Search

A key question in psycholinguistic research on second language (L2) sentence processing is how L2 users derive sentence structure and meaning from a linear string of words encountered during on-line (real-time) comprehension. Most L2 sentence processing studies have focused on late (adult) L2 learners to examine learners who began learning an L2 after childhood process sentences in the same way as native speakers do (for review, Hopp, 2021; Roberts, 2018). While it is well established that adult native speakers are able to draw upon multiple linguistic sources of information (e.g., lexical, semantic, (morpho)syntactic, and discourse-contextual) and generate predictions about the following input based on their integration, it remains unclear to what extent L2 learners, like native speakers, integrate various types of information and make predictions during sentence comprehension. Against this background, the present thesis aims to draw a detailed picture of the L2 sentence processing system by unravelling the integrative and predictive nature of real-time L2 sentence comprehension as well as the potential role of L1 background in L2 processing. This thesis focusses particularly on L2 processing of verb-specific lexical information during listening and reading comprehension. The reason why this thesis addresses the processing of verb-based information is that verbs play a central role in building sentence structure and interpretation and thus provide a fertile ground for exploring the integrative and predictive processes of real-time sentence comprehension. This thesis reports three studies: Study 1 is a visual world eye-tracking study which probed the use of verb information for predicting upcoming sentence structure. Study 2 is an eye-tracking reading study which investigated the interaction between verb-based and semantic information. In Study 3, a self-paced reading task was conducted to address the interaction of verb-based and discourse-level information sources. All these studies tested L1-German– and L1-Turkish–L2-English learners to examine the extent to which differences in L1 background modulate the L2 processing patterns. Besides the two groups of L2 learners, these studies also tested native speakers of English as the control group, which allows gauging whether and how L2 learners differ from L1 speakers in their use of verb-based lexical information during sentence processing. The results from Studies 1–3 indicated that both L1 German and L1 Turkish learners recruited verb-specific information for prediction just as native speakers do, although L1 effects were observed in the degree and timing of their sensitivity to verb information. The results, however, also showed that L1 German and L1 Turkish learners differed from native speakers in their ability to integrate verb-based lexical information with semantic and discourse-level information. These empirical findings documented that L2 learners may not perform in a native-like manner under increased processing demands, for example, when they are overtaxed by the cognitive demands of integrating multiple sources of information or when L1 effects attenuate the efficiency of processing verb information in the L2. In the context of current approaches to L2 sentence processing, I argue that the observed L1/L2 processing differences do not point to fundamental differences in the architecture and processes underlying sentence comprehension, but are due to limitations in processing resources and efficiency (Hopp, 2010, 2018; Sorace, 2011)., Eine Schlüsselfrage in der psycholinguistischen Forschung zur Verarbeitung sprachlicher Strukturen in der Zweitsprache (L2) ist, wie L2-Benutzer die Satzstruktur und -bedeutung aus einer linearen Wortfolge beim Online (Echtzeit)-Verständnis ableiten. Die meisten L2-Satzverarbeitungsstudien haben sich auf späte (erwachsene) L2 Lerner konzentriert, um zu untersuchen, ob Lernende, die nach der Kindheit mit dem Lernen einer L2 begonnen haben, Sätze auf die gleiche Weise verarbeiten wie Muttersprachler (zur Übersicht siehe Hopp, 2021; Roberts, 2018). Während es allgemein anerkannt ist, dass erwachsene Muttersprachler in der Lage sind, auf mehrere sprachliche (z.B. auf lexikalische, semantische, (morpho)syntaktische und diskurs-kontextuelle) Informationsquellen zurückzugreifen und auf Grundlage ihrer Integration Prädiktionen über den folgenden Input zu generieren, bleibt unklar, inwieweit L2-Lerner wie Muttersprachler verschiedene Typen sprachlicher Informationen integrieren und Prädiktionen während des Satzverständnisses bilden. Vor diesem Hintergrund zielt die vorliegende Dissertation darauf ab, ein detailliertes Bild des L2-Satzverarbeitungssystems zu zeichnen, indem die integrative und prädiktive Natur des Echtzeit-L2-Satzverständnisses sowie die potenzielle Rolle des L1 Hintergrunds in der L2 Verarbeitung betrachtet werden. Diese Dissertation fokussiert insbesondere auf die L2-Verarbeitung von verbspezifischen lexikalischen Informationen beim Hör- und Leseverstehen. Der Grund, warum sich diese Dissertation mit der Verarbeitung verbbasierter Informationen beschäftigt, liegt darin, dass Verben eine zentrale Rolle beim Aufbau und der Interpretation von Sätzen spielen und somit einen fruchtbaren Boden für die Erforschung integrativer und prädiktiver Prozesse des Echtzeit-Satzverständnisses bieten. In dieser Dissertation werden die folgenden drei Studien durchgeführt: Studie 1 ist eine Visual-World-Eyetracking-Studie, die die Nutzung von Verbinformationen zur Prädiktion des bevorstehenden Satzbaus untersuchte. Studie 2 ist eine Eyetracking-Lesestudie, die die Interaktion zwischen verbbasierten und semantischen Informationen erforschte. In Studie 3 wurde eine selbstgesteuerte Leseaufgabe durchgeführt, um die Interaktion von verbbasierten und diskursbasierten Informationsquellen zu untersuchen. All diese Studien testen L1-Deutsch- und L1-Türkisch-L2-Englisch-Lernende, um zu untersuchen, inwieweit Unterschiede im L1 Hintergrund die L2 Verarbeitungsmuster modulieren. Neben den beiden Gruppen von L2 Lernenden wurden in diesen Studien auch englische Muttersprachler als Kontrollgruppe getestet, wodurch beurteilt werden kann, ob und wie sich L2 Lernende von L1-Sprechern in der Nutzung verbbasierter lexikalischer Informationen bei der Satzverarbeitung unterscheiden. Die Ergebnisse aus den Studien 1–3 zeigten, dass sowohl L1-deutschsprachige als auch L1-türkischsprachige Lernende verbspezifische Informationen ebenso wie Muttersprachler zur Prädiktion rekrutierten, obwohl L1 Effekte den Grad und Zeitpunkt der Nutzung von Verbinformationen beeinflussten. Die Ergebnisse zeigten jedoch auch, dass sich L1-deutschsprachige und L1-türkischsprachige Lernende in ihrer Fähigkeit, verbbasierte lexikalische Informationen mit semantischen und diskursiven Informationen zu integrieren, von Muttersprachlern unterschieden. Diese empirischen Ergebnisse belegen, dass sich L2 Lernende unter den erhöhten Verarbeitungsanforderungen in einer L2 in der Satzverarbeitung von Muttersprachlern unterscheiden, beispielsweise wenn sie durch die kognitiven Anforderungen der Integration mehrerer Informationsquellen überfordert sind oder wenn L1 Effekte die Effizienz der Verarbeitung von Verbinformationen in der L2 dämpfen. Im Kontext aktueller Modelle zur L2 Satzverarbeitung argumentiere ich, dass die beobachteten L1/L2-Verarbeitungsunterschiede nicht auf grundlegende Unterschiede in der Architektur und den Prozessen hinweisen, die dem Satzverständnis zugrunde liegen, sondern auf Einschränkungen bei Verarbeitungsressourcen und -effizienz zurückzuführen sind (Hopp, 2010, 2018; Sorace, 2011).

Recently, the gender asterisk (“Gendersternchen”) has become more widespread in grammatical gender languages in order to represent all genders. Such gender-fair language is intended to help better address women and other genders and make their interests and achievements more visible. Critics often argue this would make the language less comprehensible and less aesthetically appealing. Two experiments examined the effects of the gender asterisk on text comprehensibility, aesthetic perception, and interest. N = 159 and N = 127 participants were randomly provided with a text in either masculine-only form or alternatively in gender-fair language with the gender asterisk. The results of the first experiment showed no impairment of comprehensibility and aesthetic evaluation of the texts by the gender asterisk and no effect on interest in the game, while the second experiment showed significant impairments of comprehensibility, aesthetic evaluation, and interest in the game by the gender asterisk. The proportion of singular forms is discussed as a possible explanation for the different results. Experiment 1 predominantly used plural forms like die Spieler*innen (∼“the fe*male players”) and did not include forms such as der*die Spieler*in (∼“the*the fe*male player”), whereas Experiment 2 included many such more complex singular forms. We argue that this issue might be crucial, and that it deserves full attention in future studies.

