In this talk I will explain the concept of continuous
integration (CI), its benefits for software development processes, as
well as practical use-cases. The term CI was originally used to refer to
the practice of integrating small changes often (daily) into a shared
code-base. Today it is often used to refer to CI-pipelines, which can be
used to automate the testing, building, and deployment of software,
which was made necessary by the practice of integrating at such small
steps. Continuous integration is meant to reduce integration problems
and thus speed up the overall development process. We propose to apply
the process of CI and automation (CI-pipelines) to the writing phase of
technical reports with many contributors. This should streamline the
process, and reduce and distribute the manual efforts, needed to compile
the individual contributions to one big report.

Over the past few years, there has been a lot of change in the development & deployment of application since the ’80s till now. From monolithic to microservices, development process, application architecture, packaging and infrastructure has been evolved. We can trace the evolution of how organizations have developed, deployed, and managed applications over distinct phases. Due to this evolution, design, monitoring and testing in the industry also has greater impact on it. In this talk, I will present practitioner & researcher perspective of how microservices systems are designed, monitored, and tested based on survey findings. I will discuss existing approaches and methods in this area, and also discuss why more research work is needed in that area to address the open challenges in microservices performance maintenance through dedicated solution. In the end, I will show the comparison between my research objectives and existing state-of-the-art.

3D point clouds are used for a wide range of geospatial applications, as they can represent almost any physical object with any precision. There has been a growing interest in using Deep Learning (DL) for 3D point cloud analysis, with a trend toward taking the raw point cloud as an input. However, most of the current works in the state-of-the-art (SOTA) are limited to small point clouds containing a few thousand elements. Most of the research regarding DL-based large-scale 3D point cloud analysis employs a data construction and training pipeline based on subsampling, downsampling, or a combination of the two. During this process, the resulting training and testing samples are treated as independent, disregarding that they come from a single, overarching scene. In the initial stages of my Ph.D. research work, I posit that some of the prevailing segmentation errors in the SOTA are due to a trade-off between considering the level of detail and preserving scene context inherent to most of the current sampling and training approaches. I then introduce research questions oriented to constructing a DL framework with a data preparation and model training pipeline better suited for large-scale 3D point clouds. Next, I present preliminary results on the proposed research questions. Finally, I outline the proposed future work in this direction and briefly present research works that utilize 3D point clouds in different application scenarios.

In this talk, I will showcase related work which is relevant to my ongoing research project, ToolsetOptimizer. I will show papers from the field of user interface optimization and papers about reverse engineering of vector graphics.

Summary of my research project: ToolsetOptimizer automatically composes special-purpose drawing programs that optimize usability for a given use case and user group. In order to accomplish this, ToolsetOptimizer needs to be provided with (a) a larger set of tools from which it selects, (b) the usability needs of the target user group in the form of weights of a cost function, and (c) a training set of vector drawings that users are supposed to produce. ToolsetOptimizer does so by performing the following algorithm: (1) It starts with an empty toolset, (2) iteratively creates bigger toolsets by adding tools, beam searching the space of possible toolsets, (3) filters the toolsets by pareto optimality, ranks them and displays them in an interactive view to the user.

Sumit Shekhar > Interactive 3D Photo Stylization on Mobile Devices

For decades, Image-based Artistic Rendering (IB-AR) has been successfully employed to simulate the appeal of traditional artistic styles for enhanced visual communication.  Recently, 3D photography has emerged as a new medium that provides an immersive dimension compared to 2D photos. The possibility to change the viewpoint, depicting parallax effects is mesmerizing.
In this talk, I will present an interactive mobile app that combines the vividness of IB-AR with immersive 3D photos. The mobile app implements an end-to-end pipeline for the acquisition of data, generation of a 3D photo in the form of a Layered-Depth-Image, and its artistic rendering. The complete pipeline runs at interactive frame rates on an iPad and the final output is obtained as a compact video, which can easily be shared. Thus, it serves as a unique interactive tool for digital artists interested in creating immersive artistic content.

