Die Innenraumpositionierung kommt in der heutigen Zeit immer mehr Aufmerksamkeit zu teil. Neben der Navigation durch das Gebäude sind vor allem Location Based Services von Bedeutung, welche Zusatzinformationen zu spezifischen Objekten zur Verfügung stellen Da für eine Innenraumortung das GPS Signal jedoch zu schwach ist, müssen andere Techniken zur Lokalisierung gefunden werden. Neben der häufig verwendeten Positionierung durch Auswertung von empfangenen Funkwellen existieren Methoden zur optischen Lokalisierung mittels Landmarken. Das kamerabasierte Verfahren bietet den Vorteil, dass eine oft zentimetergenaue Positionierung möglich ist. In dieser Masterarbeit erfolgt die Bestimmung der Position im Gebäude mittels Detektion von ArUco-Markern und Türschildern aus Bilddaten. Als Evaluationsgeräte sind zum einen die Kinect v2 von Microsoft, als auch das Lenovo Phab 2 Pro Smartphone verwendet worden. Neben den Bilddaten stellen diese auch mittels Time of Flight Sensoren generierte Tiefendaten zur Verfügung. Durch den Vergleich von aus dem Bild extrahierten Eckpunkten der Landmarke, mit den aus einer Datenbank entnommenen realen geometrischen Maßen des Objektes, kann die Entfernung zu einer gefundenen Landmarke bestimmt werden. Neben der optischen Distanzermittlung wird die Position zusätzlich anhand der Tiefendaten ermittelt. Abschließend werden beiden Verfahren miteinander verglichen und eine Aussage bezüglich der Genauigkeit und Zuverlässigkeit des in dieser Arbeit entwickelten Algorithmus getroffen. / Indoor Positioning is receiving more and more attention nowadays. Beside the navigation through a building, Location Bases Services offer the possibility to get more information about certain objects in the enviroment. Because GPS signals are too weak to penetrate buildings, other techniques for localization must be found. Beneath the commonly used positioning via the evaluation of received radio signals, optical methods for localization with the help of landmarks can be used. These camera-based procedures have the advantage, that an inch-perfect positioning is possible. In this master thesis, the determination of the position in a building is chieved through the detection of ArUco-Marker and door signs in images gathered by a camera. The evaluation is done with the Microsoft Kinect v2 and the Lenovo Phab 2 Pro Smartphone. They offer depth data gained by a time of flight sensor beside the color images. The range to a detected landmark is calculated by comparing the object´s corners in the image with the real metrics, extracted from a database. Additionally, the distance is determined by the evaluation of the depth data. Finally, both procedures are compared with each other and a statement about the accuracy and responsibility is made.

Visualization is used for stainless steel container wall and lid cross section characterization. Two specific types of containers are examined: 3013 and SAVY. The container wall examined is from a sample of the inner container of a 3013 container. The inner lid cross section examined is from a SAVY container. Laser confocal microscope data and photographic data are used to determine features of the surfaces. The surface features are then characterized by various feature statistics, such as, maximum depth, area, eccentricity, and others. The purpose of this pilot study is to demonstrate the effectiveness of using the methodology to detect potential corrosion events on the inner container surfaces. The features are used to quantify these corrosion events. An automatic image analysis system uses this methodology to classify images for possible further human analysis by flagging possible corrosion events. A manual image analysis methodology is used to determine the amount of MnS on the SAVY container lid cross section. Visualization is an integral component of the analysis methodology.

Understanding the processes of cellular development and the interplay of cell shape changes, division and migration requires investigation of developmental processes at the spatial resolution of single cell. Biomedical imaging experiments enable the study of dynamic processes as they occur in living organisms. While biomedical imaging is essential, a key component of exposing unknown biological phenomena is quantitative image analysis. Biomedical images, especially microscopy images, are usually noisy owing to practical limitations such as available photon budget, sample sensitivity, etc. Additionally, microscopy images often contain artefacts due to the optical aberrations in microscopes or due to imperfections in camera sensor and internal electronics. The noisy nature of images as well as the artefacts prohibit accurate downstream analysis such as cell segmentation. Although countless approaches have been proposed for image denoising, artefact removal and segmentation, supervised Deep Learning (DL) based content-aware algorithms are currently the best performing for all these tasks. Supervised DL based methods are plagued by many practical limitations. Supervised denoising and artefact removal algorithms require paired corrupted and high quality images for training. Obtaining such image pairs can be very hard and virtually impossible in most biomedical imaging applications owing to photosensitivity and the dynamic nature of the samples being imaged. Similarly, supervised DL based segmentation methods need copious amounts of annotated data for training, which is often very expensive to obtain. Owing to these restrictions, it is imperative to look beyond supervised methods. The objective of this thesis is to develop novel unsupervised alternatives for image denoising, and artefact removal as well as semisupervised approaches for image segmentation. The first part of this thesis deals with unsupervised image denoising and artefact removal. For unsupervised image denoising task, this thesis first introduces a probabilistic approach for training DL based methods using parametric models of imaging noise. Next, a novel unsupervised diversity denoising framework is presented which addresses the fundamentally non-unique inverse nature of image denoising by generating multiple plausible denoised solutions for any given noisy image. Finally, interesting properties of the diversity denoising methods are presented which make them suitable for unsupervised spatial artefact removal in microscopy and medical imaging applications. In the second part of this thesis, the problem of cell/nucleus segmentation is addressed. The focus is especially on practical scenarios where ground truth annotations for training DL based segmentation methods are scarcely available. Unsupervised denoising is used as an aid to improve segmentation performance in the presence of limited annotations. Several training strategies are presented in this work to leverage the representations learned by unsupervised denoising networks to enable better cell/nucleus segmentation in microscopy data. Apart from DL based segmentation methods, a proof-of-concept is introduced which views cell/nucleus segmentation from the perspective of solving a label fusion problem. This method, through limited human interaction, learns to choose the best possible segmentation for each cell/nucleus using only a pool of diverse (and possibly faulty) segmentation hypotheses as input. In summary, this thesis seeks to introduce new unsupervised denoising and artefact removal methods as well as semi-supervised segmentation methods which can be easily deployed to directly and immediately benefit biomedical practitioners with their research.

