Refine
Has Fulltext
- yes (23)
Is part of the Bibliography
- yes (23)
Year of publication
Document Type
- Doctoral Thesis (21)
- Preprint (2)
Language
- English (23) (remove)
Keywords
- Maschinelles Lernen (23) (remove)
Institute
- Betriebswirtschaftliches Institut (6)
- Institut für Informatik (6)
- Graduate School of Life Sciences (4)
- Institut für Theoretische Physik und Astrophysik (2)
- Institut für deutsche Philologie (2)
- Fakultät für Chemie und Pharmazie (1)
- Graduate School of Science and Technology (1)
- Institut für Funktionsmaterialien und Biofabrikation (1)
- Institut für Klinische Epidemiologie und Biometrie (1)
- Institut für diagnostische und interventionelle Radiologie (Institut für Röntgendiagnostik) (1)
Sonstige beteiligte Institutionen
EU-Project number / Contract (GA) number
Introduction.
Mobile health (mHealth) integrates mobile devices into healthcare, enabling remote monitoring, data collection, and personalized interventions. Machine Learning (ML), a subfield of Artificial Intelligence (AI), can use mHealth data to confirm or extend domain knowledge by finding associations within the data, i.e., with the goal of improving healthcare decisions. In this work, two data collection techniques were used for mHealth data fed into ML systems: Mobile Crowdsensing (MCS), which is a collaborative data gathering approach, and Ecological Momentary Assessments (EMA), which capture real-time individual experiences within the individual’s common environments using questionnaires and sensors. We collected EMA and MCS data on tinnitus and COVID-19. About 15 % of the world’s population suffers from tinnitus.
Materials & Methods.
This thesis investigates the challenges of ML systems when using MCS and EMA data. It asks: How can ML confirm or broad domain knowledge? Domain knowledge refers to expertise and understanding in a specific field, gained through experience and education. Are ML systems always superior to simple heuristics and if yes, how can one reach explainable AI (XAI) in the presence of mHealth data? An XAI method enables a human to understand why a model makes certain predictions. Finally, which guidelines can be beneficial for the use of ML within the mHealth domain? In tinnitus research, ML discerns gender, temperature, and season-related variations among patients. In the realm of COVID-19, we collaboratively designed a COVID-19 check app for public education, incorporating EMA data to offer informative feedback on COVID-19-related matters. This thesis uses seven EMA datasets with more than 250,000 assessments. Our analyses revealed a set of challenges: App user over-representation, time gaps, identity ambiguity, and operating system specific rounding errors, among others. Our systematic review of 450 medical studies assessed prior utilization of XAI methods.
Results.
ML models predict gender and tinnitus perception, validating gender-linked tinnitus disparities. Using season and temperature to predict tinnitus shows the association of these variables with tinnitus. Multiple assessments of one app user can constitute a group. Neglecting these groups in data sets leads to model overfitting. In select instances, heuristics outperform ML models, highlighting the need for domain expert consultation to unveil hidden groups or find simple heuristics.
Conclusion.
This thesis suggests guidelines for mHealth related data analyses and improves estimates for ML performance. Close communication with medical domain experts to identify latent user subsets and incremental benefits of ML is essential.
This work introduced the reader to all relevant fields to tap into an ultrasound-based state of charge estimation and provides a blueprint for the procedure to achieve and test the fundamentals of such an approach. It spanned from an in-depth electrochemical characterization of the studied battery cells over establishing the measurement technique, digital processing of ultrasonic transmission signals, and characterization of the SoC dependent property changes of those signals to a proof of concept of an ultrasound-based state of charge estimation.
The State of the art & theoretical background chapter focused on the battery section on the mechanical property changes of lithium-ion batteries during operation. The components and the processes involved to manufacture a battery cell were described to establish the fundamentals for later interrogation. A comprehensive summary of methods for state estimation was given and an emphasis was laid on mechanical methods, including a critical review of the most recent research on ultrasound-based state estimation. Afterward, the fundamentals of ultrasonic non-destructive evaluation were introduced, starting with the sound propagation modes in isotropic boundary-free media, followed by the introduction of boundaries and non-isotropic structure to finally approach the class of fluid-saturated porous media, which batteries can be counted to. As the processing of the ultrasonic signals transmitted through lithium-ion battery cells with the aim of feature extraction was one of the main goals of this work, the fundamentals of digital signal processing and methods for the time of flight estimation were reviewed and compared in a separate section.