The nature of crosslinguistic influence in real-time L2 processing has been explored with various methods and language combinations and so far, the findings have shown differences between lexicon and syntax. For instance, current research on the mental lexicon shows evidence for nonselective activation of languages (Dijkstra & Van Heuven, 2002). Yet, there is less evidence for crosslinguistic influence when it comes to morphosyntax, which has been found mostly for verb subcategorization properties (Frenck-Mestre & Pynte, 1997), subject-object ambiguities (Hopp, 2006) and attachment preferences (Frenck-Mestre, 1997, 2002). The current thesis aims at investigating the extent of crosslinguistic influence in L2 sentence comprehension among highly advanced L2 English learners by focusing on the differences in the realization of grammatical gender, tense and grammatical aspect in the learners L1. Therefore, the thesis focuses on crosslinguistic influence in L2 processing and it investigates if L1 lexical-grammatical information gets activated in L2 comprehension and which factors modulate it. There are three separate studies on online comprehension of anaphoric pronouns, coreference violations and temporarily ambiguous sentences with different L1 groups (i.e., L1 German, L1 Croatian and L1 Spanish). The results show limited support for crosslinguistic influence in the case of grammatical gender and no support for crosslinguistic evidence in grammatical aspect and tense. Thus, the results replicate the findings of the studies so far that there is less clear evidence of L1 activation during L2 sentence reading compared to L1 activation in the bilingual lexicon. Some of the possible factors for the lack of crosslinguistic influence in L2 processing might be the language context, L1-L2 differences and the choice of items. Additionally, the results cannot be fully explained by leading L2 processing and acquisition theories. Yet, the study on grammatical aspect does give support for the target-like use of grammatical information in L2 sentence comprehension, which is contrary to the predictions of the Shallow Structure Hypothesis (Clahsen & Felser, 2006)., Die Art des Einflusses der L1 bei der L2-Verarbeitung in Echtzeit wurde bereits mit verschiedenen Methoden und Sprachkombinationen untersucht. Bisher zeigen die Ergebnisse Unterschiede zwischen Lexikon und Syntax. Beispielsweise zeigen aktuelle Forschungen zur lexikalischen Verarbeitung Hinweise auf eine nichtselektive Aktivierung von Sprachen (Dijkstra & Van Heuven, 2002). Es gibt jedoch weniger Hinweise auf einen Einfluss der L1 auf die Morphosyntax, die hauptsächlich für Eigenschaften der Verbunterkategorisierung (Frenck-Mestre & Pynte, 1997), Subjekt-Objekt-Ambiguitäten (Hopp, 2006) und Bindungspräferenzen (Frenck-Mestre, 1997, 2002) gefunden wurde. Die aktuelle Doktorarbeit zielt darauf ab, das Ausmaß der L1 Einflusses auf das L2- Satzverständnis unter hochentwickelten L2-Englischlernenden zu untersuchen, indem die Unterschiede in der Realisierung des grammatikalischen Geschlechts, der Zeitform und des grammatikalischen Aspekts im L1 der Lernenden manipuliert werden. Daher konzentriert sich die Doktorarbeit auf den Einfluss der L1 auf das L2 Verarbeitung und untersucht, ob lexikalisch-grammatische L1-Informationen im L2 Verständnis aktiviert werden und welche Faktoren sie modulieren. Es gibt drei separate Studien zum Online-Verständnis von Anaphorikpronomen, Verstößen gegen die Koreferenzvereinbarung und vorübergehend mehrdeutigen Sätzen mit verschiedenen L1-Gruppen (d.h. L1-Deutsch, L1-Kroatisch und L1-Spanisch). Die Ergebnisse sind eine begrenzte Unterstützung für den Einflusses der L1 im Fall des grammatikalischen Geschlechts. Für den Einfluss der L1 auf grammatikalische Aspekt und Zeitform werden keine Beweise gefunden. Somit zeigen die Ergebnisse, dass es weniger eindeutige Hinweise auf eine L1-Aktivierung während des Lesens des L2-Satzes gibt als auf eine L1-Aktivierung im zweisprachigen Lexikon. Einige der möglichen Faktoren für das Fehlen eines Spracheneinflusses bei der L2-Verarbeitung könnten der Sprachkontext, die Unterschiede zwischen L1 und L2 und die Auswahl der Elemente sein. Außerdem können die Ergebnisse nicht vollständig durch führende L2-Verarbeitungs- und Erwerbstheorien erklärt werden. Die Studie zum grammatikalischen Aspekt unterstützt jedoch die zielgerichtete Verwendung grammatikalischer Informationen beim Verständnis von L2-Sätzen, was den Vorhersagen der Shallow Structure Hypothesis (Clahsen & Felser, 2006) widerspricht.

In der höfischen Literatur des Mittelalters begegnet eine Vielzahl an Ritualen und ritualisierten Handlungen. Begrüßungen, Empfänge, Feste, Taufen, Schwertleiten, Hochzeiten, Verabschiedungen, Wehklagen und Trauerrituale folgen Konventionen, kehren in unterschiedlichen Kontexten wieder, geben dem Alltags- wie dem Festhandeln eine feste Struktur und führen zu einer Veränderung zwischen Individuen und Gemeinschaften. Trotz der Häufigkeit und Bedeutung ritueller Handlungen, wie sie sich in der mittelalterlichen Schriftkultur dokumentiert, ist ihre Definition bis in die jüngere Forschung auffallend vage, plural und heterogen geblieben und sind ritualtheoretische Ansätze in der mediävistischen Literaturwissenschaft nur gelegentlich aufgegriffen worden. Die vorliegende Studie diskutiert basierend auf Ritualbegriffen von Arnold van Gennep, Gerd Althoff, Barbara Stollberg-Rilinger u.a. kritisch den Stellenwert von Ritualen in sogenannten ‚Klassikern‘ mittelhochdeutscher Literatur unter Berücksichtigung der Forschungsmeinung, Rituale seien ordnungsgebende Instanz und zentraler Bestandteil höfischer Kultur um 1200. In Anlehnung an aktuelle mediävistische Forschungsansätze analysiert die Studie dafür ritualisierte Handlungen in den Erzählungen Hartmanns von Aue. Im ‚Erec‘, ‚Iwein‘, ‚Gregorius‘ und in ‚Der arme Heinrich‘ prüft der Autor die These, dass Rituale in der literarischen Bearbeitung keine Selbstläufer sind, sondern vielmehr durch Redeszenen, spezifische Auslassungen und Ausgestaltung sowie Strategien der Distanzierung einem Diskurs zugänglich gemacht werden. In den untersuchten Texten stellt der Autor dabei eine Tendenz zur stark reduzierten Darstellung von Ritualen bis hin zu auftretenden Leerstellen fest. Statt auf einer narrativen Ausgestaltung der symbolischen Kommunikation liegt der Fokus auf deren gesellschaftspolitischen Dimensionen und wird ein Verhandlungscharakter herausgestellt, indem Rituale zwischen Redeszenen und der Darstellung der Eignung von zentral beteiligten Figuren eingebettet werden. Im ‚Erec‘ wird auf diese Weise eine gemeinschaftlich-tragfähige Norm von Minne zunächst über Rituale verhandelbar und anschließend ausgefüllt. Im ‚Iwein‘ weisen Rituale auf die Diskursivierung juristischer Fragestellungen hin, z.B. durch gestörte Rechtsordalien, die als Gottesurteil angelegt sind und letztlich zum Stillstand kommen. Im ‚Gregorius‘ ist durch die Darstellung der extremen Buße des Protagonisten eine Aushandlung religiöser Praktiken in Korrelation zur inneren Einstellung und gesellschaftlichen Auswirkungen nachweisbar. Besonders hervorgehoben werden dort die Rituale, die einer geistlichen Sphäre zuzuordnen sind, während menschengemachte Handlungen problematisiert erscheinen und einem gemeinschaftlichen Austausch bedürfen. Ähnlich wird auch in ‚Der arme Heinrich‘ – vor dem Hintergrund medizinischer Fragestellungen – eine ritualisierte Opferungshandlung als Arznei für den erkrankten Protagonisten diskursiviert, wenn explizit das Scheitern dieser Methode zur angestrebten Heilung führt. Den Texten gemein ist der betont hohe Stellenwert gemeinschaftlicher Absprachen, wenn z.B. in ‚Der arme Heinrich‘ vor der Vermählung zwischen einem Bauernmädchen und einem Ritter eine Verhandlung stattfinden soll. Die Studie kommt zu dem Schluss, dass Rituale in der literarischen Bearbeitung Hartmanns durch narrative Strategien, wie das Spiel mit Erwartungen, dem Versatz einzelner Elemente, gezielten Auslassungen bis hin zu Leerstellen und Distanzierungen, in ihrer Bedeutung und Wirkung verhandelt werden. Der Möglichkeit des Scheiterns von Ritualen wird in Hartmanns Werken jeweils mit der Forderung nach gemeinschaftlicher Absprache begegnet. Über den Text hinaus bieten Rituale so einen Zugang, um bestehende soziale Interaktionen und die persönliche Eignung der Beteiligten zu reflektieren und zu erörtern., In the courtly literature of the Middle Ages, there is a multitude of rituals and ritualized actions. Greetings, receptions, celebrations, baptisms, sword conducting, weddings, farewells, lamentations and mourning rituals follow conventions, recur in different contexts, give everyday and festive activities a firm structure and lead to a change between individuals and communities. Despite the frequency and importance of ritual acts as documented in medieval written culture, their definition has remained remarkably vague, plural and heterogeneous until recent research, and ritual-theoretical approaches in medieval literary studies have only occasionally been taken up. Based on the concepts of ritual by Arnold van Gennep, Gerd Althoff, Barbara Stollberg-Rilinger and others, this study critically discusses the importance of rituals in the so-called 'classics' of Middle High German literature, taking into account the research opinion that rituals are a regulatory authority and a central component of courtly culture around 1200. Based on current medieval research approaches, the study analyzes ritualised actions in Hartmann von Aue's stories: 'Erec', 'Iwein', 'Gregorius' and 'Der arme Heinrich' and examines the thesis that rituals in literary processing are not a sure-fire success, but rather made accessible to a discourse through speech scenes, specific omissions and design as well as strategies of distancing. In the texts examined, the author notes a tendency towards the greatly reduced representation of rituals, to the point of gaps appearing. Instead of a narrative design of symbolic communication, the focus is on its socio-political dimensions and a negotiation character is emphasized by embedding rituals between speech scenes and the depiction of the suitability of centrally involved figures. In this way, in the 'Erec', a common, sustainable norm of love is first negotiable via rituals and then fullfilled. In the 'Iwein', rituals point to the discursivization of legal issues, e.g. through disturbed legal ordinances, which are designed as a judgment of God and ultimately come to a standstill. In 'Gregorius', the portrayal of the protagonist's extreme penance reveals a negotiation of religious practices in correlation with the inner attitude and social effects. The rituals that are assigned to a spiritual sphere are particularly emphasized there, while man-made actions appear to be problematic and require a community exchange. Similarly, in 'Der arme Heinrich' – against the background of medical issues – a ritualized sacrificial act is discursed as medicine for the ill protagonist when the failure of this method explicitly leads to the intended healing. What the texts have in common is the emphasized importance of joint agreements, e.g. in 'Der arme Heinrich', a negotiation is to take place before the marriage between a peasant girl and a knight. The study concludes that rituals in Hartmann's literary treatment are negotiated in terms of their meaning and effect through narrative strategies, such as playing with expectations, the misalignment of individual elements, targeted omissions, gaps and distancing. In Hartmann's works, the possibility of rituals failing is met with the demand for collective agreement. Beyond the text, rituals therefore offer access to reflect on and discuss existing social interactions and the personal suitability of those involved.