Reaction-diffusion models have been widely successful in explaining a large variety of patterning phenomena in biology ranging from embryonic development to cancer growth and angiogenesis. Firstly proposed by Alan Turing in 1952 and applied to a simple two-component system, reaction-diffusion models describe spontaneous spatial pattern formation, driven purely by interactions of the system components and their diffusion in space. Today, access to unprecedented amounts of quantitative biological data allows us to build and test biochemically accurate reaction-diffusion models of intracellular processes. However, any increase in model complexity increases the number of unknown parameters and thus the computational cost of model analysis. To efficiently characterize the behavior and robustness of models with many unknown parameters is, therefore, a key challenge in systems biology. Here, we propose a novel computational framework for efficient high-dimensional parameter space characterization of reaction-diffusion models. The method leverages the $L_p$-Adaptation algorithm, an adaptive-proposal statistical method for approximate high-dimensional design centering and robustness estimation. Our approach is based on an oracle function, which describes for each point in parameter space whether the corresponding model fulfills given specifications. We propose specific oracles to estimate four parameter-space characteristics: bistability, instability, capability of spontaneous pattern formation, and capability of pattern maintenance. We benchmark the method and demonstrate that it allows exploring the ability of a model to undergo pattern-forming instabilities and to quantify model robustness for model selection in polynomial time with dimensionality. We present an application of the framework to reconstituted membrane domains bearing the small GTPase Rab5 and propose molecular mechanisms that potentially drive pattern formation.

Computer vision in the area of medical imaging has rapidly improved during recent years as a consequence of developments in deep learning and explainability algorithms. In addition, imaging in nuclear medicine is becoming increasingly sophisticated, with the emergence of targeted radiotherapies that enable treatment and imaging on a molecular level (“theranostics”) where radiolabeled targeted molecules are directly injected into the bloodstream. Based on our recent work, we present two use-cases in nuclear medicine as follows: first, the impact of automated organ segmentation required for personalized dosimetry in patients with neuroendocrine tumors and second, purely data-driven identification and verification of brain regions for diagnosis of Parkinson’s disease. Convolutional neural network was used for automated organ segmentation on computed tomography images. The segmented organs were used for calculation of the energy deposited into the organ-at-risk for patients treated with a radiopharmaceutical. Our method resulted in faster and cheaper dosimetry and only differed by 7% from dosimetry performed by two medical physicists. The identification of brain regions, however was analyzed on dopamine-transporter single positron emission tomography images using convolutional neural network and explainability, i.e., layer-wise relevance propagation algorithm. Our findings confirm that the extra-striatal brain regions, i.e., insula, amygdala, ventromedial prefrontal cortex, thalamus, anterior temporal cortex, superior frontal lobe, and pons contribute to the interpretation of images beyond the striatal regions. In current common diagnostic practice, however, only the striatum is the reference region, while extra-striatal regions are neglected. We further demonstrate that deep learning-based diagnosis combined with explainability algorithm can be recommended to support interpretation of this image modality in clinical routine for parkinsonian syndromes, with a total computation time of three seconds which is compatible with busy clinical workflow. Overall, this thesis shows for the first time that deep learning with explainability can achieve results competitive with human performance and generate novel hypotheses, thus paving the way towards improved diagnosis and treatment in nuclear medicine.