All available information on the interrogated battery cell and the instrumentation was collected in the Experimental methods & instrumentation chapter, including a detailed step-by-step manual of the process developed in this work to create and attach a sensor stack for ultrasonic interrogation based on low-cost off-the-shelf piezo elements.
The Results & discussion chapter opened with an in-depth electrochemical and post-mortem interrogation to reverse engineer the battery cell design and its internal structure. The combination of inductively coupled plasma-optical emission spectrometry and incremental capacity analysis applied to three-electrode lab cells, constructed from the studied battery cell’s materials, allowed to identify the SoC ranges in which phase transitions and staging occur and thereby directly links changes in the ultrasonic signal properties with the state of the active materials, which makes this work stand out among other studies on ultrasound-based state estimation. Additional dilatometer experiments were able to prove that the measured effect in ultrasonic time of flight cannot originate from the thickness increase of the battery cells alone, as this thickness increase is smaller and in opposite direction to the change in time of flight. Therefore, changes in elastic modulus and density have to be responsible for the observed effect.
The construction of the sensor stack from off-the-shelf piezo elements, its electromagnetic shielding, and attachment to both sides of the battery cells was treated in a subsequent section. Experiments verified the necessity of shielding and its negligible influence on the ultrasonic signals. A hypothesis describing the metal layer in the pouch foil to be the transport medium of an electrical coupling/distortion between sending and receiving sensor was formulated and tested. Impedance spectroscopy was shown to be a useful tool to characterize the resonant behavior of piezo elements and ensure the mechanical coupling of such to the surface of the battery cells. The excitation of the piezo elements by a raised cosine (RCn) waveform with varied center frequency in the range of 50 kHz to 250 kHz was studied in the frequency domain and the influence of the resonant behavior, as identified prior by impedance spectroscopy, on waveform and frequency content was evaluated to be uncritical. Therefore, the forced oscillation produced by this excitation was assumed to be mechanically coupled as ultrasonic waves into the battery cells.
The ultrasonic waves transmitted through the battery cell were recorded by piezo elements on the opposing side. A first inspection of the raw, unprocessed signals identified the transmission of two main wave packages and allowed the identification of two major trends: the time of flight of ultrasonic wave packages decreases with the center frequency of the RCn waveform, and with state of charge. These trends were to be assessed further in the subsequent sections. Therefore, methods for the extraction of features (properties) from the ultrasonic signals were established, compared, and tested in a dedicated section. Several simple and advanced thresholding methods were compared with envelope-based and cross-correlation methods to estimate the time of flight (ToF). It was demonstrated that the envelope-based method yields the most robust estimate for the first and second wave package. This finding is in accordance with the literature stating that an envelope-based method is best suited for dispersive, absorptive media [204], to which lithium-ion batteries are counted. Respective trends were already suggested by the heatmap plots of the raw signals vs. RCn frequency and SoC. To enable such a robust estimate, an FIR filter had to be designed to preprocess the transmitted signals and thereby attenuate frequency components that verifiably lead to a distorted shape of the envelope.
With a robust ToF estimation method selected, the characterization of the signal properties ToF and transmitted energy content (EC) was performed in-depth. A study of cycle-to-cycle variations unveiled that the signal properties are affected by a long rest period and the associated relaxation of the multi-particle system “battery cell” to equilibrium. In detail, during cycling, the signal properties don’t reach the same value at a given SoC in two subsequent cycles if the first of the two cycles follows a long rest period. In accordance with the literature, a break-in period, making up for more than ten cycles post-formation, was observed. During this break-in period, the mechanical properties of the system are said to change until a steady state is reached [25]. Experiments at different C-rate showed that ultrasonic signal properties can sense the non-equilibrium state of a battery cell, characterized by an increasing area between charge and discharge curve of the respective signal property vs. SoC plot. This non-equilibrium state relaxes in the rest period following the discharge after the cut-off voltage is reached. The relaxation in the rest period following the charge is much smaller and shows little C-rate dependency as the state is prepared by constant voltage charging at the end of charge voltage. For a purely statistical SoC estimation approach, as employed in this work, where only instantaneous measurements are taken into account and the historic course of the measurement is not utilized as a source of information, the presence of hysteresis and relaxation leads to a reduced estimation accuracy. Future research should address this issue or even utilize the relaxation to improve the estimation accuracy, by incorporating historic information, e.g., by using the derivative of a signal property as an additional feature. The signal properties were then tested for their correlation with SoC as a function of RCn frequency. This allowed identifying trends in the behavior of the signal properties as a function of RCn frequency and C-rate in a condensed fashion and thereby enabled to predict the frequency range, about 50 kHz to 125 kHz, in which the course of the signal properties is best suited for SoC estimation.
