Latente Gauß-Prozess-Modelle (latent Gaussian process models) werden von Wissenschaftlern benutzt, um verborgenen Muster in Daten zu er- kennen, Expertenwissen in probabilistische Modelle einfließen zu lassen und um Vorhersagen über die Zukunft zu treffen. Diese Modelle wurden erfolgreich in vielen Gebieten wie Robotik, Geologie, Genetik und Medizin angewendet. Gauß-Prozesse definieren Verteilungen über Funktionen und können als flexible Bausteine verwendet werden, um aussagekräftige probabilistische Modelle zu entwickeln. Dabei ist die größte Herausforderung, eine geeignete Inferenzmethode zu implementieren. Inferenz in probabilistischen Modellen bedeutet die A-Posteriori-Verteilung der latenten Variablen, gegeben der Daten, zu berechnen. Die meisten interessanten latenten Gauß-Prozess-Modelle haben zurzeit nur begrenzte Anwendungsmöglichkeiten auf großen Datensätzen. In dieser Doktorarbeit stellen wir eine neue effiziente Inferenzmethode für latente Gauß-Prozess-Modelle vor. Unser neuer Ansatz, den wir augmented variational inference nennen, basiert auf der Idee, eine erweiterte (augmented) Version des Gauß-Prozess-Modells zu betrachten, welche bedingt konjugiert (conditionally conjugate) ist. Wir zeigen, dass Inferenz in dem erweiterten Modell effektiver ist und dass alle Schritte des variational inference Algorithmus in geschlossener Form berechnet werden können, was mit früheren Ansätzen nicht möglich war. Unser neues Inferenzkonzept ermöglicht es, neue latente Gauß-Prozess- Modelle zu studieren, die zu innovativen Ergebnissen im Bereich der Sprachmodellierung, genetischen Assoziationsstudien und Quantifizierung der Unsicherheit in Klassifikationsproblemen führen. / Latent Gaussian process (GP) models help scientists to uncover hidden structure in data, express domain knowledge and form predictions about the future. These models have been successfully applied in many domains including robotics, geology, genetics and medicine. A GP defines a distribution over functions and can be used as a flexible building block to develop expressive probabilistic models. The main computational challenge of these models is to make inference about the unobserved latent random variables, that is, computing the posterior distribution given the data. Currently, most interesting Gaussian process models have limited applicability to big data. This thesis develops a new efficient inference approach for latent GP models. Our new inference framework, which we call augmented variational inference, is based on the idea of considering an augmented version of the intractable GP model that renders the model conditionally conjugate. We show that inference in the augmented model is more efficient and, unlike in previous approaches, all updates can be computed in closed form. The ideas around our inference framework facilitate novel latent GP models that lead to new results in language modeling, genetic association studies and uncertainty quantification in classification tasks.

Glioblastoma multiforme (GMB) is an extremely aggressive and invasive brain cancer with a median survival of less than one year. In addition, due to its anaplastic nature the histological classification of this cancer is not simple. These characteristics make this disease an interesting and important target for new methodologies of analysis and classification. In recent years, molecular information has been used to segregate and analyze GBM patients, but generally this methodology utilizes single-`omic' data to perform the classification or multi-’omic’ data in a sequential manner. In this project, a novel approach for the classification and analysis of patients with GBM is presented. The main objective of this work is to find clusters of patients with distinctive profiles using multi-’omic’ data with a real integrative methodology. During the last years, the TCGA consortium has made publicly available thousands of multi-’omic’ samples for multiple cancer types. Thanks to this, it was possible to obtain numerous GBM samples (> 300) with data for gene and microRNA expression, CpG sites methylation and copy-number variation (CNV). To achieve our objective, a mixture of linear models were built for each gene using its expression as output and a mixture of multi-`omic' data as covariates. Each model was coupled with a lasso penalization scheme, and thanks to the mixture nature of the model, it was possible to fit multiple submodels to discover different linear relationships in the same model. This complex but interpretable method was used to train over \numprint{10000} models. For \texttildelow \numprint{2400} cases, two or more submodels were obtained. Using the models and their submodels, 6 different clusters of patients were discovered. The clusters were profiled based on clinical information and gene mutations. Through this analysis, a clear separation between the younger patients and with higher survival rate (Clusters 1, 2 and 3) and those from older patients and lower survival rate (Clusters 4, 5 and 6) was found. Mutations in the gene IDH1 were found almost exclusively in Cluster 2, additionally, Cluster 5 presented a hypermutated profile. Finally, several genes not previously related to GBM showed a significant presence in the clusters, such as C15orf2 and CHEK2. The most significant models for each clusters were studied, with a special focus on their covariants. It was discovered that the number of shared significant models were very small and that the well known GBM related genes appeared as significant covariates for plenty of models, such as EGFR1 and TP53. Along with them, ubiquitin-related genes (UBC and UBD) and NRF1, which have not been linked to GBM previously, had a very significant role. This work showed the potential of using a mixture of linear models to integrate multi-’omic’ data and to group patients in order to profile them and find novel markers. The resulting clusters showed unique profiles and their significant models and covariates were comprised by well known GBM related genes and novel markers, which present the possibility for new approaches to study and attack this disease. The next step of the project is to improve several elements of the methodology to achieve a more detail analysis of the models and covariates, in particular taking into account the regression coefficients of the submodels.