Recent advancements in the field of Computer Vision are a by-product of breakthroughs in the domain of Artificial Intelligence. Object detection in monocular images is now realized by an amalgamation of Computer Vision and Deep Learning. While most approaches detect objects as a mere two dimensional (2D) bounding box, there are a few that exploit rather traditional representation of the 3D object. Such approaches detect an object either as a 3D bounding box or exploit its shape primitives using active shape models which results in a wireframe-like detection. Such a wireframe detection is represented as combinations of detected keypoints (or landmarks) of the desired object. Apart from a faithful retrieval of the object’s true shape, wireframe based approaches are relatively robust in handling occlusions. The central task of this thesis was to find such an approach and to implement it with the goal of its performance evaluation. The object of interest is the vehicle class (cars, mini vans, trucks etc.) and the evaluation data is monocular traffic surveillance videos collected by the supervising chair. A wireframe type detection can aid several facets of traffic analysis by improved (compared to 2D bounding box) estimation of the detected object’s ground plane. The thesis encompasses the process of implementation of the chosen approach called Occlusion-Net [40], including its design details and a qualitative evaluation on traffic surveillance videos. The implementation reproduces most of the published results across several occlusion categories except the truncated car category. Occlusion-Net’s erratic detections are mostly caused by incorrect detection of the initial region of interest. It employs three instances of Graph Neural Networks for occlusion reasoning and localization. The thesis also provides a didactic introduction to the field of Machine and Deep Learning including intuitions of mathematical concepts required to understand the two disciplines and the implemented approach.:Contents 1 Introduction 1 2 Technical Background 7 2.1 AI, Machine Learning and Deep Learning 7 2.1.1 But what is AI ? 7 2.1.2 Representational composition by Deep Learning 10 2.2 Essential Mathematics for ML 14 2.2.1 Linear Algebra 15 2.2.2 Probability and Statistics 25 2.2.3 Calculus 34 2.3 Mathematical Introduction to ML 39 2.3.1 Ingredients of a Machine Learning Problem 39 2.3.2 The Perceptron 40 2.3.3 Feature Transformation 46 2.3.4 Logistic Regression 48 2.3.5 Artificial Neural Networks: ANN 53 2.3.6 Convolutional Neural Network: CNN 61 2.3.7 Graph Neural Networks 68 2.4 Specific Topics in Computer Vision 72 2.5 Previous work 76 3 Design of Implemented Approach 81 3.1 Training Dataset 81 3.2 Keypoint Detection : MaskRCNN 83 3.3 Occluded Edge Prediction : 2D-KGNN Encoder 84 3.4 Occluded Keypoint Localization : 2D-KGNN Decoder 86 3.5 3D Shape Estimation: 3D-KGNN Encoder 88 4 Implementation 93 4.1 Open-Source Tools and Libraries 93 4.1.1 Code Packaging: NVIDIA-Docker 94 4.1.2 Data Processing Libraries 94 4.1.3 Libraries for Neural Networks 95 4.1.4 Computer Vision Library 95 4.2 Dataset Acquisition and Training 96 4.2.1 Acquiring Dataset 96 4.2.2 Training Occlusion-Net 96 4.3 Refactoring 97 4.3.1 Error in Docker File 97 4.3.2 Image Directories as Input 97 4.3.3 Frame Extraction in Parallel 98 4.3.4 Video as Input 100 4.4 Functional changes 100 4.4.1 Keypoints In Output 100 4.4.2 Mismatched BB and Keypoints 101 4.4.3 Incorrect Class Labels 101 4.4.4 Bounding Box Overlay 101 5 Evaluation 103 5.1 Qualitative Evaluation 103 5.1.1 Evaluation Across Occlusion Categories 103 5.1.2 Performance on Moderate and Heavy Vehicles 105 5.2 Verification of Failure Analysis 106 5.2.1 Truncated Cars 107 5.2.2 Overlapping Cars 108 5.3 Analysis of Missing Frames 109 5.4 Test Performance 110 6 Conclusion 113 7 Future Work 117 Bibliography 119

Progressive download video services, such as YouTube and podcasts, are responsible for a major part of the transmitted data volume in the Internet and it is expected, that they will also strongly affect mobile networks. Streaming video quality mainly depends on the sustainable throughput achieved during transmission. To ensure acceptable video quality in mobile networks (with limited capacity resources) the perceived quality by the customer (QoE) needs to be monitored by estimation. For that, the streaming video quality needs to be measured and monitored permanently. For TCP based progressive download we propose to extract the the video timestamps which are encoded within the payload of the TCP segments by decoding the video within the payload. The actual estimation is then done by play out buffer fill level calculations based on the TCP segment timestamp and their internal play out timestamp. The perceived quality for the user is derived from the number and duration of video stalls. Algorithms for decoding Flash Video, MP4 and WebM Video have already been implemented. After deriving the play out time it is compared to the timestamp of the respective TCP segment. The result of this comparison is an estimate of the fill level of the play out buffer in terms of play out time within the client. This estimation is done without access to the end device. The same measurement procedure can be applied for any TCP based progressive download Internet service. Video was simply taken as an example because of its current large share in traffic volume in operator networks.