Jotaro Shigeyama > VR at 1:1 scale in small physical space

Virtual Reality (VR) leads to the highest level of immersion if presented using a 1:1 mapping of virtual space to physical space—also known as real-walking. The advent of inexpensive consumer VR headsets, all capable of running real-tracking VR, has brought VR to the home. Surprisingly, however, only a few of these experiences feature real-walking, but instead feature a less immersive space-saving technique known as instant teleportation. The reason appears to be that home-users simply do not have the physical space available that would allow for a meaningful use of real-walking. The main hurdle to the adoption of real-walking VR in the home, thus appears to be the sheer cost of physical space.

In this talk, I present my Ph.D topic on how to run 1:1 mapped VR in physical spaces, with the sizes available to home users.
[world in miniature] We present DualPanto, a device that allows blind VR users to experience the virtual world from a 1:1 mapped miniature world perspective—by leveraging haptics. 
[arms only] We then advance to VR leveraging the user’s body, albeit just the user’s arms. By extending the movement of the user’s torso to the virtual arms, virtualArms allows users to experience arm-oriented VR applications, such as boxing in highly confined spaces, such as during train or airplane rides.
[whole body] Finally, we extend our approach to whole-body VR. We build on impossible spaces, a technique that allows squishing VR works into tiny tracking volumes. We focus on 3m2 (2m x 1.5m), which is what 75% of home users have access to. We extend impossible spaces with (a) windows that allow users to look outside the tracking volume, (b) interaction techniques that allow users to reach outside the tracking volume, and (c) hallways and staircases that allow users to traverse longer and multi-story spaces, despite the small tracking volume. 

A key difference between digital fabrication and the processing of purely virtual data is that fabrication process undergoes multiple representations: 3D virtual, 2D virtual (aka nesting), 2D physical (cutting) and 3D physical (assembly). Unfortunately, the resulting sequential process lacks many of the qualities purely virtual processes offer: in particular there is no real "what you see is what you get" that would allow users to anticipate the consequences of their actions and there is no real "forgiveness" that would allow users to take their actions back. This makes the digital fabrication process slow and, at times, frustrating.

However, a substantial body of recent research has tried to remedy this, by breaking away from the traditional edit-fabricate-assemble-test process, towards blurred boundaries between the steps by allowing users to preview or review, resulting in a new breed of fluid and vastly sped-up end-to-end fabrication systems that feel more like their virtual counterparts. In this talk I will discuss the implications of these systems and show some of my previous and current work in this domain.

Processes in healthcare are complex and data-intensive. Process mining uses data recorded during process execution to obtain an understanding of the actual execution of a process. Due to the complexity of healthcare processes, it is useful to consider and analyse the process execution of certain cohorts, such as old and young patients, separately. While such analysis is facilitated by process variant analysis techniques, existing approaches for process variant analysis only consider a comparison based on the control flow and performance perspectives. Given the large amount of event data attributes available in healthcare settings, we propose the first data-based process variant analysis approach. Our approach allows comparing process variants based on differences in event data attributes by building on statistical tests.

We applied our approach on the MIMIC-IV real-world data set on hospitalizations in the US, where we demonstrate that the approach is feasible and can actually provide relevant medical insights.

The integration of new enterprise computing practices and the emergence of sophisticated adversarial tactics has severely threatened the established security model of perimeter defense, in which controls placed at  the organization's or segment boundary shall fend off or contain attacks. While zero trust networking promises angles to cope with new workplace habits and advanced adversaries, it requires a comprehensive collection and evaluation of sensor data on network devices. While from the organization's and end user's perspective it is desirable to integrate personal devices as part of a bring-your-own-device policy, collecting the necessary sensor data on user-owned devices raises severe privacy concerns for many users.

In this paper, we investigate how zero trust may be applied in BYOD environment without requiring users to install monitoring software, but indirectly infer the device's status from sensor information passively collected in the network. We find that this collected information can provide high visibility, and can thus balance both the requirement for security and user acceptance.

Shohei Katakura:

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 kerf calibration. To achieve the desired result, users need to adjust the laser-cut model based on kerf, the width of material removed by a laser cutter. 

As you may know, we developed a mechanical device called Kerfmeter to measure kerf for a given material and laser energy. This allows users to fabricate the model with proper fit in any human calibration process. However, we found measuring kerf and compensating models is not enough to produce a proper fit. This is because laser cutters have limited positioning accuracy, precision, and laser power accuracy over the cut location. This can lead to wrong physical dimensions of laser cut joints, thus the fit of a produced model will not be proper. In this talk, I explain this problem and our work-in-progress solution.