Jahrbuch der Braunschweigischen Wissenschaftlichen Gesellschaft, vol. 2020Jahrbuch 2020 der Braunschweigischen Wissenschaftlichen Gesellschaft, vol. 2020, p. 325

Most of our knowledge related to how social exclusion affects those who ostracize and those who are being ostracized is based on questionnaires administered after the ostracism situation is over. In this research, we strived to further our understanding of the internal dynamics of an ostracism situation. We therefore examined individuals' language-specifically, function words-as a behavior indicative of psychological processes and emergent states that can be unobtrusively recorded right in the situation. In online chats, 128 participants talked about a personal topic in groups of three. In the experimental group (n = 79), two conversation partners ignored every contribution by the third. We found that, compared to the control group, these targets of ostracism used language indicative of a self-focus and worsened mood, but not of social focus or positivity, although positivity was related to a writer's likeability. Sources of ostracism used language suggesting that they were distancing themselves from the situation, and they further engaged in victim derogation. We discuss how our results highlight the severity and potential self-sustainability of ostracism.

ISBN:978-3-906916-72-9

In recent years, network bandwidth has increased at a rapid pace, moving from 10G, to 40G/100G, to 400G. CPU frequency on the other hand has been stagnant. As a result, more and more CPU cycles have to be allocated to network processing to keep up with the increasing network bandwidth. At the same time, we have seen exponential growth in data requiring distributed systems to process it. While these systems benefit from the increased network bandwidth, their ability to scale out is limited by inefficiencies in I/O and data movement. Under the premise that most of such computation is distributed, one question remains as to whether data processing operations can be offloaded to the network. In this thesis, we explore two approaches to enable in-network data processing and thereby reduce burden on the CPU. Our first approach places data processing accelerators, such as FPGAs, in the network. To enable network-attached accelerators, we introduce a scalable network stack architecture for FPGAs supporting thousands of concurrent connections, a requirement of data center applications. Based on this architecture, we implement a TCP/IP stack that allows the seamless integration of accelerators into the existing network infrastructure. In our second approach, we enable data processing operations in the network by enhancing the functionality of the network card. In particular, we present Smart Remote Memory (STROM), which extends Remote Direct Memory Access (RDMA) with the capability to deploy acceleration kernels on the network card. These acceleration kernels can, for instance, directly access the remote host memory to traverse remote data structures such as lists or indexes. In addition to that, they can execute stream-based operations, e.g., filtering or aggregation, on data that is passing through the network card to the host memory, or vice versa. As part of this thesis, we implemented and released an open source TCP/IP stack for reconfigurable hardware. The stack enables network-attached accelerators and has been deployed successfully in a number of research and commercial systems. Following on from this success, we released an open source FPGA-based RDMA NIC, which we used as a prototyping platform for STROM, to facilitate more research in this area.

3D hand pose estimation aims at recovering 3D coordinates of joints or mesh vertices of hand from visual inputs. It has important applications in human-human or human-machine interactions and computer animation. Given the large view point change, cluttered background, finger similarity and self-occlusion, accurately estimating 3D hand pose in real time remains very challenging even after decades of research. In this thesis, we propose a series of learning based methods to 1) improve the estimation accuracy, 2) enable training with less or even no human annotation and 3) apply the estimation system to human-computer interaction with passive haptic feedback. The majority of the methods focuses on estimating pose of a single hand from depth data, with the only exception of human-computer interaction application, in which monochrome inputs are used. To improve the accuracy of hand pose estimation system, the first proposed method incorporates local geometric properties, e.g., surface normal and curvature, into random forest to achieve better invariance to view-point changes. Then, we propose another method reformulating hand pose estimation as regressing dense vector field with 2D fully convolutional network. Finally, we propose to estimate 3D coordinates of both joints and mesh vertices by establishing dense correspondence between input depth map and template mesh surface with just one forward pass of an highly efficient 2D fully convolutional network. Learning based hand pose estimation methods, especially deep learning based methods, requires large amount of accurate annotation on real samples to achieve high accuracy. However, acquiring such accurate annotated real samples can be extremely difficult and expensive. To mitigate the dependency on large amount of annotated real samples, we propose to leverage unlabelled real samples from two perspectives. First, we utilize deep generative models to formulate hand pose estimation in a semi-supervision setup. Then, we bridge model based and discriminative approaches to enable training network in a self-supervision way, i.e., by using the model fitting error to train the neural network. Finally, we demonstrate an interaction application in mixed reality with hand pose estimation system. While the system works on more energy efficient and cheaper monochrome camera, the training samples of which are automatically generated by the accurate depth based system. All these methods are thoroughly evaluated and compared with state-of-arts on a set of benchmarks. We then conclude and discuss their limitations and new ideas for future work.

Automatic content scoring is an important application in the area of automatic educational assessment. Short texts written by learners are scored based on their content while spelling and grammar mistakes are usually ignored. The difficulty of automatically scoring such texts varies according to the variance within the learner answers. In this paper, we first discuss factors that influence variance in learner answers, so that practitioners can better estimate if automatic scoring might be applicable to their usage scenario. We then compare the two main paradigms in content scoring: (i) similarity-based and (ii) instance-based methods, and discuss how well they can deal with each of the variance-inducing factors described before. OA Förderung 2019

Financial traders usually interact with high amount of data visualized on several displays. The display properties of their systems as well as their previous experience are expected to have an influence on the quality of their interaction and the resulting financial decisions. However, to date, little empirical evidence is available addressing financial user interfaces. Research in human factors with professionals from data-rich domains has shown that display clutter can degrade user performance, lead to confusion, and influence the distribution of visual attention. Furthermore, previous findings indicate that the impact of display clutter varies depending on the expertise of the system user. This dissertation presents four papers investigating how display clutter and expertise can have an impact on visual attention and financial judgment performance. The first paper approaches display clutter theoretically in connection with the concept of user experience, and lays out how display clutter can influence the user experience. The second paper investigates main effects and the interaction effect of display clutter and expertise in a controlled two-by-two experiment. It shows that not only expertise, but also the combination of expertise and display clutter have an effect on judgment performance as well as the visual attention control. The third manuscript confirms the results from paper two with more ecologically valid stimuli and explores the impact of display clutter on two different financial tasks, risk judgment performance and investment decisions. It shows that the type of task has an influence on the consistency, concordance as well as confidence of risk judgments and investment decisions; the visual attention measures, however, are not affected. In the fourth manuscript the focus lies on the influence of expertise on risk judgment performance in relation to judgment confidence. The results show confidence differences between experts and novices, despite similar judgment performances.