The complexity of computer networks has been rising over the last decades. Increasing interconnectivity between multiple devices, growing complexity of performed tasks and a strong collaboration between nodes are drivers for this phenomenon. An example is represented by Internet-of-Things devices, whose relevance has been rising in recent years. The increasing number of devices requiring updates and supervision makes maintenance more difficult. Human interaction, in this case, is costly and requires a lot of time. To overcome this, self-adaptive software systems (SAS) can be used. SAS are a subset of autonomous systems which can monitor themselves and their environment to adapt to changes without human interaction. In the literature, different approaches for engineering SAS were proposed, including techniques for executing adaptations on multiple devices based on generated plans for reacting to changes. Among those solutions, also decentralised approaches can be found. To the best of our knowledge, no approach for engineering a SAS exists which tolerates errors during the execution of adaptation in a decentralised setting. While some approaches for role-based execution reset the application in case of a single failure during the adaptation process, others do not make assumptions about errors or do not consider an erroneous environment. In a real-world environment, errors will likely occur during run-time, and the adaptation process could be disturbed. This work aims to perform adaptations in a decentralised way on role-based systems with a relaxed consistency constraint, i.e., errors during the adaptation phase are tolerated. This increases the availability of nodes since no rollbacks are required in case of a failure. Moreover, a subset of applications, such as drone swarms, would benefit from an approach with a relaxed consistency model since parts of the system that adapted successfully can already operate in an adapted configuration instead of waiting for other peers to apply the changes in a later iteration. Moreover, if we eliminate the need for an atomic adaptation execution, asynchronous execution of adaptation would be possible. In that case, we can supervise the adaptation process for a long time and ensure that every peer takes the planned actions as soon as the internal task execution allows it. To allow for a relaxed consistent way of adaptation execution, we develop a decentralised adaptation execution protocol, which supports the notion of eventual consistency. As soon as devices reconnect after network congestion or restore their internal state after local failures, our protocol can coordinate the recovery process among multiple devices to attempt recovery of a globally consistent state after errors occur. By superseding the need for a central instance, every peer who received information about failing peers can start the recovery process. The developed approach can restore a consistent global configuration if almost all peers fail. Moreover, the approach supports asynchronous adaptations, i.e., the peers can execute planned adaptations as soon as they are ready, which increases overall availability in case of delayed adaptation of single nodes. The developed protocol is evaluated with the help of a proof-of-concept implementation. The approach was run in five different experiments with thousands of iterations to show the applicability and reliability of this novel approach. The time for execution of the protocol and the number of exchanged messages has been measured to compare the protocol for different error cases and system sizes, as well as to show the scalability of the approach. The developed solution has been compared to a blocking approach to show the feasibility compared to an atomic approach. The applicability in a real-world scenario has been described in an empirical study using an example of a fire-extinguishing drone swarm. The results show that an optimistic approach to adaptation is suitable and specific scenarios can benefit from the improved availability since no rollbacks are required. Systems can continue their work regardless of the failures of participating nodes in large-scale systems.:Abstract VI 1. Introduction 1 1.1. Motivational Use-Case 2 1.2. Problem Definition 3 1.3. Objectives 4 1.4. Research Questions 5 1.5. Contributions 5 1.6. Outline 6 2. Foundation 7 2.1. Role Concept 7 2.2. Self-Adaptive Software Systems 13 2.3. Terminology for Role-Based Self-Adaptation 15 2.4. Consistency Preservation and Consistency Models 17 2.5. Summary 20 3. Related Work 21 3.1. Role-Based Approaches 22 3.2. Actor Model of Computation and Akka 23 3.3. Adaptation Execution in Self-Adaptive Software Systems 24 3.4. Change Consistency in Distributed Systems 33 3.5. Comparison of the Evaluated Approaches 40 4. The Decentralised Consistency Compensation Protocol 43 4.1. System and Error Model 43 4.2. Requirements to the Concept 44 4.3. The Usage of Roles in Adaptations 45 4.4. Protocol Overview 47 4.5. Protocol Description 51 4.6. Protocol Corner- and Error Cases 64 4.7. Summary 66 5. Prototypical Implementation 67 5.1. Technology Overview 67 5.2. Reused Artifacts 68 5.3. Implementation Details 70 5.4. Setup of the Prototypical Implementation 76 5.5. Summary 77 6. Evaluation 79 6.1. Evaluation Methodology 79 6.2. Evaluation Setup 80 6.3. Experiment Overview 81 6.4. Default Case: Successful Adaptation 84 6.5. Compensation on Disconnection of Peers 85 6.6. Recovery from Failed Adaptation 88 6.7. Impact of Early Activation of Adaptations 91 6.8. Comparison with a Blocking Approach 92 6.9. Empirical Study: Fire Extinguishing Drones 95 6.10. Summary 97 7. Conclusion and Future Work 99 7.1. Recap of the Research Questions 99 7.2. Discussion 101 7.3. Future Work 101 A. Protocol Buffer Definition 103 Acronyms 108 Bibliography 109