Adrian Jobst:

Dynamic transaction shows many aspects of Big data, which causes, among other factors, a high level of technical complexity in their processing, analysis and visualization. In general, it does not have a natural geometry embedding, shape or gestalt. Visual Analytics provides a general process model for mining knowledge from data by combining visualization techniques and data analysis. 

In this talk, we  present two visual analytics approaches for mining order book data, as an example for dynamic transaction data. Our first approach allows the comparison of a discrete sequence of order books across several exchanges or time stamps for monitoring market dynamics. Our second approach combines probabilistic topic models for text data to correlate financial time series with discrete events. Besides first results, we discuss challenges and future work.

In this session I will first give a short presentation on Babylonian Programming and the next steps in its research. After that, I will talk about conferences including related activities such as student volunteering. The latter talk will include interactive discussions.

Babylonian Programming aims to provide dynamic feedback directly in the code itself through live examples. This enables faster exploration with less context switches and can also assist future developers in understanding a system more quickly and easily. It can also enhance the design process by providing experts with concrete artifacts that may act as boundary objects.

However, Babylonian Programming currently requires domain experts to create examples. In order to automatically provide examples for a given system, techniques from Example Mining as well as example-driven program inference can be used. Additionally, examples respond to changes in a program’s outcome down to primitive user interactions (liveness). We plan to extend Babylonian Programming with features that draw attention to changes, explain how behavior changed in response to a code change, and allow programmers to express their intent to preserve some behavior (similar to tests) while re-working other behavior.

Due to the temporally linear nature of videos, it’s often inefficient to search for and locate specific parts of interest within videos, e.g., moments when certain objects, persons, or events appear. This problem becomes more severe as videos become longer and fine-grained  actions occur within a short period of time. 

There is a fundamental gap between the ability to formulate meaningful queries for text (e.g. natural language queries) and the ability to retrieve the parts of a video that correspond to such queries for videos. Filling this gap represents a key component for a video summarization system. Recently, unified image-to-language models have been successfully used for approaching downstream tasks such as image labeling, text-to-image retrieval and generation, visual question answering, etc. 

In this talk, we explore the application of one such unified vision-language model, i.e.,  (BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation), for the task of Video Moment Retrieval (VMR), which involves finding moments from a single video described by a given natural language query. We present a conceptually simple approach that can be performed in parallel, discuss its computational aspects, present a demo and address further steps for improvement.

Weronika Wrazen: Towards trustworthy and transparent AI: Explainability

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 shortly remain what is Explainable AI and present my current research on that topic. I will also run a discussion on how to increase trust in AI models and how to make them more understandable to humans.

Tom Beckmann: Generating Structured Editors from Grammars

In this talk, we will be taking a brief look at the state of the art and the history of structured editing. We will then derive possible reasons why structured editing has so far not reached the mainstream. I will then describe our most recent work around Sandblocks, a prototypical structured editor designed to address these potential shortcomings, by generating the editors based on grammars. We will dive into the grammars of some of your favorite programming languages to understand why some languages pose more challenges when it comes to structured editing via our approach then others.

The Internet of Things (IoT) stands for cheap, massively deployed devices, that are introducing pervasive sensing as well as remote-control of our environment. Mass-produced and with little profit margin to the manufacturer, these devices often make little if any provision for security due to the lack of incentives. Shipped with insecure default configurations left without security updates and directly exposed to the Internet, they are prone to be hacked. This comes with severe implications, because despite their low cost, these devices are still relatively powerful. For instance, an ARM Cortex-A53 processor, a typical system-on-chip, retails at only few dollars, but features four cores and provides enough performance to run a full standard operating system and several applications concurrently. This means that once compromised, these devices can provide a versatile platform for adversaries to launch further attacks.

Despite the significance and scale of this malware threat, it is surprising that these Android worms have received almost no attention to date. While several studies extensively analyze the Mirai IoT botnet, no academic work has systematically studied the ADB malware. This paper addresses this void and provides a systematic and comprehensive study of the malware and the resulting ecosystem. In a 4 month study, we tracked and quantified infections, investigated actors and attack vectors and experimentated with containing the worm.