The final section provided a proof of concept of the ultrasound-based SoC estimation, by applying a support vector regression (SVR) to before thoroughly studied ultrasonic signal properties, as well as current and battery cell voltage. The included case study was split into different parts that assessed the ability of an SVR to estimate the SoC in a variety of scenarios. Seven battery cells, prepared with sensor stacks attached to both faces, were used to generate 14 datasets. First, a comparison of self-tests, where a portion of a dataset is used for training and another for testing, and cross-tests, which use the dataset of one cell for training and the dataset of another for testing, was performed. A root mean square error (RMSE) of 3.9% to 4.8% SoC and 3.6% to 10.0% SoC was achieved, respectively. In general, it was observed that the SVR is prone to overestimation at low SoCs and underestimation at high SoCs, which was attributed to the pronounced hysteresis and relaxation of the ultrasonic signal properties in this SoC ranges. The fact that higher accuracy is achieved, if the exact cell is known to the model, indicates that a variation between cells exists. This variation between cells can originate from differences in mechanical properties as a result of production variations or from differences in manual sensor placement, mechanical coupling, or resonant behavior of the ultrasonic sensors. To mitigate the effect of the cell-to-cell variations, a test was performed, where the datasets of six out of the seven cells were combined as training data, and the dataset of the seventh cell was used for testing. This reduced the spread of the RMSE from (3.6 - 10.0)% SoC to (5.9 – 8.5)% SoC, respectively, once again stating that a databased approach for state estimation becomes more reliable with a large data basis. Utilizing self-tests on seven datasets, the effect of additional features on the state estimation result was tested. The involvement of an additional feature did not necessarily improve the estimation accuracy, but it was shown that a combination of ultrasonic and electrical features is superior to the training with these features alone. To test the ability of the model to estimate the SoC in unknown cycling conditions, a test was performed where the C-rate of the test dataset was not included in the training data. The result suggests that for practical applications it might be sufficient to perform training with the boundary of the use cases in a controlled laboratory environment to handle the estimation in a broad spectrum of use cases.
In comparison with literature, this study stands out by utilizing and modifying off-the-shelf piezo elements to equip state-of-the-art lithium-ion battery cells with ultrasonic sensors, employing a range of center frequencies for the waveform, transmitted through the battery cell, instead of a fixed frequency and by allowing the SVR to choose the frequency that yields the best result. The characterization of the ultrasonic signal properties as a function of RCn frequency and SoC and the assignment of characteristic changes in the signal properties to electrochemical processes, such as phase transitions and staging, makes this work unique. By studying a range of use cases, it was demonstrated that an improved SoC estimation accuracy can be achieved with the aid of ultrasonic measurements – thanks to the correlation of the mechanical properties of the battery cells with the SoC.
One consequence of the recent coronavirus pandemic is increased demand and use of online services around the globe. At the same time, performance requirements for modern technologies are becoming more stringent as users become accustomed to higher standards. These increased performance and availability requirements, coupled with the unpredictable usage growth, are driving an increasing proportion of applications to run on public cloud platforms as they promise better scalability and reliability.
With data centers already responsible for about one percent of the world's power consumption, optimizing resource usage is of paramount importance. Simultaneously, meeting the increasing and changing resource and performance requirements is only possible by optimizing resource management without introducing additional overhead. This requires the research and development of new modeling approaches to understand the behavior of running applications with minimal information.