Ein üblicher Ansatzpunkt zur Verbesserung der Energieeffizienz im High Performance Computing (HPC) ist, neben Verbesserungen an der Hardware oder einer effizienteren Nachnutzung der Wärme des Systems, die Optimierung der ausgeführten Programme. Dazu können zum Beispiel energieoptimale Einstellungen, wie die Frequenzen des Prozessors, für verschiedene Programmfunktionen bestimmt werden, um diese dann im späteren Verlauf des Programmes anwenden zu können. Mit jeder Änderung des Programmes kann sich dessen optimale Einstellung ändern, weshalb diese zeitaufwendig neu bestimmt werden muss. Das stellt eine wesentliche Hürde für die Anwendung solcher Verfahren dar. Dieser Prozess des Bestimmens der optimalen Frequenzen kann mithilfe von Machine-Learning-Methoden vereinfacht werden, wie in dieser Arbeit gezeigt wird. So lässt sich mithilfe von sogenannten Performance-Events ein neuronales Netz erstellen, mit dem während der Ausführung des Programmes die optimalen Frequenzen automatisch geschätzt werden können. Performance-Events sind prozessorintern und können Einblick in die Abläufe im Prozessor gewähren. Bei dem Einsatz von Performance-Events gilt es einige Fallstricke zu vermeiden. So werden die Performance-Events von Performance-Countern gezählt. Die Anzahl der Counter ist allerdings begrenzt, womit auch die Anzahl der Events, die gleichzeitig gezählt werden können, limitiert ist. Eine für diese Arbeit wesentliche Fragestellung ist also: Welche dieser Events sind relevant und müssen gezählt werden? Bei der Beantwortung dieser Frage sind Merkmalsauswahlverfahren hilfreich, besonders sogenannte Filtermethoden, bei denen die Merkmale vor dem Training ausgewählt werden. Viele bekannte Methoden gehen dabei entweder davon aus, dass die Zusammenhänge zwischen den Merkmalen linear sind, wie z. B. bei Verfahren, die den Pearson-Korrelationskoeffizienten verwenden, oder die Daten müssen in Klassen eingeteilt werden, wie etwa bei Verfahren, die auf der Transinformation beruhen. Beides ist für Performance-Events nicht ideal. Auf der einen Seite können keine linearen Zusammenhänge angenommen werden. Auf der anderen Seite bedeutet das Einteilen in Klassen einen Verlust an Information. Um diese Probleme zu adressieren, werden in dieser Arbeit bestehende Merkmalsauswahlverfahren mit den dazugehörigen Algorithmen analysiert, neue Verfahren entworfen und miteinander verglichen. Es zeigt sich, dass mit neuen Verfahren, die auf sogenannten Copulas basieren, auch nichtlineare Zusammenhänge erkannt werden können, ohne dass die Daten in Klassen eingeteilt werden müssen. So lassen sich schließlich einige Events identifiziert, die zusammen mit neuronalen Netzen genutzt werden können, um die Energieeffizienz von HPC-Anwendung zu steigern. Das in dieser Arbeit erstellte Framework erfüllt dabei neben der Auswahl der Performance-Events weitere Aufgaben: Es stellt sicher, dass diverse Programmteile mit verschiedenen optimalen Einstellungen voneinander unterschieden werden können. Darüber hinaus sorgt das Framework dafür, dass genügend Daten erzeugt werden, um ein neuronales Netz zu trainieren, und dass dieses Netz später einfach genutzt werden kann. Dabei ist das Framework so flexibel, dass auch andere Machine-Learning-Methoden getestet werden können. Die Leistungsfähigkeit des Frameworks wird abschließend in einer Ende-zu-Ende-Evaluierung an einem beispielhaften Programm demonstriert. Die Evaluierung il­lus­t­riert, dass bei nur 7% längerer Laufzeit eine Energieeinsparung von 24% erzielt werden kann und zeigt damit, dass mit Machine-Learning-Methoden wesentliche Energieeinsparungen erreicht werden können.:1 Einleitung und Motiovation 2 Energieeffizienz und Machine-Learning – eine thematische Einführung 2.1 Energieeffizienz von Programmen im Hochleistungsrechnen 2.1.1 Techniken zur Energiemessung oder -abschätzung 2.1.2 Techniken zur Beeinflussung der Energieeffizienz in der Hardware 2.1.3 Grundlagen zur Performanceanalyse 2.1.4 Regionsbasierte Ansätze zur Erhöhung der Energieeffizienz 2.1.5 Andere Ansätze zur Erhöhung der Energieeffizienz 2.2 Methoden zur Merkmalsauswahl 2.2.1 Merkmalsauswahlmethoden basierend auf der Informationstheorie 2.2.2 Merkmalsauswahl für stetige Merkmale 2.2.3 Andere Verfahren zur Merkmalsauswahl 2.3 Machine-Learning mit neuronalen Netzen 2.3.1 Neuronale Netze 2.3.2 Backpropagation 2.3.3 Aktivierungsfunktionen 3 