Refine
Year of publication
Document Type
- Preprint (759)
- Article (402)
- Working Paper (119)
- Doctoral Thesis (93)
- Diploma Thesis (47)
- Conference Proceeding (41)
- Book (37)
- Bachelor Thesis (36)
- diplomthesis (28)
- Report (25)
Has Fulltext
- yes (1619)
Is part of the Bibliography
- no (1619)
Keywords
Institute
- Informatik (1619) (remove)
Co-design of a trustworthy AI system in healthcare: deep learning based skin lesion classifier
(2021)
This paper documents how an ethically aligned co-design methodology ensures trustworthiness in the early design phase of an artificial intelligence (AI) system component for healthcare. The system explains decisions made by deep learning networks analyzing images of skin lesions. The co-design of trustworthy AI developed here used a holistic approach rather than a static ethical checklist and required a multidisciplinary team of experts working with the AI designers and their managers. Ethical, legal, and technical issues potentially arising from the future use of the AI system were investigated. This paper is a first report on co-designing in the early design phase. Our results can also serve as guidance for other early-phase AI-similar tool developments.
Point-based geometry representations have become widely used in numerous contexts, ranging from particle-based simulations, over stereo image matching, to depth sensing via light detection and ranging. Our application focus is on the reconstruction of curved line structures in noisy 3D point cloud data. Respective algorithms operating on such point clouds often rely on the notion of a local neighborhood. Regarding the latter, our approach employs multi-scale neighborhoods, for which weighted covariance measures of local points are determined. Curved line structures are reconstructed via vector field tracing, using a bidirectional piecewise streamline integration. We also introduce an automatic selection of optimal starting points via multi-scale geometric measures. The pipeline development and choice of parameters was driven by an extensive, automated initial analysis process on over a million prototype test cases. The behavior of our approach is controlled by several parameters — the majority being set automatically, leaving only three to be controlled by a user. In an extensive, automated final evaluation, we cover over one hundred thousand parameter sets, including 3D test geometries with varying curvature, sharp corners, intersections, data holes, and systematically applied varying types of noise. Further, we analyzed different choices for the point of reference in the co-variance computation; using a weighted mean performed best in most cases. In addition, we compared our method to current, publicly available line reconstruction frameworks. Up to thirty times faster execution times were achieved in some cases, at comparable error measures. Finally, we also demonstrate an exemplary application on four real-world 3D light detection and ranging datasets, extracting power line cables.
High-energy physics experiments aim to deepen our understanding of the fundamental structure of matter and the governing forces. One of the most challenging aspects of the design of new experiments is data management and event selection. The search for increasingly rare and intricate physics events asks for high-statistics measurements and sophisticated event analysis. With progressively complex event signatures, traditional hardware-based trigger systems reach the limits of realizable latency and complexity. The Compressed Baryonic Matter experiment (CBM) employs a novel approach for data readout and event selection to address these challenges. Self-triggered, free-streaming detectors push all data to a central compute cluster, called First-level Event Selector (FLES), for software-based event analysis and selection. While this concept solves many issues present in classical architectures, it also sets new challenges for the design of the detector readout systems and online event selection.
This thesis presents an efficient solution to the data management challenges presented by self-triggered, free-streaming particle detectors. The FLES must receive asynchronously streamed data from a heterogeneous detector setup at rates of up to 1 TB/s. The real-time processing environment implies that all components have to deliver high performance and reliability to record as much valuable data as possible. The thesis introduces a time-based data model to partition the input streams into containers of fixed length in experiment time for efficient data management. These containers provide all necessary metadata to enable generic, detector-subsystem-agnostic data distribution across the entire cluster. An analysis shows that the introduced data overhead is well below 1 % for a wide range of system parameters.
Furthermore, a concept and the implementation of a detector data input interface for the CBM FLES, optimized for resource-efficient data transport, are presented. The central element of the architecture is an FPGA-based PCIe extension card for the FLES entry nodes. The hardware designs developed in the thesis enable interfacing with a diverse set of detector systems. A custom, high-throughput DMA design structures data in a way that enables low-overhead access and efficient software processing. The ability to share the host DMA buffers with other devices, such as an InfiniBand HCA, allows for true zero-copy data distribution between the cluster nodes. The discussed FLES input interface is fully implemented and has already proven its reliability in production operation in various physics experiments.
Ein großer Tag für das Institut für Informatik der Goethe-Universität: Es feiert sein 40-jähriges Jubiläum mit einem Festakt am Freitag, den 13. Oktober, ab 14 Uhr im Casino auf dem Campus Westend. Die Universität hat zusammen mit dem Alumni- und Förderverein hochkarätige Gastredner eingeladen, die ein Bild von der Bedeutung der Informatik und ihrer Forschung zeichnen werden.