However, the emergence of modern software paradigms makes it increasingly difficult to derive such models and renders previous performance modeling techniques infeasible. Modern cloud applications are often deployed as a collection of fine-grained and interconnected components called microservices. Microservice architectures offer massive benefits but also have broad implications for the performance characteristics of the respective systems. In addition, the microservices paradigm is typically paired with a DevOps culture, resulting in frequent application and deployment changes. Such applications are often referred to as cloud-native applications. In summary, the increasing use of ever-changing cloud-hosted microservice applications introduces a number of unique challenges for modeling the performance of modern applications. These include the amount, type, and structure of monitoring data, frequent behavioral changes, or infrastructure variabilities. This violates common assumptions of the state of the art and opens a research gap for our work.
In this thesis, we present five techniques for automated learning of performance models for cloud-native software systems. We achieve this by combining machine learning with traditional performance modeling techniques. Unlike previous work, our focus is on cloud-hosted and continuously evolving microservice architectures, so-called cloud-native applications. Therefore, our contributions aim to solve the above challenges to deliver automated performance models with minimal computational overhead and no manual intervention. Depending on the cloud computing model, privacy agreements, or monitoring capabilities of each platform, we identify different scenarios where performance modeling, prediction, and optimization techniques can provide great benefits. Specifically, the contributions of this thesis are as follows:
Monitorless: Application-agnostic prediction of performance degradations.
To manage application performance with only platform-level monitoring, we propose Monitorless, the first truly application-independent approach to detecting performance degradation. We use machine learning to bridge the gap between platform-level monitoring and application-specific measurements, eliminating the need for application-level monitoring. Monitorless creates a single and holistic resource saturation model that can be used for heterogeneous and untrained applications. Results show that Monitorless infers resource-based performance degradation with 97% accuracy. Moreover, it can achieve similar performance to typical autoscaling solutions, despite using less monitoring information.
SuanMing: Predicting performance degradation using tracing.
We introduce SuanMing to mitigate performance issues before they impact the user experience. This contribution is applied in scenarios where tracing tools enable application-level monitoring. SuanMing predicts explainable causes of expected performance degradations and prevents performance degradations before they occur. Evaluation results show that SuanMing can predict and pinpoint future performance degradations with an accuracy of over 90%.
SARDE: Continuous and autonomous estimation of resource demands.
We present SARDE to learn application models for highly variable application deployments. This contribution focuses on the continuous estimation of application resource demands, a key parameter of performance models. SARDE represents an autonomous ensemble estimation technique. It dynamically and continuously optimizes, selects, and executes an ensemble of approaches to estimate resource demands in response to changes in the application or its environment. Through continuous online adaptation, SARDE efficiently achieves an average resource demand estimation error of 15.96% in our evaluation.
DepIC: Learning parametric dependencies from monitoring data.
DepIC utilizes feature selection techniques in combination with an ensemble regression approach to automatically identify and characterize parametric dependencies. Although parametric dependencies can massively improve the accuracy of performance models, DepIC is the first approach to automatically learn such parametric dependencies from passive monitoring data streams. Our evaluation shows that DepIC achieves 91.7% precision in identifying dependencies and reduces the characterization prediction error by 30% compared to the best individual approach.
Baloo: Modeling the configuration space of databases.
To study the impact of different configurations within distributed DBMSs, we introduce Baloo. Our last contribution models the configuration space of databases considering measurement variabilities in the cloud. More specifically, Baloo dynamically estimates the required benchmarking measurements and automatically builds a configuration space model of a given DBMS. Our evaluation of Baloo on a dataset consisting of 900 configuration points shows that the framework achieves a prediction error of less than 11% while saving up to 80% of the measurement effort.
Although the contributions themselves are orthogonally aligned, taken together they provide a holistic approach to performance management of modern cloud-native microservice applications.
Our contributions are a significant step forward as they specifically target novel and cloud-native software development and operation paradigms, surpassing the capabilities and limitations of previous approaches.
In addition, the research presented in this paper also has a significant impact on the industry, as the contributions were developed in collaboration with research teams from Nokia Bell Labs, Huawei, and Google.
Overall, our solutions open up new possibilities for managing and optimizing cloud applications and improve cost and energy efficiency.