Merkmalsauswahl für mehrdimensionale nichtlineare Abhängigkeiten 3.1 Analyse der Problemstellung, Merkmale und Zielgröße 3.2 Merkmalsauswahl mit mehrdimensionaler Transinformation für stetige Merkmale 3.2.1 Mehrdimensionale Copula-Entropie und mehrdimensionale Transinformation 3.2.2 Schätzung der mehrdimensionalen Transinformation basierend auf Copula-Dichte 3.3 Normierung 3.4 Vergleich von Copula-basierten Maßzahlen mit der klassischen Transinformation und dem Pearson-Korrelationskoeffizienten 3.4.1 Deterministische Abhängigkeit zweier Variablen 3.4.2 UnabhängigkeitVergleich verschiedener Methoden zur Auswahl stetiger Merkmale 3.5 Vergleich verschiedener Methoden zur Auswahl stetiger Merkmale 3.5.1 Erzeugung synthetischer Daten 3.5.2 Szenario 1 – fünf relevante Merkmale 3.5.3 Szenario 2 – fünf relevante Merkmale, fünf wiederholte Merkmale 3.5.4 Schlussfolgerungen aus den Simulationen 3.6 Zusammenfassung 4 Entwicklung und Umsetzung des Frameworks 4.1 Erweiterungen der READEX Runtime Library 4.1.1 Grundlegender Aufbau der READEX Runtime Library 4.1.2 Call-Path oder Call-Tree 4.1.3 Calibration-Module 4.2 Testsystem 4.2.1 Architektur 4.2.2 Bestimmung des Offsets zur Energiemessung mit RAPL 4.3 Verwendete Benchmarks zur Erzeugung der Datengrundlage 4.3.1 Datensatz 1: Der Stream-Benchmark 4.3.2 Datensatz 2: Eine Sammlung verschiedener Benchmarks 4.4 Merkmalsauswahl und Modellgenerierung 4.4.1 Datenaufbereitung 4.4.2 Merkmalsauswahl Algorithmus 4.4.3 Performance-Events anderer Arbeiten zum Vergleich 4.4.4 Erzeugen und Validieren eines Modells mithilfe von TensorFlow und Keras 4.5 Zusammenfassung 5 Evaluierung des Ansatzes 5.1 Der Stream-Benchmark 5.1.1 Analyse der gewählten Merkmale 5.1.2 Ergebnisse des Trainings 5.2 Verschiedene Benchmarks 5.2.1 Ausgewählte Merkmale 5.2.2 Ergebnisse des Trainings 5.3 Energieoptimierung einer Anwendung 6 Zusammenfassung und Ausblick Literatur Abbildungsverzeichnis Tabellenverzeichnis Quelltextverzeichnis / There are a variety of different approaches to improve energy efficiency in High Performance Computing (HPC). Besides advances to the hardware or cooling systems, optimising the executed programmes' energy efficiency is another a promising approach. Determining energy-optimal settings of program functions, such as the processor frequency, can be applied during the program's execution to reduce energy consumption. However, when the program is modified, the optimal setting might change. Therefore, the energy-optimal settings need to be determined again, which is a time-consuming process and a significant impediment for applying such methods. Fortunately, finding the optimal frequencies can be simplified using machine learning methods, as shown in this thesis. With the help of so-called performance events, a neural network can be trained, which can automatically estimate the optimal processor frequencies during program execution. Performance events are processor-specific and can provide insight into the procedures of a processor. However, there are some pitfalls to be avoided when using performance events. Performance events are counted by performance counters, but as the number of counters is limited, the number of events that can be counted simultaneously is also limited. This poses the question of which of these events are relevant and need to be counted. % Though the issue has received some attention in several publications, a convincing solution remains to be found. In answering this question, feature selection methods are helpful, especially so-called filter methods, where features are selected before the training. Unfortunately, many feature selection methods either assume a linear correlation between the features, such as methods using the Pearson correlation coefficient or require data split into classes, particularly methods based on mutual information. Neither can be applied to performance events as linear correlation cannot be assumed, and splitting the data into classes would result in a loss of information. In order to address that problem, this thesis analyses existing feature selection methods together with their corresponding algorithms, designs new methods, and compares different feature selection methods. By utilising new methods based on the mathematical concept of copulas, it was possible to detect non-linear correlations without splitting the data into classes. Thus, several performance events could be identified, which can be utilised together with neural networks to increase the energy efficiency of HPC applications. In addition to selecting performance events, the created framework ensures that different programme parts, which might have different optimal settings, can be identified. Moreover, it assures that sufficient data for the training of the neural networks is generated and that the network can easily be applied. Furthermore, the framework is flexible enough to evaluate other machine learning methods. Finally, an end-to-end evaluation with a sample application demonstrated the framework's performance. The evaluation illustrates that, while extending the runtime by only 