In this introductory talk Eva Krebs will present some of the previous projects she worked on. There will be live demonstrations of the respective tools and programs. Topics will include font creation and rendering, live tools for simulations, polyglot programming for Jupyter-style notebooks and automated creativity assessments. The talk will also touch upon online user studies and study participant acquisition.

At the current state of the art, practical FPGA designs combine hardware descriptions on several abstraction levels as well as predefined IP-blocks, to specify the application logic with an acceptable degree of convenience and portability. In consequence, an end-to-end simulation of such a design involves the accurate modeling of all these design parts, even if the developer would have considered them as part of the system environment. Cycle accurate hardware models for complex environment resources such as DRAM, HBM or network ports can add an overwhelming overhead with respect to the simulation time of the pure 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 demonstrate an alternative approach implemented in the DFACCTO-SIM simulation framework. Here hardware components outside the particular application focus are represented as more abstract behavioural models in software. I will go into architectural details and the intended workflow of DFACCTO-SIM, as well as its application to evaluate a dynamic data stream switch component, which is central to an autonomously scheduled multi-tenant operator network.

Process discovery is an important area in the field of process mining, where most discovery algorithms focus on process control-flow, giving little attention to the data-flow perspective. 

As a result, the discovered process models lack information about data dependencies, and process experts need to manually enrich the discovered process models accordingly. This requires deep domain knowledge, is not scalable, and is error-prone. To overcome this limitation, this paper proposes an approach that aims to discover the data objects and their behavior by investigating how event attributes are manipulated during process execution. The resulting data objects are used to enhance the discovered process model. The feasibility of the proposed approach is evaluated with two real-life event logs: Road Traffic Fine Management and Hospital Billing.

Modern mobile devices, equipped with high-end cameras and GPU-support, have the capabilities to perform real-time video processing tasks such as video enhancement or stylization. With respect to this, optical flow estimation can be used to increase the temporal coherency of the output, and thus its overall quality. There are a number of high-accuracy learning-based estimation methods, however, their application to mobile-devices is limited by computational resources. To this end, we apply neural network compression techniques to an existing flow estimation network thereby significantly improving its run-time performance on both mobile and desktop systems

Recent advancements in the field of machine learning (ML), especially deep learn- ing have led to high interest in the field in the medical domain. Many machine learning models have been and are being developed each year to address various health-related questions. However, to be able to bring these advancements to the end-user, be it patient or healthcare professional, outside of clinical trial settings, models need to be approved as medical products. This approval process is regulated by European as well as local laws, it is a very expensive and slow-moving formal process, which makes it unfit for the fast- moving field of machine learning and artificial intelligence. Sofar regulators in the EU do not explicitly address the use of machine learning in medical products, how- ever, approval of medical products containing AI is already possible, by applying the existing standards used to approve software. The field of Machine Learning Operations (MLOPs) tries to apply the principles of DevOps to machine learning. It combines the fields of machine learning, data engineering, and DevOps. MLOPs aims to make the deployment of ML systems more reliable and efficient. These formalization of best practices for the development and testing of machine learning software could help to speed up the process whilst being an important factor to fullfill the strict regulatory requirements required for medical products.

Process mining bridges the gap between process management and data science by discovering process models using event logs derived from real-world data. Besides mandatory event attributes, additional attributes can be part of an event representing domain data, such as human resources and costs. Data-enhanced process models provide a visualization of domain data associated to process activities directly in the process model, allowing to monitor the actual values of domain data in the form of event attribute aggregations. However, event logs can have so many attributes that it is difficult to decide, which one is of interest to observe throughout the process.

In this talk, I will introduce three mechanisms to support domain data selection, allowing process analysts and domain experts to progressively reach their information of interest. The proposed technique was applied on the MIMIC-IV real-world data set on hospitalizations in the US.

Learning causal relationships from observational data is insightful for researchers in multiple domains. For example, in genetic research, gene regulatory networks can be derived from gene expression data using causal structure learning. Or, in manufacturing, the learned causal relationships provide decision support for effective troubleshooting.
Yet, algorithms for learning causal structures have high computational complexity and consequently are currently limited by their long runtimes. In this talk, I will present work on the efficient execution of causal structure learning algorithms in heterogeneous computing systems by leveraging the parallel computing capabilities of GPUs to constitute a solution to the long runtime. In this context, the design of parallel causal structure learning algorithms, e.g., following Foster’s methodology, requires considering GPU hardware characteristics such as the size of shared memory for local communication within thread blocks and further constraints stemming from different data distributions.

Continuous digital planning of control command and signaling with the DB format PlanPro is the next generation of planning of railway networks. This approach finally wants to get rid of paper-based planning. Such digital planning contains a huge number of details about the planned railway network. In our research, we want to use this planning as a basis for the automated commissioning of new railway infrastructure devices.

Our approach is to analyze the planning, create a virtual simulated copy of the railway network. In this simulated network, each infrastructure element can either be a virtualized object controller or a real hardware system. This creates a flexible, scalable but realistic test environment.

To create a huge number of test cases, we want to generate the operating procedures of such a network and virtually play all these scenarios as test cases.

This presentation shows a suggestion for the system architecture and the current status of the simulated virtual copy of the railway network.