We study online secretary problems with returns in combinatorial packing domains with n candidates that arrive sequentially over time in random order. The goal is to determine a feasible packing of candidates of maximum total value. In the first variant, each candidate arrives exactly twice. All 2n arrivals occur in random order. We propose a simple 0.5‐competitive algorithm. For the online bipartite matching problem, we obtain an algorithm with ratio at least 0.5721 − o(1), and an algorithm with ratio at least 0.5459 for all n ≥ 1. We extend all algorithms and ratios to k ≥ 2 arrivals per candidate. In the second variant, there is a pool of undecided candidates. In each round, a random candidate from the pool arrives. Upon arrival a candidate can be either decided (accept/reject) or postponed. We focus on minimizing the expected number of postponements when computing an optimal solution. An expected number of Θ(n log n) is always sufficient. For bipartite matching, we can show a tight bound of O(r log n), where r is the size of the optimum matching. For matroids, we can improve this further to a tight bound of O(r′ log(n/r′)), where r′ is the minimum rank of the matroid and the dual matroid.
In this dissertation the formal abstraction and verification of analog circuit is examined. An approach is introduced that automatically abstracts a transistor level circuit with full Spice accuracy into a hybrid automaton (HA) in various output languages. The generated behavioral model exhibits a significant simulation speed-up compared to the original netlist, while maintaining an acceptable accuracy, and can be therefore used in various verification and validation routines. On top of that, the generated models can be formally verified against their Spice netlists, making the obtained models correct by construction.
The generated abstract models can be extended to enclose modeling as well as technology dependent parameter variations with little over approximations. As these models enclose the various behaviors of the sampled netlists, the obtained models are of significant importance as they can replace several simulations with just a single reachability analysis or symbolic simulation. Moreover, these models can be as well be used in different verification routines as demonstrated in this dissertation.
As the obtained models are described by HAs with linear behaviors in the locations, the abstract models can be as well compositionally linked, allowing thereby the abstraction of complex analog circuits.
Depending on the specified modeling settings, including for example the number of locations of the HA and the description of the system behavior, the accuracy, speedup, and various additional properties of the HA can be influenced. This is examined in detail in this dissertation. The underlying abstraction process is first covered in detail. Several extensions are then handled including the modeling of the HAs with parameter variations. The obtained models are then verified using various verification methodologies. The accuracy and speed-up of the abstraction methodology is finally evaluated on several transistor level circuits ranging from simple operational amplifiers up to a complex circuits.
In this talk we presented a novel technique, based on Deep Learning, to determine the impact parameter of nuclear collisions at the CBM experiment. PointNet based Deep Learning models are trained on UrQMD followed by CBMRoot simulations of Au+Au collisions at 10 AGeV to reconstruct the impact parameter of collisions from raw experimental data such as hits of the particles in the detector planes, tracks reconstructed from the hits or their combinations. The PointNet models can perform fast, accurate, event-by-event impact parameter determination in heavy ion collision experiments. They are shown to outperform a simple model which maps the track multiplicity to the impact parameter. While conventional methods for centrality classification merely provide an expected impact parameter distribution for a given centrality class, the PointNet models predict the impact parameter from 2–14 fm on an event-by-event basis with a mean error of −0.33 to 0.22 fm.
The ongoing digitalization of educational resources and the use of the internet lead to a steady increase of potentially available learning media. However, many of the media which are used for educational purposes have not been designed specifically for teaching and learning. Usually, linguistic criteria of readability and comprehensibility as well as content-related criteria are used independently to assess and compare the quality of educational media. This also holds true for educational media used in economics. This article aims to improve the analysis of textual learning media used in economic education by drawing on threshold concepts. Threshold concepts are key terms in knowledge acquisition within a domain. From a linguistic perspective, however, threshold concepts are instances of specialized vocabularies, exhibiting particular linguistic features. In three kinds of (German) resources, namely in textbooks, in newspapers, and on Wikipedia, we investigate the distributive profiles of 63 threshold concepts identified in economics education (which have been collected from threshold concept research). We looked at the threshold concepts' frequency distribution, their compound distribution, and their network structure within the three kinds of resources. The two main findings of our analysis show that firstly, the three kinds of resources can indeed be distinguished in terms of their threshold concepts' profiles. Secondly, Wikipedia definitely shows stronger associative connections between economic threshold concepts than the other sources. We discuss the findings in relation to adequate media use for teaching and learning—not only in economic education.