7%, energy savings of 24% can be achieved, showing that substantial energy savings can be attained using machine learning approaches.:1 Einleitung und Motiovation 2 Energieeffizienz und Machine-Learning – eine thematische Einführung 2.1 Energieeffizienz von Programmen im Hochleistungsrechnen 2.1.1 Techniken zur Energiemessung oder -abschätzung 2.1.2 Techniken zur Beeinflussung der Energieeffizienz in der Hardware 2.1.3 Grundlagen zur Performanceanalyse 2.1.4 Regionsbasierte Ansätze zur Erhöhung der Energieeffizienz 2.1.5 Andere Ansätze zur Erhöhung der Energieeffizienz 2.2 Methoden zur Merkmalsauswahl 2.2.1 Merkmalsauswahlmethoden basierend auf der Informationstheorie 2.2.2 Merkmalsauswahl für stetige Merkmale 2.2.3 Andere Verfahren zur Merkmalsauswahl 2.3 Machine-Learning mit neuronalen Netzen 2.3.1 Neuronale Netze 2.3.2 Backpropagation 2.3.3 Aktivierungsfunktionen 3 Merkmalsauswahl für mehrdimensionale nichtlineare Abhängigkeiten 3.1 Analyse der Problemstellung, Merkmale und Zielgröße 3.2 Merkmalsauswahl mit mehrdimensionaler Transinformation für stetige Merkmale 3.2.1 Mehrdimensionale Copula-Entropie und mehrdimensionale Transinformation 3.2.2 Schätzung der mehrdimensionalen Transinformation basierend auf Copula-Dichte 3.3 Normierung 3.4 Vergleich von Copula-basierten Maßzahlen mit der klassischen Transinformation und dem Pearson-Korrelationskoeffizienten 3.4.1 Deterministische Abhängigkeit zweier Variablen 3.4.2 UnabhängigkeitVergleich verschiedener Methoden zur Auswahl stetiger Merkmale 3.5 Vergleich verschiedener Methoden zur Auswahl stetiger Merkmale 3.5.1 Erzeugung synthetischer Daten 3.5.2 Szenario 1 – fünf relevante Merkmale 3.5.3 Szenario 2 – fünf relevante Merkmale, fünf wiederholte Merkmale 3.5.4 Schlussfolgerungen aus den Simulationen 3.6 Zusammenfassung 4 Entwicklung und Umsetzung des Frameworks 4.1 Erweiterungen der READEX Runtime Library 4.1.1 Grundlegender Aufbau der READEX Runtime Library 4.1.2 Call-Path oder Call-Tree 4.1.3 Calibration-Module 4.2 Testsystem 4.2.1 Architektur 4.2.2 Bestimmung des Offsets zur Energiemessung mit RAPL 4.3 Verwendete Benchmarks zur Erzeugung der Datengrundlage 4.3.1 Datensatz 1: Der Stream-Benchmark 4.3.2 Datensatz 2: Eine Sammlung verschiedener Benchmarks 4.4 Merkmalsauswahl und Modellgenerierung 4.4.1 Datenaufbereitung 4.4.2 Merkmalsauswahl Algorithmus 4.4.3 Performance-Events anderer Arbeiten zum Vergleich 4.4.4 Erzeugen und Validieren eines Modells mithilfe von TensorFlow und Keras 4.5 Zusammenfassung 5 Evaluierung des Ansatzes 5.1 Der Stream-Benchmark 5.1.1 Analyse der gewählten Merkmale 5.1.2 Ergebnisse des Trainings 5.2 Verschiedene Benchmarks 5.2.1 Ausgewählte Merkmale 5.2.2 Ergebnisse des Trainings 5.3 Energieoptimierung einer Anwendung 6 Zusammenfassung und Ausblick Literatur Abbildungsverzeichnis Tabellenverzeichnis Quelltextverzeichnis

In this thesis I will use state-of-the-art (SOTA) image denoising methods to denoise electron microscopy (EM) data. Then, I will present NoiseVoid a deep learning based self-supervised image denoising approach which is trained on single noisy observations. Eventually, I approach the missing wedge problem in tomography and introduce a novel image encoding, based on the Fourier transform which I am using to predict missing Fourier coefficients directly in Fourier space with Fourier Image Transformer (FIT). In the next paragraphs I will summarize the individual contributions briefly. Electron microscopy is the go to method for high-resolution images in biological research. Modern scanning electron microscopy (SEM) setups are used to obtain neural connectivity maps, allowing us to identify individual synapses. However, slow scanning speeds are required to obtain SEM images of sufficient quality. In (Weigert et al. 2018) the authors show, for fluorescence microscopy, how pairs of low- and high-quality images can be obtained from biological samples and use them to train content-aware image restoration (CARE) networks. Once such a network is trained, it can be applied to noisy data to restore high quality images. With SEM-CARE I present how this approach can be directly applied to SEM data, allowing us to scan the samples faster, resulting in $40$- to $50$-fold imaging speedups for SEM imaging. In structural biology cryo transmission electron microscopy (cryo TEM) is used to resolve protein structures and describe molecular interactions. However, missing contrast agents as well as beam induced sample damage (Knapek and Dubochet 1980) prevent acquisition of high quality projection images. Hence, reconstructed tomograms suffer from low signal-to-noise ratio (SNR) and low contrast, which makes post-processing of such data difficult and often has to be done manually. To facilitate down stream analysis and manual data browsing of cryo tomograms I present cryoCARE a Noise2Noise (Lehtinen et al. 2018) based denoising method which is able to restore high contrast, low noise tomograms from sparse-view low-dose tilt-series. An implementation of cryoCARE is publicly available as Scipion (de la Rosa-Trevín et al. 2016) plugin. Next, I will discuss the