Jahrbuch der Braunschweigischen Wissenschaftlichen Gesellschaft, vol. 2018Jahrbuch 2018 der Braunschweigischen Wissenschaftlichen Gesellschaft, vol. 2018, p. 144

Trapped-ions form a promising platform to realize a future large scale quantum computing device. Qubits are typically stored in internal electronic states, which are coupled using their joint motion in the trap potential. In this thesis this control paradigm is reversed. The harmonic motion of a trapped calcium ion forms the main subject of studies, which is controlled via the internal electronic states. A number of new techniques are introduced and examined, primarily based on the implementation of modular variable measurements. These are realized combining an internal state dependent optical dipole force with readout of the internal states. Modular measurements are used to investigate large "Schrödinger cat'' states of the ion's motion, to violate Leggett-Garg tests of macroscopic realism, and finally to realize a logical qubit encoded in an error-correcting code based on the trapped-ion oscillator. The latter offers an alternative to the standard qubit based quantum information processing approach, which when embedded in systems of coupled oscillators could lead to a large-scale quantum computer. Measurements of a particle's modular position and momentum have been the focus of various discussions of foundational quantum mechanics. Such modular measurements of the trapped-ion's motion are studied in depth in this thesis, in particular their ability to commute, which forms a key element for the latter work on error-correcting codes. Here we make use of the ability to investigate sequences of measurements on a single harmonic oscillator, and study correlations between their results, as well as quantum measurement disturbances between the measurements. In order to achieve the major results of the thesis, it was necessary to characterize and control multiple wave packets in phase space. On the characterization side, the need to cope with states with high energy occupations led to the development of multiple new methods for quantum state tomography, including the use of a squeezed eigenstate basis, and the direct extraction of the characteristic function of the oscillator using state-dependent forces. These were used to analyze some of the largest oscillator "Schrödinger cat'' states which have been produced to date. The main result of this thesis is encoding and full control of a logical qubit in the motional oscillator space using a code proposed 18 years ago by Gottesman, Kitaev and Preskill. Logical code states are realized and manipulated using sequences of up to five modular measurements applied to an ion initially prepared in a squeezed motional state. Such sequences realize superpositions of multiple squeezed wave packets, which form the code words. The usage of the oscillator enables to encode and in principle correct a logical qubit within a single trapped ion, which when compared to typical qubit-array based approaches simplifies control and hardware. While the discussion above focuses on the new physics in this thesis, in addition the work required technical upgrades to the system, improving control of both qubit and oscillator. These form important components which have impact on all experiments in our setup, beyond the bounds of the current thesis., ISBN:5800134927809

Some computational tasks, such as solving the discrete logarithm problem, can be computed with exponential better scaling on quantum computers than using classical hardware. Due to these algorithms, a digital quantum computer is strongly sought after. A scalable quantum computer needs fault-tolerant protocols to correct for errors and deviations due to the environment. These protocols add additional layers of complexity because of their non-trivial logical operations which need to be decomposed into operations on many physical qubits. This thesis covers an essential part for the endeavor of building a quantum computer: classical control software for a large, fault-tolerant quantum computer. It will explain how to implement a quantum algorithm by manipulations of logical qubits, i.e. degrees of freedom that the algorithm uses for its computation. These logical qubits are encoded in many qubits, such that a single logical operation requires operations on many different qubits. A quantum compiler can perform a translation from the high-level quantum algorithm into individual hardware instructions. The surface code is one of the most promising error-correction protocols because of its 2D nearest-neighbor interactions and high threshold. The topological nature of this code allows logical operations to be understood by simple geometric manipulations. Therefore, the surface code and its variants are used as the underlying error-correcting codes throughout this thesis. Lattice surgery is a means to perform logical operations while using boundary operations between different planar surface codes. Other methods, such as braiding, have already been investigated, but their optimization seems to be more difficult than lattice surgery. This thesis presents an alternative way to perform computation using the fundamental operations of lattice surgery directly, instead of simulating a predefined set of universal gates. To this end, a representation is developed that is similar to measurement-based quantum computation in an encoded environment. The optimization of this method is still NP-hard. While the previous error-correction protocols are hardware agnostic, the surface code can be reformulated to better match a specific hardware architecture. In particular fault-tolerant computation on linear-optics quantum computers are explored. On such architectures, measurements are used to teleport information through an entangled lattice, called the Raus\-sendorf lattice. This lattice has intrinsic error-correction capabilities, and the previous lattice surgery protocols can be applied to this measurement-based setting. Finally, it is investigated how fault-tolerance can be maintained with realistic failure rates for the non-deterministic creation of the lattice.

Many algorithms used in hardware applications across disciplines, such as the fast Fourier transform, the Walsh-Hadamard transform or sorting networks share a common structure. They consist of stages of parallel and identical processing elements that each operates on two inputs with different data permutations in between. The symmetry in this structure enables folding to an area-efficient in a streaming architecture that accepts the input over several cycles. However, necessary for folding are streaming permutation circuits that require memory and routing components. In this dissertation, we focus on the optimal implementation of these algorithms. We provide lower bounds for the implementation of streaming permutations, and propose algorithms that produce solutions that match it (under certain assumptions). We apply these algorithms in the context of a generator capable of implementing the hardware applications previously mentioned. Finally, for a given problem, we address the question of finding an optimal algorithm, i.e., an algorithm which streaming implementations are the cheapest to implement.

Mittels der Fusion der Ansätze der konzeptuellen Integration nach Fauconnier & Turner (2003), als Ansatz über die allgemeine kognitive Verarbeitung von Wissen, und der Diasystematischen Konstruktionsgrammatik (DCxG) nach Höder (z.B. 2012, 2018), als Ansatz zum diachronen Sprachwandel, wird in dieser Arbeit ein Modell vorgeschlagen, das - die kognitive Sprachverarbeitung multilingualer SprecherInnen berücksichtigend - auf unterschiedlichste Sprachkontaktphänomene, oder auch (Ad-hoc-)Innovationen, angewandt werden kann. Dabei wird die traditionelle kategorische Unterscheidung zwischen einzelnen Kontaktphänomenen aufgeweicht, wodurch auch Phänomene beschrieben werden können, die in der traditionellen Sprachkontaktforschung bisher als Sonderfälle gelten mussten. Zusätzlich zeigt die Modellierung, dass eine prinzipielle Unterscheidung zwischen Ad-hoc-Innovationen und konventionalisierten Innovationen für die Anwendung des Modells nicht notwendig ist, womit die DCxG um eine synchrone Perspektive erweitert wird. Diese Arbeit leistet insgesamt einen Beitrag zur Modellierung von Sprachkontaktphänomenen aus der Perspektive der Kognitiven Linguistik im Allgemeinen.

The thesis examines typical errors in the pronunciation of German second language (L2) Danish learners and their evaluation by native Danish speakers. The first part of the thesis investigates into typical errors in empirical L2 speech data based on a contrastive analysis of the German and Danish sound systems. In the second part of the thesis, the typical pronunciation errors are analyzed in terms of their different error gravity. Danish native speakers rated different errors in the dimensions comprehensibility, accentedness and acceptability. In the summary, the findings of both studies are discussed with regard to their relevance for Danish language teaching in schools

Today’s hardware is increasingly parallel, and modern programming languages must thus allow a programmer to use this parallelism effectively. For languages that depend on a virtual machine (vm), it is the responsibility of the vm to execute code efficiently and in parallel. In the case of Python, a popular dynamic language, several vms exist, but none of them can deliver high performance and parallel execution at the same time. The reason for the lack of such a high-performance, parallel vm for Python lies with Python’s concurrency semantics, which is based on a strong memory model and atomic operations on large entities. Such a concurrency semantics does not map well to modern hardware architectures, which is why parallel vms must emulate this semantics under parallel execution through expensive synchronization. This dissertation introduces a new approach to the design and construction of high-performance, parallel vms with a concurrency semantics such as Python’s. The introduction of large-scale atomicity to the implementation language of a vm lets a vm developer specify the concurrency semantics independently from the vm’s synchronization mechanism. Thereby, the used synchronization mechanism becomes an exchangeable vm component. For the high-performance execution of Python code, just-in-time compilation is an essential concern. Unfortunately, Python’s strong memory model inhibits basic compiler optimizations under concurrency. Hence, to allow a compiler to optimize effectively, the concept of Parallel Worlds is introduced. Parallel Worlds transparently isolate concurrent computations from each other, and thereby allow for effective optimizations under the assumption of no concurrency. The transparent isolation of Parallel Worlds is supported by a special-purpose software transactional memory system (stm). Apart from isolation, this stm is the key enabler for the efficient parallel execution of Python code. Parallel execution builds on the speculative execution capability of the stm. The product of this dissertation is PyPy-stm, a high-performance, parallel vm for Python. With PyPy-stm, multi-threaded Python programs can take advantage of the parallelism in modern hardware. On a set of benchmark programs, PyPy-stm outperforms established Python vms such as CPython, Jython, IronPython, and PyPy. Compared with PyPy, the current top-of-class in program performance, PyPy-stm achieves speedups in the range of 1.5 to 8.0× with 8 threads available. These results confirm the viability of the approach and show that PyPy-stm deserves the designation as a high-performance, parallel vm for Python.

