is the unmitigated unrealiability of predictive models that are exposed to various types of changes inherent to the more complex and evolving reality of the production environments.

In this talk I will argue that there are essential and accidental sources for these changes, which can only be partially mitigated by existing methods. Hence, there is a need for new methods, in particular, that can systematically discover model invariants across different environments. In other words, that can categorize rules as either structural (essential/causal) or contingent (accidental/spuriously correlational). I will exemplify these methods in the context of my research on "causal models of imperfect fault understanding" and their deployment as software components of a recommender system for bug inspection tasks.

Currently, user interfaces for laser-cutting are designed for expert users. However, as laser cutting becomes widespread, non-expert users (such as high school teachers) will want to use these interfaces too. 
Using a typical laser cutting interface requires advanced knowledge of the machine and the material. This lack of knowledge demands complicated calibration procedures. A non-expert would not be willing to spend time on.
The domain of desktop publishing has solved this same problem with a page description language called PostScript. With this description language and its standardized interpreter, users no longer need to consider setting up a machine dependent parameter.
In this talk, I will show the history of PostScript and how it has changed computer printing. Based on this history, we will discuss current issues with laser cutters and what the description language in the laser cutter will look like.

The rapid progress in machine learning has brought people huge benefits and conveniences. These algorithms have been embedded into almost everything in our daily life, from a cellphone, smartwatch to important infrastructure. However, the widely used smart systems also brought risks and challenges, especially for "smart" safety-critical systems. There seems to be an inherent conflict between smart and safe/reliable. In this talk, I want to introduce the challenges brought by AI components to the safety-critical system(e.g. distributional shift, unsafe exploration), and the potential approaches(e.g. fail-safe mechanism, redundancy) to handle them. 

Various web-based image-editing tools and web-based collaborative tools exist in isolation. Research focusing to bridge the gap between these two domains is sparse. We respond to the above and develop prototype groupware for real-time collaborative editing of raster images in a web browser.  To better understand the requirements, we conduct a preliminary user study and establish communication and synchronization as key elements.  The existing groupware for text documents, presentations, and vector graphics handles the above through well-established techniques. However, those cannot be extended as it is for raster graphics manipulation. To this end, we develop a document model that is maintained by a server and is delivered and synchronized to multiple clients. Our prototypical implementation is based on a scalable client-server architecture: using WebGL for interactive browser-based rendering and WebSocket connections to maintain synchronization. We evaluate our work qualitatively through a post-deployment user study for three different scenarios.

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 anomaly. 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. So, automated anomaly detection and failure analysis reduces the overall dependency on expert knowledge and provide an effective solution to mitigate failure propagation in Microservice Architecture. In this talk, I will present my research idea as compared to existing solution, our methodology, what are those KPI metrics that are being involved in anomaly detection and failure analysis.

We present virtualArms, an interaction technique that allows users to interact with virtual worlds by means of a pair of virtual arms that users control using their shoulders. This allows users’ physical hands to remain in their laps—saving physical space and thus allowing VR interaction to take place in highly confined spaces. We implemented virtualArms as a driver that allows controlling arbitrary applications running on a VR headset. The driver implements a simple kinematic model that maps shoulder movement to human arm motion. When punching, users perform the same whole-body movement they normally would, only suppress moving the arm. As the rest of the user’s body performs the same movement they would perform in the physical world, i.e., a wave traveling up the “kinematic chain”, virtualArms preserve a high degree of realism and immersion. In our user study, participants controlling a boxing game and a rhythm game rated their interacting more immersive and natural when using virtualArms than when using a game pad.