Learning to solve graph tasks is one of the key prerequisites of acquiring domain-specific knowledge in most study domains. Analyses of graph understanding often use eye-tracking and focus on analyzing how much time students spend gazing at particular areas of a graph—Areas of Interest (AOIs). To gain a deeper insight into students’ task-solving process, we argue that the gaze shifts between students’ fixations on different AOIs (so-termed transitions) also need to be included in holistic analyses of graph understanding that consider the importance of transitions for the task-solving process. Thus, we introduced Epistemic Network Analysis (ENA) as a novel approach to analyze eye-tracking data of 23 university students who solved eight multiple-choice graph tasks in physics and economics. ENA is a method for quantifying, visualizing, and interpreting network data allowing a weighted analysis of the gaze patterns of both correct and incorrect graph task solvers considering the interrelations between fixations and transitions. After an analysis of the differences in the number of fixations and the number of single transitions between correct and incorrect solvers, we conducted an ENA for each task. We demonstrate that an isolated analysis of fixations and transitions provides only a limited insight into graph solving behavior. In contrast, ENA identifies differences between the gaze patterns of students who solved the graph tasks correctly and incorrectly across the multiple graph tasks. For instance, incorrect solvers shifted their gaze from the graph to the x-axis and from the question to the graph comparatively more often than correct solvers. The results indicate that incorrect solvers often have problems transferring textual information into graphical information and rely more on partly irrelevant parts of a graph. Finally, we discuss how the findings can be used to design experimental studies and for innovative instructional procedures in higher education
Volatility clustering and fat tails are prominently observed in financial markets. Here, we analyze the underlying mechanisms of three agent-based models explaining these stylized facts in terms of market instabilities and compare them on empirical grounds. To this end, we first develop a general framework for detecting tail events in stock markets. In particular, we introduce Hawkes processes to automatically identify and date onsets of market turmoils which result in increased volatility. Second, we introduce three different indicators to predict those onsets. Each of the three indicators is derived from and tailored to one of the models, namely quantifying information content, critical slowing down or market risk perception. Finally, we apply our indicators to simulated and real market data. We find that all indicators reliably predict market events on simulated data and clearly distinguish the different models. In contrast, a systematic comparison on the stocks of the Forbes 500 companies shows a markedly lower performance. Overall, predicting the onset of market turmoils appears difficult, yet, over very short time horizons high or rising volatility exhibits some predictive power.
Monitoring is an indispensable tool for the operation of any large installation of grid or cluster computing, be it high energy physics or elsewhere. Usually, monitoring is configured to collect a small amount of data, just enough to enable detection of abnormal conditions. Once detected, the abnormal condition is handled by gathering all information from the affected components. This data is processed by querying it in a manner similar to a database.
This contribution shows how the metaphor of a debugger (for software applications) can be transferred to a compute cluster. The concepts of variables, assertions and breakpoints that are used in debugging can be applied to monitoring by defining variables as the quantities recorded by monitoring and breakpoints as invariants formulated via these variables. It is found that embedding fragments of a data extracting and reporting tool such as the UNIX tool awk facilitates concise notations for commonly used variables since tools like awk are designed to process large event streams (in textual representations) with bounded memory. A functional notation similar to both the pipe notation used in the UNIX shell and the point-free style used in functional programming simplify the combination of variables that commonly occur when formulating breakpoints.
We test the hypothesis that the extent to which one obtains information on a given topic through Wikipedia depends on the language in which it is consulted. Controlling the size factor, we investigate this hypothesis for a number of 25 subject areas. Since Wikipedia is a central part of the web-based information landscape, this indicates a language-related, linguistic bias. The article therefore deals with the question of whether Wikipedia exhibits this kind of linguistic relativity or not. From the perspective of educational science, the article develops a computational model of the information landscape from which multiple texts are drawn as typical input of web-based reading. For this purpose, it develops a hybrid model of intra- and intertextual similarity of different parts of the information landscape and tests this model on the example of 35 languages and corresponding Wikipedias. In the way it measures the similarities of hypertexts, the article goes beyond existing approaches by examining their structural and semantic aspects intra- and intertextually. In this way it builds a bridge between reading research, educational science, Wikipedia research and computational linguistics.