Epilepsy is the most common neurological disorder and an accurate forecast of seizures would help to overcome the patient’s uncertainty and helplessness. In this contribution, we present and discuss a novel methodology for the classification of intracranial electroencephalography (iEEG) for seizure prediction. Contrary to previous approaches, we categorically refrain from an extraction of hand-crafted features and use a convolutional neural network (CNN) topology instead for both the determination of suitable signal characteristics and the binary classification of preictal and interictal segments. Three different models have been evaluated on public datasets with long-term recordings from four dogs and three patients. Overall, our findings demonstrate the general applicability. In this work we discuss the strengths and limitations of our methodology.

The use of computers and algorithms to deal with human language, in both spoken and written form, is summarized by the term natural language processing (nlp). Modeling language in a way that is suitable for computers plays an important role in nlp. One idea is to use formalisms from theoretical computer science for that purpose. For example, one can try to find an automaton to capture the valid written sentences of a language. Finding such an automaton by way of examples is called training. In this work, we also consider the structure of sentences by making use of trees. We use weighted tree automata (wta) in order to deal with such tree structures. Those devices assign weights to trees in order to, for example, distinguish between good and bad structures. The well-known expectation-maximization algorithm can be used to train the weights for a wta while the state behavior stays fixed. As a way to adapt the state behavior of a wta, state splitting, i.e. dividing a state into several new states, and state merging, i.e. replacing several states by a single new state, can be used. State splitting, state merging, and the expectation maximization algorithm already were combined into the state splitting and merging algorithm, which was successfully applied in practice. In our work, we formalized this approach in order to show properties of the algorithm. We also examined a new approach – the count-based state merging algorithm – which exclusively relies on state merging. When dealing with trees, another important tool is binarization. A binarization is a strategy to code arbitrary trees by binary trees. For each of three different binarizations we showed that wta together with the binarization are as powerful as weighted unranked tree automata (wuta). We also showed that this is still true if only probabilistic wta and probabilistic wuta are considered.:How to Read This Thesis 1. Introduction 1.1. The Contributions and the Structure of This Work 2. Preliminaries 2.1. Sets, Relations, Functions, Families, and Extrema 2.2. Algebraic Structures 2.3. Formal Languages 3. Language Formalisms 3.1. Context-Free Grammars (CFGs) 3.2. Context-Free Grammars with Latent Annotations (CFG-LAs) 3.3. Weighted Tree Automata (WTAs) 3.4. Equivalences of WCFG-LAs and WTAs 4. Training of WTAs 4.1. Probability Distributions 4.2. Maximum Likelihood Estimation 4.3. Probabilities and WTAs 4.4. The EM Algorithm for WTAs 4.5. Inside and Outside Weights 4.6. Adaption of the Estimation of Corazza and Satta [CS07] to WTAs 5. State Splitting and Merging 5.1. State Splitting and Merging for Weighted Tree Automata 5.1.1. Splitting Weights and Probabilities 5.1.2. Merging Probabilities 5.2. The State Splitting and Merging Algorithm 5.2.1. Finding a Good π-Distributor 5.2.2. Notes About the Berkeley Parser 5.3. Conclusion and Further Research 6. Count-Based State Merging 6.1. Preliminaries 6.2. The Likelihood of the