The Synchronized-By-Default (SBD) concurrency model synchronizes all accesses to shared memory by default. Synchronization takes place by ensuring that all program code runs inside atomic sections even if this program code has external side effects. Threads are mapped to atomic sections by default. A programmer must explicitly break this atomicity to increase concurrency by splitting an atomic section. The SBD concurrency model differs from explicit synchronization where a programmer must add atomicity where needed. Under explicit synchronization, missing atomicity causes data races and race conditions, while under the SBD concurrency model missing atomicity breaks cause serialization. A naive implementation of this approach incurs a large amount of overhead. Previous work on similar approaches concentrated on simulated hardware transactional memory to reduce the overhead and on the formal semantics of such a concurrency model. In the thesis, we show how to reduce the overhead to make the approach suitable for realistic application programs on existing hardware using software transactional memory (STM). For this reason, we present an explicit semantics for SBD and novel STM techniques that in combination with a corresponding programming discipline reduce the overhead of the SBD approach in terms of performance and memory overhead. At the core of SBD lies an explicit semantics. This semantics allows a programmer to avoid performance overhead by reducing the amount of speculation typically found in transactional memory. We show that exploiting the combination of these explicit semantics, large atomic sections, properties of managed, object-oriented programming languages, intraprocedural static analyses, and transactional I/O results in good scaling properties and moderate sequential overhead.

The goal in this dissertation is the efficient synthesis of photorealistic images on a computer. Currently, by far the most popular approach for photorealistic image synthesis is path tracing, a Monte Carlo simulation of the integral equations that describe light transport. We investigate several data-driven approaches for improving the convergence of path tracing, leveraging increasingly sophisticated machine-learning models. Our first approach focuses on the specific setting of ``multiple scattering in translucent materials'' whereas the following approaches operate in the more general ``path-guiding'' framework. The appearance of bright translucent materials is dominated by light that scatters beneath the material surface hundreds to thousands of times. We sidestep an expensive, repeated simulation of such long light paths by precomputing the large-scale characteristics of material-internal light transport, which we use to accelerate rendering. Our method employs ``white Monte Carlo'', imported from biomedical optics, to precompute in a single step the exitant radiance on the surface of large spherical shells that can be filled with a wide variety of translucent materials. Constructing light paths by utilizing these shells is similarly efficient as popular diffusion-based approaches while introducing significantly less error. We combine this technique with prior work on rendering granular materials such that heterogeneous arrangements of polydisperse grains can be rendered efficiently. The computational cost of path construction is not the only factor in rendering efficiency. Equally important is the distribution of constructed paths, because it determines the stochastic error of the simulation. We present two path-guiding techniques that aim to improve this distribution by systematically guiding paths towards scene regions with large energy contribution. To this end, we introduce a framework that learns a path construction scheme on line during rendering while optimally balancing the computational rendering and learning cost. In this framework, we use two novel path-generation models:\ a performance-optimized spatio-directional tree (``SD-tree'') and a neural-network-based generative model that utilizes normalizing flows. Our SD-tree is designed to learn the 5-D light field in a robust manner, making it suitable for production environments. Our neural networks, on the other hand, are able to learn the full 7-D integrand of the rendering equation, leading to higher-quality path guiding, albeit at increased computational cost. Our neural architecture generalizes beyond light-transport simulation and permits importance sampling of other high-dimensional integration problems.

Every second the world around us changes in appearance and geometry. Buildings get torn down, businesses open and close while light and shadow play throughout the day. But even low-frequency variation from seasonal change influence how the environment looks like and challenge us anew to recognize the path to work and back home. While building a mental model of our environment and recognizing it on our next visit seems effortless to us, teaching machines to do so is a huge undertaking. The main theme of this thesis is to research algorithms and systems that allow robots and mobile devices to understand their position in an environment over long-periods. With such understanding, we hope, robots can become more autonomous in building and maintaining their map of the environment s.t. they can collaborate and become more useful for humans. The goal of the research is the question of how we can efficiently capture the ever changing world around us s.t. we can localize agents in that model while concurrently updating it from newly captured data. A typical method for capturing the environment is to use photo collections and run "Structure from Motion, a class of algorithms which jointly estimates the position and orientation of the cameras at time of exposure as well as the geometry of the environment in the form of sparse 3d points. Each such 3d point corresponds to a salient part in the environment and is associated with a fingerprint that describes its appearance. The resulting model allows computers to recognize a scene by comparing visual fingerprints between the model and camera live stream. If we were able to build the model, identify changes in the environment and communicate updates to all other agents in real-time, the above posed question would be considered solved. Building the model and identifying the dynamic parts however is computationally costly and sharing the resulting model requires substantial network bandwidth. Most use-cases additionally require localization and state-estimation to run in real-time for user feedback or robotic control which adds computational load on the agent. We address these limitations and challenges by proposing a set of algorithms that span the cycle from map collection, change identification, compression of the environment to real-time localization. The proposed approaches target at systems that are resource constraint given the size of their operating environment such as found in robotics and mobile applications. Depending on the use-case this ranges from a single processor platforms to data-center sized infrastructure; localizing a group of robots or millions of mobile phones. In the first part of the thesis we present algorithms which provide real-time pose estimates which is the basis for navigating and capturing the environment. We start from a extended Kalman filter formulation and extend it with the capability to fuse multiple sensors with time delays and add robustness to sensor signal loss. To ensure constant and real-time performance the approach keeps only limited memory of the environment, which is sufficient to locally prevent positional drift. Given noisy sensor signals and modeling errors, the positioning accuracy of such approach however degrades as the platform moves through the environment. To correct for the resulting drift in pose we combine the pose estimator with a loop-closure module which is able to recognize previously visited places and thus can correct for accumulated error. We propose algorithms that change the definition of a place s.t. it becomes continuous instead of the arbitrary discretization applied previously in literature which improves recognition rates substantially. To accelerate the model-capturing process we combined the algorithms in a multi-agent mapping system where agents share data to collaboratively build a joint model of the environment. In our second project we take the information from multiple agents and fuse it into a common lifelong map to improve robustness and coverage. We investigate which signals can be used to identify overlapping, redundant or outdated information about the environment and propose optimization strategies to reduce the model size while maintaining localization performance. Combined with efficient indexing schemes we show that localization queries can run in real-time even against models spanning large environments such as entire cities. Our experiments demonstrate that artifacts from high compression rates can be mitigated by fusing real-time the localization signal with the on-board state estimation. The overall research is motivated by the vision to localize a large number of agents in a dynamic, real-world non lab environment while concurrently updating the model of it from new experience captured by the agents over their lifetime. As a proof of concept the algorithms form the foundation of Google's "Visual Positioning Service" where they allow the real-time localization of mobile phones for both small and large scale Augmented Reality applications.