A new method of event characterization based on Deep Learning is presented. The PointNet models can be used for fast, online event-by-event impact parameter determination at the CBM experiment. For this study, UrQMD and the CBM detector simulation are used to generate Au+Au collision events at 10 AGeV which are then used to train and evaluate PointNet based architectures. The models can be trained on features like the hit position of particles in the CBM detector planes, tracks reconstructed from the hits or combinations thereof. The Deep Learning models reconstruct impact parameters from 2-14 fm with a mean error varying from -0.33 to 0.22 fm. For impact parameters in the range of 5-14 fm, a model which uses the combination of hit and track information of particles has a relative precision of 4-9% and a mean error of -0.33 to 0.13 fm. In the same range of impact parameters, a model with only track information has a relative precision of 4-10% and a mean error of -0.18 to 0.22 fm. This new method of event-classification is shown to be more accurate and less model dependent than conventional methods and can utilize the performance boost of modern GPU processor units.
The impact of columnar file formats on SQL‐on‐hadoop engine performance: a study on ORC and Parquet
(2019)
Columnar file formats provide an efficient way to store data to be queried by SQL‐on‐Hadoop engines. Related works consider the performance of processing engine and file format together, which makes it impossible to predict their individual impact. In this work, we propose an alternative approach: by executing each file format on the same processing engine, we compare the different file formats as well as their different parameter settings. We apply our strategy to two processing engines, Hive and SparkSQL, and evaluate the performance of two columnar file formats, ORC and Parquet. We use BigBench (TPCx‐BB), a standardized application‐level benchmark for Big Data scenarios. Our experiments confirm that the file format selection and its configuration significantly affect the overall performance. We show that ORC generally performs better on Hive, whereas Parquet achieves best performance with SparkSQL. Using ZLIB compression brings up to 60.2% improvement with ORC, while Parquet achieves up to 7% improvement with Snappy. Exceptions are the queries involving text processing, which do not benefit from using any compression.
Es sollte eine Simulationsumgebung mit einem Straßennetz und eine KDNA, die Autos auf diesem Straßennetz kontrolliert, implementiert werden. Für die Simulation wurde eine einfache graphische Darstellung entwickelt auf der eine variable Anzahl Autos auf einem vorprogrammierten Straßennetz fahren. Eine KDNA steuert diese Autos über Kontrolle von Gas-, Bremse- und Steuerradpositionen, wobei Geschwindigkeits- und Richtungskontrolle unabhängig stattfinden. Bei der Analyse der KDNA für mehrere Autos traten Leistungsprobleme auf, deren Quelle genauer untersucht wurde. Die Last wurde primär durch die Kommunikation zur Verwaltung der KDNA-Tasks im AHS erzeugt.
Das Ziel dieser Arbeit ist, einen Text automatisch darauf zu untersuchen, ob er Gebäude beschreibt, und diese gegebenenfalls zu visualisieren. Zu diesem Zweck wurde ein Prototyp entwickelt, der mithilfe von NLP-Software auf Basis einer UIMA-Pipeline einen Text auf Gebäudedaten untersucht und diese anschließend als 3D-Modelle auf einer Karte visualisiert. Um die Güte des Projekts zu bestimmen wurde eine Evaluation durchgeführt, in der die Aufgabe darin bestand, Paragraphen ihren zugehörigen 3D-Modellen zuzuordnen. Die Ergebnisse wiesen eine Erkennungsrate von 88.67\% auf. Jedoch wurden auch Schwächen im Standardisierungsverfahren der Parameter und in der einseitigen Art zu Visualisieren aufgezeigt. Zum Schluss wird vorgestellt, wie diese Schwachstellen mithilfe eines ontologischen Modells behoben werden können und wie mit dem Projekt weiterverfahren werden kann.
In diesem Beitrag untersuchen wir Entwicklungstendenzen von Infrastrukturen in den Digitalen Geisteswissenschaften. Wir argumentieren, dass infolge (1) der Verfügbarkeit von immer mehr Daten über sozial-semiotische Netzwerke, (2) der Methodeninflation in geisteswissenschaftlichen Disziplinen, (3) der zunehmend hybriden Arbeitsteilung zwischen Mensch und Maschine und (4) der explosionsartigen Vermehrung künstlicher Texte ein erheblicher Anpassungsdruck auf die Weiterentwicklung solcher Infrastrukturen entstanden ist. In diesem Zusammenhang beschreiben wir drei Informationssysteme, die sich unter anderem durch die Interaktionsmöglichkeiten unterscheiden, die sie ihren Nutzern bieten, um solchen Herausforderungen zu begegnen. Dabei skizzieren wir mit VienNA eine neuartige Architektur solcher Systeme, welche aufgrund ihrer Flexibilität die Möglichkeit bieten könnte, letztere Herausforderungen zu bewältigen.