Maximum Likelihood Estimate and Its Behavior While Merging 6.3. The Count-Based State Merging Algorithm 6.3.1. Further Adjustments for Practical Implementations 6.4. Implementation of Count-Based State Merging 6.5. Experiments with Artificial Automata and Corpora 6.5.1. The Artificial Automata 6.5.2. Results 6.6. Experiments with the Penn Treebank 6.7. Comparison to the Approach of Carrasco, Oncina, and Calera-Rubio [COC01] 6.8. Conclusion and Further Research 7. Binarization 7.1. Preliminaries 7.2. Relating WSTAs and WUTAs via Binarizations 7.2.1. Left-Branching Binarization 7.2.2. Right-Branching Binarization 7.2.3. Mixed Binarization 7.3. The Probabilistic Case 7.3.1. Additional Preliminaries About WSAs 7.3.2. Constructing an Out-Probabilistic WSA from a Converging WSA 7.3.3. Binarization and Probabilistic Tree Automata 7.4. Connection to the Training Methods in Previous Chapters 7.5. Conclusion and Further Research A. Proofs for Preliminaries B. Proofs for Training of WTAs C. Proofs for State Splitting and Merging D. Proofs for Count-Based State Merging Bibliography List of Algorithms List of Figures List of Tables Index Table of Variable Names

Diese Dissertation bietet eine empirische Untersuchung deutscher Verben auf der Grundlage statistischer Beschreibungen, die aus einem großen deutschen Textkorpus gewonnen wurden. In einem kurzen Überblick über linguistische Theorien zur lexikalischen Semantik von Verben skizziere ich die Idee, dass die Verbbedeutung wesentlich von seiner Argumentstruktur (der Anzahl und Art der Argumente, die zusammen mit dem Verb auftreten) und seiner Aspektstruktur (Eigenschaften, die den zeitlichen Ablauf des vom Verb denotierten Ereignisses bestimmen) abhängt. Anschließend erstelle ich statistische Beschreibungen von Verben, die auf diesen beiden unterschiedlichen Bedeutungsfacetten basieren. Insbesondere untersuche ich verbale Subkategorisierung, Selektionspräferenzen und Aspekt. Alle diese Modellierungsstrategien werden anhand einer gemeinsamen Aufgabe, der Verbklassifikation, bewertet. Ich zeige, dass im Rahmen von maschinellem Lernen erworbene Merkmale, die verbale lexikalische Aspekte erfassen, für eine Anwendung von Vorteil sind, die Argumentstrukturen betrifft, nämlich semantische Rollenkennzeichnung. Darüber hinaus zeige ich, dass Merkmale, die die verbale Argumentstruktur erfassen, bei der Aufgabe, ein Verb nach seiner Aspektklasse zu klassifizieren, gut funktionieren. Diese Ergebnisse bestätigen, dass diese beiden Facetten der Verbbedeutung auf grundsätzliche Weise zusammenhängen. / This dissertation provides an empirical investigation of German verbs conducted on the basis of statistical descriptions acquired from a large corpus of German text. In a brief overview of the linguistic theory pertaining to the lexical semantics of verbs, I outline the idea that verb meaning is composed of argument structure (the number and types of arguments that co-occur with a verb) and aspectual structure (properties describing the temporal progression of an event referenced by the verb). I then produce statistical descriptions of verbs according to these two distinct facets of meaning: In particular, I examine verbal subcategorisation, selectional preferences, and aspectual type. All three of these modelling strategies are evaluated on a common task, automatic verb classification. I demonstrate that automatically acquired features capturing verbal lexical aspect are beneficial for an application that concerns argument structure, namely semantic role labelling. Furthermore, I demonstrate that features capturing verbal argument structure perform well on the task of classifying a verb for its aspectual type. These findings suggest that these two facets of verb meaning are related in an underlying way.