Technical advances and the decreasing costs of photovoltaic (PV) and battery (B) systems are key drivers for the consumer-prosumer transition, and crucial components in many countries' energy strategies for reducing the emission intensity in energy systems and supporting the phase-out of nuclear energy generation. The build-up and expansion of solar capacities in response to the gradual reduction of electricity generated from fossil or nuclear fuels leads to an increasingly decentralized energy supply system, which consists of many small utility- and residential-scale generators spread across a service area alongside larger centralized power plants that compensate for the intermittent nature of renewable energy sources. Consequently, the decentralization of the energy system also implies a shift from large (mostly publicly owned) investors in centralized power stations to smaller, private investors in small-scale PV systems. During the past decade, remuneration rates for PV systems in many countries have been continuously reduced in response to falling PV module costs. In some service areas, the remuneration rate is considerably less than the retail rate for energy, making it more attractive to self-consume a large share of the generated electricity. Self-consumption is defined as the share of the annually generated energy that can be directly consumed by the prosumer. The fraction that cannot be self-consumed is injected into the grid at a remuneration rate. The achieved rate of self-consumption is an important factor that influences the profitability of photovoltaic battery (PVB) systems and depends (among other factors) on how well the load profile of a household concurs with its solar production. However, the influence of the individual load profiles on the profitability of PVB systems has received little attention in the academic literature. Therefore, the first objective of this thesis is the assessment of the impact of load profile heterogeneity on the profitability of solar battery systems under various system cost scenarios. A techno-economic simulation model has been developed to account for load profile heterogeneity by processing a dataset consisting of more than 4000 real world load profiles with today's costs (without subsidies) and future cost scenarios. The simulation results for the location of Zurich suggest that the profitability of PVB systems varies considerably between households, even for households with comparable total annual demands. About 40% of the households can reach profitability under today's cost assumption (2000 €/kWp, 1000 €/kWh. Solar becomes profitable for nearly all (94%) analyzed households if costs drop to 1000 €/kWp. For battery storage to be profitable for prosumer households, battery system costs must fall to 250-500 €/kWh. Utility companies and service providers have started to offer information systems that enable households to estimate the profitability of PVB systems based on some basic inputs, such as annual demand or orientation of the solar arrays. However, these information systems omit household specific load profile data, which also contains information about the shape of the load profile. The integration of load profile data can provide more personalized and reliable estimation of profitability, optimal PV and battery size, as well as annual performance factors, like expected self-consumption and self-sufficiency rates. However, household specific smart-meter data is often not available because utilities do not store smart-meter data with sub-hourly resolution by default. Thus, the estimation of household specific profitability, optimal system sizes and expected annual performance factors need to rely on smart-meter data recorded over a time period comparable to the duration of the tendering process for a PVB system, which typically lasts a few months. At the beginning or throughout the tendering phase, smart-meter data can be collected to estimate profitability, optimal sizes and annual performance factors would allow households to make more cautious and confident investment decisions. The academic literature already provides statistical methods that help predict self-consumption or self-sufficiency from annual demand or installed solar power. However, very few existing studies include the load profile and its shape as an input factor and apply it to an unseen dataset large enough to verify its application in the tendering phase of a PVB system. Given the importance of the load profile for the personalized estimation of profitability, the second objective of the dissertation is the development and evaluation of a machine learning algorithm for the prediction of PVB profitability. The proposed machine learning algorithm uses the load profile as an exclusive input factor and predicts the PVB profitability, along with the optimal PV size, optimal battery size, achieved self-consumption and self-sufficiency rates. Good prediction accuracy was found if the smart-meter data is collected over a minimum timeframe of 30 days, which is acceptable for the prediction of profitability, optimal system sizes and annual performance factors during the tendering phase of a PVB system. Given the growing gap between the remuneration of injecting generated electricity into the grid and retail rates, one promising alternative might be electricity trading in peer-to-peer (P2P) networks. P2P energy networks may offer an alternative system in which prosumers market their surplus energy directly to consumers at a rate closer to the retail tariff that households face when they consume energy from the utility. Prosumers are therefore incentivized to maximize the share of their surplus generation that can be traded within a P2P network. However, little is known about optimal network or community configurations (i.e. number of households, prosumer share, generation and storage capacity) that reduce imports and exports out of the P2P network compared to stand-alone households. Hence, the third objective of the dissertation is the quantification of optimal P2P network configurations regarding optimal number of households, generation and storage capacity. To this end, the previously mentioned techno-economic model is extended to account for energy flows between prosumers and consumers that are part of a P2P network. Energy flows between peers lead to reductions in imports and exports at the community-level and can be compared against the sum of individual household-level imports and exports under various community configurations. The simulation results indicate that increasing the number of households within a community reduces the imports and exports out of the community, thus increasing the self-consumption and self-sufficiency rate of the entire community and enabling prosumers to market parts of their surplus generation directly to peers. However, the results also indicate that the self-consumption and self-sufficiency rate of the community already levels off with 10 participating households. Hence, adding more households to a community will only marginally increase its self-consumption and self-sufficiency rate. Therefore, larger communities are only marginally more effective in reducing imports and exports compared to many small communities consisting of at least 10 households. Prosumer-dominated communities are less effective in reducing imports and exports compared to communities with smaller prosumer shares, because the surplus generation and demand concur less frequently in communities consisting predominantly of prosumers. In a P2P community network, households should store energy only if that unit of energy cannot be simultaneously consumed elsewhere in the community. The more surplus energy generated by prosumers in a network that can be directly consumed by other households within the community, the less storage capacity is required. The simulations indicate that battery storage systems -- sized to maximize the net present value -- are not required if the prosumer ratio in the community is not larger than 25%. Even in the extreme scenario where all members of the community are prosumers, the battery units of the individual households can be reduced by 26% if the batteries cooperate to reduce imports and exports at the community-level. Overall, the results of this thesis are relevant for households making individual investment decisions as well as for utility companies to more effectively identify and approach relevant customers for the installation of PVB systems. The results indicate that large spreads in profitability will continue to persist if costs for PVB systems fall in the future. Furthermore, the dissertation develops and evaluates a prototypical machine learning algorithm for the prediction of profitability, optimal system sizes and annual performance, even if incomplete input data, such as load profiles, are available. Beyond that, the dissertation provides simple design guidelines for P2P energy networks, suggesting that small communities of 10 households already maximize the self-consumption and self-sufficiency rate by effectively reducing imports and exports compared to individual households. Given the small community size required to reduce imports and exports, communities should be embedded locally on the low-voltage grid. The results of this dissertation highlight the need to discuss and develop regulations that facilitate community building and tariff schemes that incentivize local balancing.

The design of machine learning algorithms is often conducted in a system-agnostic manner. As a consequence, established methods may not be well aligned with the particularities of the systems on which they are deployed. In this thesis we demonstrate that there is a huge potential for improving the performance and efficiency of machine learning applications by incorporating system characteristics into the algorithm design. We develop new principled tools and methods for training machine learning models that are theoretically sound and enable the systematic utilization of individual hardware resources available in heterogeneous systems. In particular, we focus on lowering the impact of slow interconnects on distributed training and exploiting hierarchical memory structures, compute parallelism and accelerator units.

Even though capacitive touchscreens represent nowadays a prevalent technology, Body Channel Communication (BCC) is relatively unknown – nevertheless, these two technologies share the same core idea. Both concepts work with weak electric signals and rely on the capacity of the human body to act as a conductor. However, while touchscreens work with the disrupted signals at the place of origin, BCC proposes to employ the human body as carrier medium for digital communication purposes in a tangible interaction setup. In fact, BCC reinvents touch interactions by augmenting its user’s touch events with digital content. When a person holds two BCC-enabled objects in her/his hands, she/he creates a physical transmission path between them, which allows general data transmission between those devices. The user naturally controls that data transmission by simply touching or releasing the objects. BCC holds various exciting opportunities for the future generation of applications in terms of interactivity or simply as a communication alternative. Yet, the exact prerequisites, operating conditions, and possibilities of using BCC on application level have not been explored. Opposed to most work in the space, this dissertation is not primarily concerned with the BCC transmission mechanism itself, but is motivated by the vision of having BCC as an accessible component in complex end-user application design. In order to take a leap forward towards this vision, we propose a systematic breakdown of the problem space. By following a typical network stack approach, we can analyze separately the channel, the base system requirements, various networking concepts, and eventually, application level integration – all while keeping not just the common objective in mind, but also building an understanding how each of these layers affect each other.

Pipelining is a well-known technique that enables parallel execution of loops with cross-iteration dependences. Loop iterations are partitioned into a sequence of stages, and parallelism arises from the overlapped execution of different loop iterations. A large class of loops can be expressed as linear pipelines if data flow from earlier to later stages. Various pipelining techniques have been explored but reconciling load-balancing and efficient execution is still challenging. In particular, pipelining techniques that rely on static scheduling algorithms perform poorly for load-imbalanced linear pipelines. Furthermore, previous research efforts that achieve load-balancing are restricted to work-stealing solutions, and they imply high overhead for fine-grained linear pipelines. Hence, the design of pipelining techniques should combine the efficiency of static scheduling algorithms with the load-imbalance tolerance of work-stealing. In this dissertation, we investigate the performance of different techniques that achieve load-balancing for load-imbalanced linear pipelines. We explore the design space by unifying different mapping schemes, data structures, algorithms, and mechanisms to achieve efficient and scalable execution on multi-core multi-processors. The design of all presented techniques is tailored to linear pipelines, i.e., it considers their properties to make efficient execution of fine-grained workloads possible. In particular, the design of the pipelining techniques exploits the structure of linear pipelines by using knowledge that is explicit in domain-specific abstractions. Thus, we introduce Proteas, a parallel programming model that provides directive-based transformations for the presented pipelining techniques, and we have developed these transformations on a prototype source-to-source compiler that generates optimized parallel code. Finally, the evaluation on a 44-core machine, using programs selected from three popular benchmark suites, shows the parallelization tradeoffs of the different design decisions. In other words, different pipelining techniques perform well for different situations. Nevertheless, provided that the runtime system is properly configured, all techniques achieve good performance for all considered programs, and they significantly outperform the state-of-the-art that relies on work-stealing for the execution of fine-grained linear pipelines.