This dissertation contributes to the empirical analysis of economic development. The continuing poverty in many Sub-Saharan-African countries as well as the declining trend in growth in the advanced economies that was initiated around the turn of the millennium raises a number of new questions which have received little attention in recent empirical studies. Is culture a decisive factor for economic development? Do larger financial markets trigger positive stimuli with regard to incomes, or is the recent increase in their size in advanced economies detrimental to economic growth? What causes secular stagnation, i.e. the reduction in growth rates of the advanced economies observable over the past 20 years? What is the role of inequality in the growth process, and how do governmental attempts to equalize the income distribution affect economic development? And finally: Is the process of democratization accompanied by an increase in living standards? These are the central questions of this doctoral thesis.
To facilitate the empirical analysis of the determinants of economic growth, this dissertation introduces a new method to compute classifications in the field of social sciences. The approach is based on mathematical algorithms of machine learning and pattern recognition. Whereas the construction of indices typically relies on arbitrary assumptions regarding the aggregation strategy of the underlying attributes, utilization of Support Vector Machines transfers the question of how to aggregate the individual components into a non-linear optimization problem.
Following a brief overview of the theoretical models of economic growth provided in the first chapter, the second chapter illustrates the importance of culture in explaining the differences in incomes across the globe. In particular, if inhabitants have a lower average degree of risk-aversion, the implementation of new technology proceeds much faster compared with countries with a lower tendency towards risk. However, this effect depends on the legal and political framework of the countries, their average level of education, and their stage of development.
The initial wealth of individuals is often not sufficient to cover the cost of investments in both education and new technologies. By providing loans, a developed financial sector may help to overcome this shortage. However, the investigations in the third chapter show that this mechanism is dependent on the development levels of the economies. In poor countries, growth of the financial sector leads to better education and higher investment levels. This effect diminishes along the development process, as intermediary activity is increasingly replaced by speculative transactions. Particularly in times of low technological innovation, an increasing financial sector has a negative impact on economic development. In fact, the world economy is currently in a phase of this kind. Since the turn of the millennium, growth rates in the advanced economies have experienced a multi-national decline, leading to an intense debate about "secular stagnation" initiated at the beginning of 2015. The fourth chapter deals with this phenomenon and shows that the growth potentials of new technologies have been gradually declining since the beginning of the 2000s.
If incomes are unequally distributed, some individuals can invest less in education and technological innovations, which is why the fifth chapter identifies an overall negative effect of inequality on growth. This influence, however, depends on the development level of countries. While the negative effect is strongly pronounced in poor economies with a low degree of equality of opportunity, this influence disappears during the development process. Accordingly, redistributive polices of governments exert a growth-promoting effect in developing countries, while in advanced economies, the fostering of equal opportunities is much more decisive.
The sixth chapter analyzes the growth effect of the political environment and shows that the ambiguity of earlier studies is mainly due to unsophisticated measurement of the degree of democratization. To solve this problem, the chapter introduces a new method based on mathematical algorithms of machine learning and pattern recognition. While the approach can be used for various classification problems in the field of social sciences, in this dissertation it is applied for the problem of democracy measurement. Based on different country examples, the chapter shows that the resulting SVMDI is superior to other indices in modeling the level of democracy. The subsequent empirical analysis emphasizes a significantly positive growth effect of democracy measured via SVMDI.
Traditional fashion retailers are increasingly hard-pressed to keep up with their digital competitors. In this context, the re-invention of brick-and-mortar stores as smart retail environments is being touted as a crucial step towards regaining a competitive edge. This thesis describes a design-oriented research project that deals with automated product tracking on the sales floor and presents three smart fashion store applications that are tied to such localization information: (i) an electronic article surveillance (EAS) system that distinguishes between theft and non-theft events, (ii) an automated checkout system that detects customers’ purchases when they are leaving the store and associates them with individual shopping baskets to automatically initiate payment processes, and (iii) a smart fitting room that detects the items customers bring into individual cabins and identifies the items they are currently most interested in to offer additional customer services (e.g., product recommendations or omnichannel services). The implementation of such cyberphysical systems in established retail environments is challenging, as architectural constraints, well-established customer processes, and customer expectations regarding privacy and convenience pose challenges to system design. To overcome these challenges, this thesis leverages Radio Frequency Identification (RFID) technology and machine learning techniques to address the different detection tasks. To optimally configure the systems and draw robust conclusions regarding their economic value contribution, beyond technological performance criteria, this thesis furthermore introduces a service operations model that allows mapping the systems’ technical detection characteristics to business relevant metrics such as service quality and profitability. This analytical model reveals that the same system component for the detection of object transitions is well suited for the EAS application but does not have the necessary high detection accuracy to be used as a component of an automated checkout system.