Die Analyse von NGS Daten ist ein zentraler Aspekt der modernen genomischen Forschung. Bei der Extraktion von Daten aus den beiden am häufigsten verwendeten Quellorganismen bestehen jedoch vielfältige Problemstellungen. Im ersten Kapitel wird ein neuartiger Ansatz vorgestellt welcher einen Abstand zwischen Krebszellinienkulturen auf Grundlage ihrer kleinen genomischen Varianten bestimmt um die Kulturen zu identifizieren. Eine Voll-Exom sequenzierte Kultur wird durch paarweise Vergleiche zu Referenzdatensätzen identifiziert so ein gemessener Abstand geringer ist als dies bei nicht verwandten Kulturen zu erwarten wäre. Die Wirksamkeit der Methode wurde verifiziert, jedoch verbleiben Einschränkung da nur das Sequenzierformat des Voll-Exoms unterstützt wird. Daher wird im zweiten Kapitel eine publizierte Modifikation des Ansatzes vorgestellt welcher die Unterstützung der weitläufig genutzten Bulk RNA sowie der Panel-Sequenzierung ermöglicht. Die Ausweitung der Technologiebasis führt jedoch zu einer Verstärkung von Störeffekten welche zu Verletzungen der mathematischen Konditionen einer Abstandsmetrik führen. Daher werden die entstandenen Verletzungen durch statistische Verfahren zuerst quantifiziert und danach durch dynamische Schwellwertanpassungen erfolgreich kompensiert. Das dritte Kapitel stellt eine neuartige Daten-Aufwertungsmethode (Data-Augmentation) vor welche das Trainieren von maschinellen Lernmodellen in Abwesenheit von neoplastischen Trainingsdaten ermöglicht. Ein abstraktes Abstandsmaß wird zwischen neoplastischen Entitäten sowie Entitäten gesundem Ursprungs mittels einer transkriptomischen Dekonvolution hergestellt. Die Ausgabe der Dekonvolution erlaubt dann das effektive Vorhersagen von klinischen Eigenschaften von seltenen jedoch biologisch vielfältigen Krebsarten wobei die prädiktive Kraft des Verfahrens der des etablierten Goldstandard ebenbürtig ist. / The analysis of NGS data is a central aspect of modern Molecular Genetics and Oncology. The first scientific contribution is the development of a method which identifies Whole-exome-sequenced CCL via the quantification of a distance between their sets of small genomic variants. A distinguishing aspect of the method is that it was designed for the computer-based identification of NGS-sequenced CCL. An identification of an unknown CCL occurs when its abstract distance to a known CCL is smaller than is expected due to chance. The method performed favorably during benchmarks but only supported the Whole-exome-sequencing technology. The second contribution therefore extended the identification method by additionally supporting the Bulk mRNA-sequencing technology and Panel-sequencing format. However, the technological extension incurred predictive biases which detrimentally affected the quantification of abstract distances. Hence, statistical methods were introduced to quantify and compensate for confounding factors. The method revealed a heterogeneity-robust benchmark performance at the trade-off of a slightly reduced sensitivity compared to the Whole-exome-sequencing method. The third contribution is a method which trains Machine-Learning models for rare and diverse cancer types. Machine-Learning models are subsequently trained on these distances to predict clinically relevant characteristics. The performance of such-trained models was comparable to that of models trained on both the substituted neoplastic data and the gold-standard biomarker Ki-67. No proliferation rate-indicative features were utilized to predict clinical characteristics which is why the method can complement the proliferation rate-oriented pathological assessment of biopsies. The thesis revealed that the quantification of an abstract distance can address sources of erroneous NGS data analysis.

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

The industrial production of semiconductor assemblies is subject to high requirements. As a result, several tests are needed in terms of component quality. In the long run, manual quality assurance (QA) is often connected with higher expenditures. Using a technique based on machine learning, some of these tests may be carried out automatically. Deep neural networks (NN) have shown to be very effective in a diverse range of computer vision applications. Especially convolutional neural networks (CNN), which belong to a subset of NN, are an effective tool for image classification. Deep NNs have the disadvantage of requiring a significant quantity of training data to reach excellent performance. When the dataset is too small a phenomenon known as overfitting can occur. Massive amounts of data cannot be supplied in certain contexts, such as the production of semiconductors. This is especially true given the relatively low number of rejected components in this field. In order to prevent overfitting, a variety of image augmentation methods may be used to the process of artificially creating training images. However, many of those methods cannot be used in certain fields due to their inapplicability. For this thesis, Infineon Technologies AG provided the images of a semiconductor component generated by an ultrasonic microscope. The images can be categorized as having a sufficient number of good and a minority of rejected components, with good components being defined as components that have been deemed to have passed quality control and rejected components being components that contain a defect and did not pass quality control. The accomplishment of the project, the efficacy with which it is carried out, and its level of quality may be dependent on a number of factors; however, selecting the appropriate tools is one of the most important of these factors because it enables significant time and resource savings while also producing the best results. We demonstrate a data augmentation graphical user interface (GUI) tool that has been widely used in the domain of image processing. Using this method, the dataset size has been increased while maintaining the accuracy-time trade-off and optimizing the robustness of deep learning models. The purpose of this work is to develop a user-friendly tool that incorporates traditional, advanced, and smart data augmentation, image processing, and machine learning (ML) approaches. More specifically, the technique mainly uses are zooming, rotation, flipping, cropping, GAN, fusion, histogram matching, autoencoder, image restoration, compression etc. This focuses on implementing and designing a MATLAB GUI for data augmentation and ML models. The thesis was carried out for the Infineon Technologies AG in order to address a challenge that all semiconductor industries experience. The key objective is not only to create an easy- to-use GUI, but also to ensure that its users do not need advanced technical experiences to operate it. This GUI may run on its own as a standalone application. Which may be implemented everywhere for the purposes of data augmentation and classification. The objective is to streamline the working process and make it easy to complete the Quality assurance job even for those who are not familiar with data augmentation, machine learning, or MATLAB. In addition, research will investigate the benefits of data augmentation and image processing, as well as the possibility that these factors might contribute to an improvement in the accuracy of AI models.