problem of self-supervised image denoising. With cryoCARE I exploited the fact that modern cryo TEM cameras acquire multiple low-dose images, hence the Noise2Noise (Lehtinen et al. 2018) training paradigm can be applied. However, acquiring multiple noisy observations is not always possible e.g. in live imaging, with old cryo TEM cameras or simply by lack of access to the used imaging system. In such cases we have to fall back to self-supervised denoising methods and with Noise2Void I present the first self-supervised neural network based image denoising approach. Noise2Void is also available as an open-source Python package and as a one-click solution in Fiji (Schindelin et al. 2012). In the last part of this thesis I present Fourier Image Transformer (FIT) a novel approach to image reconstruction with Transformer networks. I develop a novel 1D image encoding based on the Fourier transform where each prefix encodes the whole image at reduced resolution, which I call Fourier Domain Encoding (FDE). I use FIT with FDEs and present proof of concept for super-resolution and tomographic reconstruction with missing wedge correction. The missing wedge artefacts in tomographic imaging originate in sparse-view imaging. Sparse-view imaging is used to keep the total exposure of the imaged sample to a minimum, by only acquiring a limited number of projection images. However, tomographic reconstructions from sparse-view acquisitions are affected by missing wedge artefacts, characterized by missing wedges in the Fourier space and visible as streaking artefacts in real image space. I show that FITs can be applied to tomographic reconstruction and that they fill in missing Fourier coefficients. Hence, FIT for tomographic reconstruction solves the missing wedge problem at its source.:Contents Summary iii Acknowledgements v 1 Introduction 1 1.1 Scanning Electron Microscopy . . . . . . . . . . . . . . . . . . . . 3 1.2 Cryo Transmission Electron Microscopy . . . . . . . . . . . . . . . 4 1.2.1 Single Particle Analysis . . . . . . . . . . . . . . . . . . . . 5 1.2.2 Cryo Tomography . . . . . . . . . . . . . . . . . . . . . . . 7 1.3 Tomographic Reconstruction . . . . . . . . . . . . . . . . . . . . . 8 1.4 Overview and Contributions . . . . . . . . . . . . . . . . . . . . . 11 2 Denoising in Electron Microscopy 15 2.1 Image Denoising . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.2 Supervised Image Restoration . . . . . . . . . . . . . . . . . . . . 19 2.2.1 Training and Validation Loss . . . . . . . . . . . . . . . . 19 2.2.2 Neural Network Architectures . . . . . . . . . . . . . . . . 21 2.3 SEM-CARE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.3.1 SEM-CARE Experiments . . . . . . . . . . . . . . . . . . 23 2.3.2 SEM-CARE Results . . . . . . . . . . . . . . . . . . . . . 25 2.4 Noise2Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 2.5 cryoCARE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 2.5.1 Restoration of cryo TEM Projections . . . . . . . . . . . . 27 2.5.2 Restoration of cryo TEM Tomograms . . . . . . . . . . . . 29 2.5.3 Automated Downstream Analysis . . . . . . . . . . . . . . 31 2.6 Implementations and Availability . . . . . . . . . . . . . . . . . . 32 2.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 2.7.1 Tasks Facilitated through cryoCARE . . . . . . . . . . . 33 3 Noise2Void: Self-Supervised Denoising 35 3.1 Probabilistic Image Formation . . . . . . . . . . . . . . . . . . . . 37 3.2 Receptive Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 3.3 Noise2Void Training . . . . . . . . . . . . . . . . . . . . . . . . . 39 3.3.1 Implementation Details . . . . . . . . . . . . . . . . . . . . 41 3.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 3.4.1 Natural Images . . . . . . . . . . . . . . . . . . . . . . . . 43 3.4.2 Light Microscopy Data . . . . . . . . . . . . . . . . . . . . 44 3.4.3 Electron Microscopy Data . . . . . . . . . . . . . . . . . . 47 3.4.4 Errors and Limitations . . . . . . . . . . . . . . . . . . . . 48 3.5 Conclusion and Followup Work . . . . . . . . . . . . . . . . . . . 50 4 Fourier Image Transformer 53 4.1 Transformers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 4.1.1 Attention Is All You Need . . . . . . . . . . . . . . . . . . 55 4.1.2 Fast-Transformers . . . . . . . . . . . . . . . . . . . . . . . 56 4.1.3 Transformers in Computer Vision . . . . . . . . . . . . . . 57 4.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 4.2.1 Fourier Domain Encodings (FDEs) . . . . . . . . . . . . . 57 4.2.2 Fourier Coefficient Loss . . . . . . . . . . . . . . . . . . . . 59 4.3 FIT for Super-Resolution . . . . . . . . . . . . . . . . . . . . . . . 60 4.3.1 Super-Resolution Data . . . . . . . . . . . . . . . . . . . . 60 4.3.2 Super-Resolution Experiments . . . . . . . . . . . . . . . . 61 4.4 FIT for Tomography . . . . . . . . . . . . . . . . . . . . . . . . . 63 4.4.1 Computed Tomography Data . . . . . . . . . . . . . . . . 64 4.4.2 Computed Tomography Experiments . . . . . . . . . . . . 66 4.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 5 Conclusions and Outlook 71