My specific objective is to create a technological basis that allows designers and engineers of laser cut 3D models to build on each other’s work, as I see this as being instrumental in allowing this nascent field to increase in model complexity and adoption—a progression that is currently held back by the use of exchange formats that disregard mechanical differences between machines and therefore overlook implications with respect to how well parts mechanically fit together (aka engineering fit). I thus explore better ways for representing models.

I started by preserving the 2D format currently in use, by writing software tools that replace machine-specific features (press-fit and loose-fit) with more forgiving mechanical elements, i.e., specific types of springs (full paper ACM UIST’19) and a novel type of bearing I engineered (full paper ACM UIST’20). While these allow users to reproduce 3D models across machines, the field needs access to parametric modifications in 3D to build on work of others. I created elements of a software system that allows modeling laser-cut models in 3D (volume-based construction, full paper CHI’19 and the integration with plate-based construction in submission to CHI’22). Users then export the 3D models to their specific machine, which solves the portability question. It does raise a question of what to do with the vast amount of existing 2D cutting plans. To handle this legacy, I have written a series of software tools that allow users to convert existing 2D cutting plans into 3D models, by reconstructing the 3D nature of the model using a set of interactive manual tools (full paper CHI’20), I then progressed to an automatic approach based on efficient graph algorithms (full paper UIST’21), which I will revisit yet again using computer vision to cover most models shared online.

I have integrated these tools into the above-mentioned system (kyub, 140,000 lines of code). To test each of my technologies with hundreds of models from this model repository, and by running workshops with school classes to see that my software lets users progress to more complex models such as furniture and guitars. This system integration also allows me to push forward to actual use. I believe that by simplifying sharing and re-use and the resulting increase in model complexity, this line of work and the resulting system will ultimately help personal fabrication scale past the maker phenomenon it currently is and towards a mainstream phenomenon—the same way that other fields, such as print (postscript) and ultimately computing itself (portable programming languages, etc.) to reach mass adoption.

Block-based Editors are typically confined to the space of domain-specific languages or, more generally, languages with a small design. Yet, even in more complex, general-purpose programming languages, enjoying the interactivity and affordances of block-based editors may be desirable.

We have previously demonstrated designs for block-based editors both for domain-specific languages, as well as Smalltalk, a general-purpose language with a minimal language design. In this presentation, we will look at our first steps to supporting a broader set of programming languages, including C, Java, or Python. We will discuss some details of the implementation, challenges we encountered, and give an outlook on what we hope this exploration will enable.

Commands to interactive application programs tend to consist of multiple elements, such as what to do (aka the "verb"), what to apply it to (aka the "noun"), as well as typically a number of additional arguments, such as options.

In this talk, I will look at a number of historical application examples from the domain of 3D editing. I find that different programs take different approaches to ordering these elements. I try to understand how these orders affect efficiency, accuracy, and learnability.

Laser cutting is a fast fabrication technique. 3D modeling software like flatfab enables users to make 3D laser cut models quickly. A software system like Roadkill makes assembling fabricated models more straightforward and faster. However, there is another time-consuming process that has not been discussed yet. It’s called calibration. To achieve the desired result, users need to adjust laser energy and the laser-cut model based on kerf, the width of material removed by a laser cutter. My research challenge is automating this calibration process. In this talk, we focus on kerf. I will talk about the mechanical device to measure kerf, which I‘m developing now.

In this talk, I will introduce a framework that can improve AI-based system safety. The main structure of the framework is based on the typical model@runtime system and MAPE-K loop. It consists of three layers, which perform (1)model management, (2)analyzing, monitoring, planning, reasoning and executing, and (3)goal, safety, and common-sense management, respectively. In the framework, a machine common sense component is adopted to enhance the reasoning part to eliminate mistakes often made by machines but not humans. For the planning part of the structure, an ensemble learning mechanism is introduced to resist the rare but severe mistake made by a single AI component.

Nowadays, AI techniques are gaining more and more popularity because of their performance. With them, we can model complex phenomena receiving high accuracy of results. However, most AI models are considered black-boxes making their decisions neither understandable nor interpretable by humans. This limitation is especially significant when these decisions contradict end-users ' expectations based on state-of-the-art knowledge. It results in a lack of trust in the model. 

In this presentation, I will focus on describing Explainable AI methods. With them, we can understand the behavior of models and explain why a defined decision was made. I will try to answer the question of what we, as end-users, can gain by deploying it. I will also present the main Explainable AI techniques, their limitation, and present the results of my current research. Further, I will shortly introduce the uncertainty estimation of the AI model, which tells us with the given probability how far from the mean prediction score the real prediction score is expected to be. This estimation allows practitioners to decide whether they can trust the decision of the model and whether there is the need for an additional human decision. In the end, I will present the direction of my further research.

Virtual reality hardware has become inexpensive and thus accessible to the broad public. As a result, the constraining resource now tends to be physical space. In my research, I am tackling this limitation, i.e., I explore how to modify virtual experiences so that they an be fit into the small spaces users have at their disposal. One of my prototypes explores to what extent we can complement users' physical space with additional virtual space outside the tracking volume. While users can never walk there, such virtual space still allows for a wide range of interactions, such as users throwing objects into the virtual space, interaction with objects in virtual space using virtual tools, interacting with virtual characters in virtual space, etc. I investigate how this “off-space interaction” effects user’s perception of space, and what we can do to maximize the utility of virtual space to users. In this talk, I present and discuss the first prototype of psychometric experiment of how to measure the value of off-space interaction.