High-content screens are widely used to get insight on mechanistic organization of biological systems. Chemical and/or genomic interferences are used to modulate molecular machinery, then light microscopy and quantitative image analysis yield a large number of parameters describing phenotype. However, extracting functional information from such high-content datasets (e.g. links between cellular processes or functions of unknown genes) remains challenging. This work is devoted to the analysis of a multi-parametric image-based genomic screen of endocytosis, the process whereby cells uptake cargoes (signals and nutrients) and distribute them into different subcellular compartments. The complexity of the quantitative endocytic data was approached using different Machine Learning techniques, namely, Clustering methods, Bayesian networks, Principal and Independent component analysis, Artificial neural networks. The main goal of such an analysis is to predict possible modes of action of screened genes and also to find candidate genes that can be involved in a process of interest. The degree of freedom for the multidimensional phenotypic space was identified using the data distributions, and then the high-content data were deconvolved into separate signals from different cellular modules. Some of those basic signals (phenotypic traits) were straightforward to interpret in terms of known molecular processes; the other components gave insight into interesting directions for further research. The phenotypic profile of perturbation of individual genes are sparse in coordinates of the basic signals, and, therefore, intrinsically suggest their functional roles in cellular processes. Being a very fundamental process, endocytosis is specifically modulated by a variety of different pathways in the cell; therefore, endocytic phenotyping can be used for analysis of non-endocytic modules in the cell. Proposed approach can be also generalized for analysis of other high-content screens.:Contents Objectives Chapter 1 Introduction 1.1 High-content biological data 1.1.1 Different perturbation types for HCS 1.1.2 Types of observations in HTS 1.1.3 Goals and outcomes of MP HTS 1.1.4 An overview of the classical methods of analysis of biological HT- and HCS data 1.2 Machine learning for systems biology 1.2.1 Feature selection 1.2.2 Unsupervised learning 1.2.3 Supervised learning 1.2.4 Artificial neural networks 1.3 Endocytosis as a system process 1.3.1 Endocytic compartments and main players 1.3.2 Relation to other cellular processes Chapter 2 Experimental and analytical techniques 2.1 Experimental methods 2.1.1 RNA interference 2.1.2 Quantitative multiparametric image analysis 2.2 Detailed description of the endocytic HCS dataset 2.2.1 Basic properties of the endocytic dataset 2.2.2 Control subset of genes 2.3 Machine learning methods 2.3.1 Latent variables models 2.3.2 Clustering 2.3.3 Bayesian networks 2.3.4 Neural networks Chapter 3 Results 3.1 Selection of labeled data for training and validation based on KEGG information about genes pathways 3.2 Clustering of genes 3.2.1 Comparison of clustering techniques on control dataset 3.2.2 Clustering results 3.3 Independent components as basic phenotypes 3.3.1 Algorithm for identification of the best number of independent components 3.3.2 Application of ICA on the full dataset and on separate assays of the screen 3.3.3 Gene annotation based on revealed phenotypes 3.3.4 Searching for genes with target function 3.4 Bayesian network on endocytic parameters 3.4.1 Prediction of pathway based on parameters values using Naïve Bayesian Classifier 3.4.2 General Bayesian Networks 3.5 Neural networks 3.5.1 Autoencoders as nonlinear ICA 3.5.2 siRNA sequence motives discovery with deep NN 3.6 Biological results 3.6.1 Rab11 ZNF-specific phenotype found by ICA 3.6.2 Structure of BN revealed dependency between endocytosis and cell adhesion Chapter 4 Discussion 4.1 Machine learning approaches for discovery of phenotypic patterns 4.1.1 Functional annotation of unknown genes based on phenotypic profiles 4.1.2 Candidate genes search 4.2 Adaptation to other HCS data and generalization Chapter 5 Outlook and future perspectives 5.1 Handling sequence-dependent off-target effects with neural networks 5.2 Transition between machine learning and systems biology models Acknowledgements References Appendix A.1 Full list of cellular and endocytic parameters A.2 Description of independent components of the full dataset A.3 Description of independent components extracted from separate assays of the HCS