Large scale systems are forecasted to greatly impact our future lives thanks to their wide ranging applications including cooperative robotics, mobility on demand, resource and task allocation, supply chain management, and many more. While technological developments have paved the way for the realization of such futuristic systems, we have a limited grasp on how to coordinate the behavior of their individual components to achieve the desired global objective. With the objective of advancing our understanding, this thesis focus on the analysis and coordination of large scale systems without the need of a centralized authority. At a high level, we distinguish these systems depending on wether they are composed of cooperative or non-cooperative subsystems. In regard to the first class, a key challenge is the design of local decision rules for the individual components to guarantee that the collective behavior is desirable with respect to a global objective. Non-cooperative systems, on the other hand, require a more careful thinking in that the designer needs to take into account the self-interested nature of the agents. In both cases, the need for distributed protocols stems from the observation that centralized decision making is prohibited due to the scale and privacy requirement associated with typical systems. In the first part of this thesis, we focus on the coordination of a large number of non-cooperative agents. More specifically, we consider strategic decision making problems where each agent's objective is a function of the aggregate behavior of the population. Examples are ubiquitous and include social and traffic networks, demand-response markets, vaccination campaigns, to name just a few. We present two cohesive contributions. First, we compare the performance of an equilibrium allocation with that of an optimal allocation, that is an allocation where a common welfare function is maximized. We propose conditions under which all Nash equilibrium allocations are efficient, i.e., are desirable from a macroscopic standpoint. In the journey towards this goal, we prove a novel result bounding the distance between the strategies at a Nash and at a Wardrop equilibrium that might be of independent interest. Second, we show how to derive scalable algorithms that guide agents towards an equilibrium allocation, i.e., a stable configuration where no agent has any incentive to deviate. When the corresponding equilibria are efficient, these algorithms attain the global objective and respect the agents' selfish nature. In the second part of this thesis, we focus on the coordination of cooperative agents. We consider large-scale resource allocation problems, where a number of agents need to be allocated to a set of resources, with the goal of jointly maximizing a given submodular or supermodular set function. Applications include sensor allocation problems, distributed caching, data summarization, and many more. Since this class of problems is computationally intractable, we aim at deriving tractable algorithms for attaining approximate solutions, ideally with the best possible approximation ratio. We approach the problem from a game-theoretic perspective and ask the following question: how should we design agents' utilities so that any equilibrium configuration recovers a large fraction of the optimum welfare? In order to answer this question, we introduce a novel framework providing a tight expression for the worst-case performance (price of anarchy) as a function of the chosen utilities. Leveraging this result, we show how to design utility functions so as to optimize the price of anarchy by means of a tractable linear program. The upshot of our contribution is the design of algorithms that are distributed, efficient, and whose performance is certified to be on par or better than that of existing (and centralized) schemes., ISBN:978-3-906916-47-7

The virtual world is increasingly merging with the real one. Consequently a proper human representation in the virtual world is becoming more important as well. Despite recent technological advances in making the virtual human presence more realistic, we are still far from having a fully immersive experience in the virtual world, in part due to the lack of proper capturing and modeling of a virtual double. Thus, new methods and techniques are needed to obtain and recover a realistic virtual doppelg¨anger. This thesis aims to make virtual human representation accessible for every person, by showcasing how it can be obtained under inexpensive minimalistic sensor requirements. Potential fields of application of the findings could be the estimation of body shape from selfies, health monitoring and garment fitting. In this thesis we investigate the problem of reconstructing the 3D virtual human from monocular imagery, mainly coming from an RGB sensor. Instead of focusing on the full avatar at once, we separately consider three constituting parts of it: the naked body, clothing and the human hand. The preeminent focus is on the estimation of the 3D shape and pose from 2D images, e.g. taken from a smart-phone, making use of data-driven priors in order to alleviate this ill-posed problem. We utilize discriminative methods, with a focus on CNNs, and leverage existing and new realistically rendered synthetic datasets to learn important statistics. In this way, our presented data-driven methods can generalize well and provide accurate reconstructions on unseen real input data. Our research is not only based on single views and annotated groundtruth data for supervised learning, but also shows how to utilize multiple views simultaneously, or leverage from them during training time, in order to boost performance achieved from a single view at inference time. In addition, we demonstrate how to train and refine unsupervised with unlabeled real data, by integrating lightweight differentiable renderers into CNNs. In the first part of the thesis, we aim to estimate the intrinsic body shape, regardless of the adopted pose. Under assumptions of uniform background colours and poses under minimal self-occlusion, we show three different approaches for estimating the body shape: Firstly, by basing our estimation on handcrafted features in combination with CCA and random forest regressors, secondly by basing it on simple standard CNNs, and thirdly by basing it on more involved CNNs with generative and cross-modal components. We show robustness to pose changes, silhouette noise and state-of-the-art performance on existing datasets, outperforming also optimization based approaches. The second part of the thesis tackles the estimation of garment shape from one or two images. Two possible estimations of the garment shape are provided: one that gets deformed from a template garment (i.e. from a t-shirt or a dress) and second one that gets deformed from the underlying body. Our analysis includes empirical evidence which shows the advantages and disadvantages of utilizing either of the estimation methods. We adopt lightweight CNNs in combination with a new realistically rendered garment dataset, synthesized under physically correct dynamic assumptions, in order to tackle the very difficult problem of estimating 3D shape from an image. Training purely on synthetic data, we are the first to show that garment shape estimation from real images is possible through CNNs. The last and concluding part of the thesis focuses on the problem of inferring a 3D hand pose from an RGB or depth image. To this end, our proposal is an end-to-end CNN system that leverages data from our newly proposed realistically rendered hand dataset, consisting of 3 million samples of hands in various poses, orientations, textures and illuminations. Utilizing this dataset in a supervised training setting, helped us not only with pose inference tasks, but also with hand segmentation. We additionally introduce network components based on differentiable renderers that enabled us to train and refine our networks with unlabeled real images in an unsupervised fashion, showing clear improvements. We demonstrate on-par and improved performance over state-of-the-art methods for two input modalities, under various tasks varying from 3D pose estimation to gesture recognition.

Image sensors allow capturing information-rich data from the environment. They provide frames at high update rate and work under a wide range of illumination. At the same time image sensors are inexpensive, light-weight and consume only little power. A combination with image sensors and a mobile central processing unit as present in today's smartphones, creates a flexible platform for mobile robotics. Inertial sensors that are included in smartphones offer another valuable data source to support any motion and position estimation algorithm. While mobile central processing units feature to record and playback image data at high frame rate and resolution, they are not capable of fully processing image data in real-time. Dedicated hardware components allow to overcome this issue and process image data in real-time at full update rate. Exploiting a hardware-software co-design, fixed latency image processing pipelines can be created for integrating vision data in mobile robot control loops. This thesis presents a multi camera and inertial system based on a central single field-programmable gate array. Image data of up to ten cameras can be processed in real-time while maintaining a low power budget. Global shutter cameras and simultaneous readout guarantee pixel level synchronised image and inertial data. The system is tailored for any mobile robotic application thanks to its light weight and small form factor. All cameras can be arranged individually to create custom configurations. Besides traditional two camera stereo vision also three camera stereo vision with perpendicular baselines or multi baseline stereo vision setups can be created. Perpendicular baseline stereo systems are less affected by matching failures with structures that are parallel to the epipolar line. Multi baseline stereo systems exploit a more accurate depth resolution compared to single baseline systems and provide more robust matches. As a consequence a multi baseline stereo system allows using simpler matching algorithms while providing comparable results to a single baseline setup. Besides multi camera matching algorithms the system also supports custom field of view setups. The ten cameras arranged as five stereo pairs in a pentagon shape create a sensing unit with omnidirectional stereo view. Lenses with a horizontal field of view of 80 degrees are used to create a full horizontal field of view with a decent overlap between neighbouring cameras. Omnidirectional obstacle detection based on stereo vision allows for maneuvers in any direction. Mobile robots like micro aerial vehicles can fully exploit this sensing setup and fly in cluttered environments in any direction while avoiding obstacles. The pentagon camera setup with full horizontal field of view is applicable as data source for visual slam algorithms. The individual cameras with small field of view lenses provide steady resolution image data. In contrast image data of a single camera with a fisheye lens, that also provides a full horizontal field of view, is heavily affected by distortion and non-uniform resolution. The multi camera system adds complexity in the general hardware system setup but enables the usage of simpler lens models.