Acceleration is a central aim of clinical and technical research in magnetic resonance imaging (MRI) today, with the potential to increase robustness, accessibility and patient comfort, reduce cost, and enable entirely new kinds of examinations. A key component in this endeavor is image reconstruction, as most modern approaches build on advanced signal and image processing. Here, deep learning (DL)-based methods have recently shown considerable potential, with numerous publications demonstrating benefits for MRI reconstruction. However, these methods often come at the cost of an increased risk for subtle yet critical errors. Therefore, the aim of this thesis is to advance DL-based MRI reconstruction, while ensuring high quality and fidelity with measured data. A network architecture specifically suited for this purpose is the variational network (VN). To investigate the benefits these can bring to non-Cartesian cardiac imaging, the first part presents an application of VNs, which were specifically adapted to the reconstruction of accelerated spiral acquisitions. The proposed method is compared to a segmented exam, a U-Net and a compressed sensing (CS) model using qualitative and quantitative measures. While the U-Net performed poorly, the VN as well as the CS reconstruction showed good output quality. In functional cardiac imaging, the proposed real-time method with VN reconstruction substantially accelerates examinations over the gold-standard, from over 10 to just 1 minute. Clinical parameters agreed on average.
Generally in MRI reconstruction, the assessment of image quality is complex, in particular for modern non-linear methods. Therefore, advanced techniques for precise evaluation of quality were subsequently demonstrated.
With two distinct methods, resolution and amplification or suppression of noise are quantified locally in each pixel of a reconstruction. Using these, local maps of resolution and noise in parallel imaging (GRAPPA), CS, U-Net and VN reconstructions were determined for MR images of the brain. In the tested images, GRAPPA delivers uniform and ideal resolution, but amplifies noise noticeably. The other methods adapt their behavior to image structure, where different levels of local blurring were observed at edges compared to homogeneous areas, and noise was suppressed except at edges. Overall, VNs were found to combine a number of advantageous properties, including a good trade-off between resolution and noise, fast reconstruction times, and high overall image quality and fidelity of the produced output. Therefore, this network architecture seems highly promising for MRI reconstruction.
Context-specific Consistencies in Information Extraction: Rule-based and Probabilistic Approaches
(2015)
Large amounts of communication, documentation as well as knowledge and information are stored in textual documents. Most often, these texts like webpages, books, tweets or reports are only available in an unstructured representation since they are created and interpreted by humans. In order to take advantage of this huge amount of concealed information and to include it in analytic processes, it needs to be transformed into a structured representation. Information extraction considers exactly this task. It tries to identify well-defined entities and relations in unstructured data and especially in textual documents.
Interesting entities are often consistently structured within a certain context, especially in semi-structured texts. However, their actual composition varies and is possibly inconsistent among different contexts. Information extraction models stay behind their potential and return inferior results if they do not consider these consistencies during processing. This work presents a selection of practical and novel approaches for exploiting these context-specific consistencies in information extraction tasks. The approaches direct their attention not only to one technique, but are based on handcrafted rules as well as probabilistic models.
A new rule-based system called UIMA Ruta has been developed in order to provide optimal conditions for rule engineers. This system consists of a compact rule language with a high expressiveness and strong development support. Both elements facilitate rapid development of information extraction applications and improve the general engineering experience, which reduces the necessary efforts and costs when specifying rules.
The advantages and applicability of UIMA Ruta for exploiting context-specific consistencies are illustrated in three case studies. They utilize different engineering approaches for including the consistencies in the information extraction task. Either the recall is increased by finding additional entities with similar composition, or the precision is improved by filtering inconsistent entities. Furthermore, another case study highlights how transformation-based approaches are able to correct preliminary entities using the knowledge about the occurring consistencies.
The approaches of this work based on machine learning rely on Conditional Random Fields, popular probabilistic graphical models for sequence labeling. They take advantage of a consistency model, which is automatically induced during processing the document. The approach based on stacked graphical models utilizes the learnt descriptions as feature functions that have a static meaning for the model, but change their actual function for each document. The other two models extend the graph structure with additional factors dependent on the learnt model of consistency. They include feature functions for consistent and inconsistent entities as well as for additional positions that fulfill the consistencies.