Machine Learning (ML) profitiert erheblich von der Verfügbarkeit großer Mengen an Trainingsdaten, sowohl im Bezug auf die Anzahl an Datenpunkten, als auch auf die Anzahl an Features pro Datenpunkt. Es ist allerdings oft weder möglich, noch gewollt, mehr Daten unter zentraler Kontrolle zu aggregieren. Multi-Party-Computation (MPC)-Protokolle stellen eine Lösung dieses Dilemmas in Aussicht, indem sie es mehreren Parteien erlauben, ML-Modelle auf der Gesamtheit ihrer Daten zu trainieren, ohne die Eingabedaten preiszugeben. Generische MPC-Ansätze bringen allerdings erheblichen Mehraufwand in der Kommunikations- und Laufzeitkomplexität mit sich, wodurch sie sich nur beschränkt für den Einsatz in der Praxis eignen. Das Ziel dieser Arbeit ist es, Privatsphäreerhaltendes Machine Learning mittels MPC praxistauglich zu machen. Zuerst fokussieren wir uns auf zwei Anwendungen, lineare Regression und Klassifikation von Dokumenten. Hier zeigen wir, dass sich der Kommunikations- und Rechenaufwand erheblich reduzieren lässt, indem die aufwändigsten Teile der Berechnung durch Sub-Protokolle ersetzt werden, welche auf die Zusammensetzung der Parteien, die Verteilung der Daten, und die Zahlendarstellung zugeschnitten sind. Insbesondere das Ausnutzen dünnbesetzter Datenrepräsentationen kann die Effizienz der Protokolle deutlich verbessern. Diese Beobachtung verallgemeinern wir anschließend durch die Entwicklung einer Datenstruktur für solch dünnbesetzte Daten, sowie dazugehöriger Zugriffsprotokolle. Aufbauend auf dieser Datenstruktur implementieren wir verschiedene Operationen der Linearen Algebra, welche in einer Vielzahl von Anwendungen genutzt werden. Insgesamt zeigt die vorliegende Arbeit, dass MPC ein vielversprechendes Werkzeug auf dem Weg zu Privatsphäre-erhaltendem Machine Learning ist, und die von uns entwickelten Protokolle stellen einen wesentlichen Schritt in diese Richtung dar. / Machine learning (ML) greatly benefits from the availability of large amounts of training data, both in terms of the number of samples, and the number of features per sample. However, aggregating more data under centralized control is not always possible, nor desirable, due to security and privacy concerns, regulation, or competition. Secure multi-party computation (MPC) protocols promise a solution to this dilemma, allowing multiple parties to train ML models on their joint datasets while provably preserving the confidentiality of the inputs. However, generic approaches to MPC result in large computation and communication overheads, which limits the applicability in practice. The goal of this thesis is to make privacy-preserving machine learning with secure computation practical. First, we focus on two high-level applications, linear regression and document classification. We show that communication and computation overhead can be greatly reduced by identifying the costliest parts of the computation, and replacing them with sub-protocols that are tailored to the number and arrangement of parties, the data distribution, and the number representation used. One of our main findings is that exploiting sparsity in the data representation enables considerable efficiency improvements. We go on to generalize this observation, and implement a low-level data structure for sparse data, with corresponding secure access protocols. On top of this data structure, we develop several linear algebra algorithms that can be used in a wide range of applications. Finally, we turn to improving a cryptographic primitive named vector-OLE, for which we propose a novel protocol that helps speed up a wide range of secure computation tasks, within private machine learning and beyond. Overall, our work shows that MPC indeed offers a promising avenue towards practical privacy-preserving machine learning, and the protocols we developed constitute a substantial step in that direction.