Multi-robot systems offer several advantages over their single-robot counterpart such as robustness to robot failure and faster exploration in time-critical search and rescue missions. In order to collaborate in these scenarios, the robots need to jointly build a unified map representation where they can co-localize each other. The goal of this thesis is to develop a real-time solution to the Simultaneous Localization and Mapping (SLAM) problem for multiple robots equipped with 3D sensors. In particular, we focus on LiDAR sensors which can be used to generate precise reconstructions of the environment and are robust to changes in lighting conditions. There exist multiple challenges with respect to the multi-robot SLAM problem with 3D point clouds. First, a global place recognition technique is often required as the relative transformations between the robots are not always known. Second, multi- robot systems generate large quantities of data which need to be processed efficiently for achieving real-time performance. Finally, such systems often operate under bandwidth-limited wireless communication channels. A compact representation that can easily be stored and transmitted is therefore required. This thesis specifically targets addressing these three challenges. In our work, we perform global localization using a novel segment extraction and matching algorithm. In essence, 3D point cloud measurements are segmented and each segment is compressed to a compact descriptor. Matching descriptors are retrieved in a map and subsequently filtered based on geometric consistency. The output of this algorithm is a 6 Degrees of Freedom (DoF) pose in a global map, without using prior position information. Globally recognizing places on the basis of segments can be more efficient than using key-point descriptors, as fewer descriptors are usually required to describe places. Additionally, we have developed a set of incremental and time-effective algorithms that exploit the inherent sequential nature of 3D LiDAR measurements. To further address the real-time requirement, we present an ego-motion estimator which attains efficiency by non-uniformly sampling knots over a continuous-time trajectory. A compact map representation is achieved by using a novel data-driven descriptor for 3D point clouds. This descriptor can be extracted by a Convolutional Neural Network (CNN) with an autoencoder-like architecture. The novelty of this approach is that it simultaneously allows us to perform robot localization, 3D environment reconstruction, and semantic extraction. These compact point cloud descriptors can easily be transmitted and used, for example to provide structural feedback to end-users operating in remote locations. We have incorporated all these functionalities in a complete multi-robot SLAM solution that can operate in real-time. The effectiveness of our system has been demonstrated in multiple experiments both in urban driving and search and rescue environments. Specifically, we achieve LiDAR-based global localization at 10Hz in the largest map of the KITTI dataset. Using our mapping approach, a single computer can process in real-time data generated by five Velodyne LiDAR sensors. When retrieving matching descriptors, our data-driven approach leads to an increase of area under the ROC curve of 28.3% over state of the art eigenvalue-based descriptors. Finally, this descriptor allows us to generate dense reconstructions while offering a compression ratio up to 43.5x.

The alternating direction method of multipliers (ADMM) is a first-order optimization algorithm for solving convex problems. We study ADMM under a specialization paradigm, which means that we shape the algorithm to customize it to the optimization problem at hand. This specialization paradigm comes as a contrast to using an ‘off-the-shelf’ or general-purpose solver, which follows a ‘one size fits all’ policy. We show that algorithm specialization makes it possible to synthesize a range of desirable algorithm features and characteristics, which promotes ADMM as a powerful and versatile tool in optimization and control. We study specialized ADMM in a variety of forms and for a range of applications. We first consider an optimal bidding strategy for energy reserve markets. In such markets, guarantees for providing grid-balancing power capacity are traded. We devise an ADMM-driven negotiation protocol that coordinates aggregations of market participants to place a joint energy reserve bid. We show that this algorithmic negotiation equally distributes the computational burden to the participants and that it provides further beneficial features, such as low set-up costs and a large participant autonomy. In a second specialized ADMM formulation, we contribute to the area of autonomous coordination and collision avoidance. We present a decentralized ADMM-based navigation protocol for the coordination of moving agents. A challenge arises from the nonconvex collision avoidance constraints, which we handle with successive linearization at each ADMM iteration. A main feature of the resulting algorithm is that it operates from an agent’s point of view, which means that each agent uses a local coordinate system and interacts with its nearest neighbors only. We show that this fully decentralized perspective provides a range of favorable properties to handle the coordination task. Besides tailor-fitting ADMM for specific areas, we also propose an ADMM specialization framework for a general class of model predictive control (MPC) problems. More specifically, we devise a method that uses the structure in the controlled system as an additional source of efficiency for the optimization routine. The algorithm then mimics the structural properties of the system, which leads to an improved performance of the overall procedure. The resulting ADMM formulations are highly parallelizable and particularly suited for embedded implementation. We also present a novel measure for system structure, called the separation tendency. With the separation tendency, we can decide whether a system is sufficiently structured for utilizing our ADMM framework. In the last part of this thesis, we provide a detailed application study for structure-exploiting ADMM. We show how to model and control the power flow in a variable speed drive (VSD). It turns out that the resulting VSD model is a structured system, which makes controlling the power flow in the VSD an excellent example for algorithm specialization., ISBN:978-3-906916-44-6

How cortical systems self-organise and function is one of the great unresolved questions science currently faces. The combination of billions of neurons, each, by themselves, relatively simple units, can bring forth rich collective dynamics which we witness and experience on a day to day basis. The observation of neuronal avalanches measured in vitro and in vivo in different cortical systems provided the first evidence of scale-free neuronal dynamics, suggesting that the observed complexity in neuronal systems could originate from the same principles which are described by the physics of second order phase transitions in classical statistical physics. The mechanism of self-organised criticality provides an explanation how such scale-invariant dynamics, which reveals itself by the observation of power-law distributed neuronal avalanches, can be observed consistently in cortical systems. The experimentally measured critical exponents are typical for a mean field self-organised branching process. These results have been reproduced by neuronal models based on self-organised criticality which, interestingly, let to the observation of mean-field exponents on different network topologies, including regular lattices, which is not expected for conventional critical systems. This raises the question about the origin of the mean field behaviour observed experimentally. This thesis provides an answer to this open question by investigating the effect of activity-dependent plasticity in combination with the neuronal refractory time in a neuronal network. Our results show that the refractory time hinders backward avalanches forcing a directed propagation of neuronal activity. Hebbian plastic adaptation then plays the role of sculpting these directed avalanche patterns into the topology of the network and slowly changes it into a branched structure where loops are marginal. The observed mean field behaviour of neuronal systems can, therefore, be considered a robust feature of spontaneous neuronal activity. Many additional neurobiological ingredients, such as inhibitory neurons, enrich the dynamics. They play a crucial role in the functioning of cortical systems. The neurotransmitter of inhibitory synapses, such as Gamma-Aminobutyric acid (GABA), reduce the membrane potential in post-synaptic neurons and lower their probability to activate, therefore, reducing the overall activity in the network. The introduction of inhibitory neurons into a self-organised critical neuronal model modifies the dynamics and one can observe that neuronal avalanches do not scale with the system size. Since experimental measurements show scaling avalanches with inhibition, we introduce the concept of dynamical synapses. Dynamical synapses are a form of short-term plasticity and allow the direct modelling of neurotransmitter at the synapses. This provides a neuronal model which features non-conserving dynamics and enables the modelling of synaptic plasticity on two time-scales. The results show that with both short-term and long-term plasticity the neuronal network exhibits scale-free dynamics with inhibitory effects and, furthermore, reproduces the 1/f power spectrum found in many experimental measurements. Finally, we examine the learning capabilities of the neuronal network by studying pattern recognition. Pattern recognition is a vital neuronal process which enables a cortical system to interpret the external world through visual stimuli. Visual recognition tasks have been studied extensively in artificial neural networks. The frequently employed method of back-propagation excels at this task but has been claimed to be biologically implausible in many aspects. How the brain learns to recognize patterns is, therefore, an unsolved problem. The dynamics of visual stimuli differ significantly from spontaneous activity and instead of triggering avalanches by randomly spiking individual neurons, entire patterns of many neurons lead to large avalanches propagating through the system. We address for the first time the pattern recognition task in a fully biological framework with negative feedback signals. These are chemical signals such as dopamine which mediate synaptic plasticity and sculpt the network to achieve certain tasks. The system can distinguish patterns such as horizontal and vertical lines with high accuracy as well as to perform well at the more complicated task of handwritten digit recognition. Resulting from the learning mechanism, spatially separate activity regions emerge, as are observed experimentally in the visual cortex where different visual stimuli activate distinct neuronal columns. The results, therefore, suggest that negative feedback signals offer an explanation for the emergence of these distinct activity areas.