The presented approaches are evaluated in three real-world domains: segmentation of scientific references, template extraction in curricula vitae, and identification and categorization of sections in clinical discharge letters. They are able to achieve remarkable results and provide an error reduction of up to 30% compared to usually applied techniques.
Deep Learning (DL) models are trained on a downstream task by feeding (potentially preprocessed) input data through a trainable Neural Network (NN) and updating its parameters to minimize the loss function between the predicted and the desired output. While this general framework has mainly remained unchanged over the years, the architectures of the trainable models have greatly evolved. Even though it is undoubtedly important to choose the right architecture, we argue that it is also beneficial to develop methods that address other components of the training process. We hypothesize that utilizing domain knowledge can be helpful to improve DL models in terms of performance and/or efficiency. Such model-agnostic methods can be applied to any existing or future architecture. Furthermore, the black box nature of DL models motivates the development of techniques to understand their inner workings. Considering the rapid advancement of DL architectures, it is again crucial to develop model-agnostic methods.
In this thesis, we explore six principles that incorporate domain knowledge to understand or improve models. They are applied either on the input or output side of the trainable model. Each principle is applied to at least two DL tasks, leading to task-specific implementations. To understand DL models, we propose to use Generated Input Data coming from a controllable generation process requiring knowledge about the data properties. This way, we can understand the model’s behavior by analyzing how it changes when one specific high-level input feature changes in the generated data. On the output side, Gradient-Based Attribution methods create a gradient at the end of the NN and then propagate it back to the input, indicating which low-level input features have a large influence on the model’s prediction. The resulting input features can be interpreted by humans using domain knowledge.
To improve the trainable model in terms of downstream performance, data and compute efficiency, or robustness to unwanted features, we explore principles that each address one of the training components besides the trainable model. Input Masking and Augmentation directly modifies the training input data, integrating knowledge about the data and its impact on the model’s output. We also explore the use of Feature Extraction using Pretrained Multimodal Models which can be seen as a beneficial preprocessing step to extract useful features. When no training data is available for the downstream task, using such features and domain knowledge expressed in other modalities can result in a Zero-Shot Learning (ZSL) setting, completely eliminating the trainable model. The Weak Label Generation principle produces new desired outputs using knowledge about the labels, giving either a good pretraining or even exclusive training dataset to solve the downstream task. Finally, improving and choosing the right Loss Function is another principle we explore in this thesis. Here, we enrich existing loss functions with knowledge about label interactions or utilize and combine multiple task-specific loss functions in a multitask setting.
We apply the principles to classification, regression, and representation tasks as well as to image and text modalities. We propose, apply, and evaluate existing and novel methods to understand and improve the model. Overall, this thesis introduces and evaluates methods that complement the development and choice of DL model architectures.
Active galactic nuclei (AGN) are among the brightest and most frequent sources on the extragalactic X-ray and gamma-ray sky. Their central supermassive blackhole generates an enormous luminostiy through accretion of the surrounding gas. A few AGN harbor highly collimated, powerful jets in which are observed across the entire electromagnetic spectrum. If their jet axis is seen in a small angle to our line-of-sight (these objects are then called blazars) jet emission can outshine any other emission component from the system. Synchrotron emission from electrons and positrons clearly prove the existence of a relativistic leptonic component in the jet plasma. But until today, it is still an open question whether heavier particles, especially protons, are accelerated as well. If this is the case, AGN would be prime candidates for extragalactic PeV neutrino sources that are observed on Earth. Characteristic signatures for protons can be hidden in the variable high-energy emission of these objects. In this thesis I investigated the broadband emission, particularly the high-energy X-ray and gamma-ray emission of jetted AGN to address open questions regarding the particle acceleration and particle content of AGN jets, or the evolutionary state of the AGN itself. For this purpose I analyzed various multiwavelength observations from optical to gamma-rays over a period of time using a combination of state-of-the-art spectroscopy and timing analysis. By nature, AGN are highly variable. Time-resolved spectral analysis provided a new dynamic view of these sources which helped to determine distinct emission processes that are difficult to disentangle from spectral or timing methods alone.
Firstly, this thesis tackles the problem of source classification in order to facilitate the search for interesting sources in large data archives and characterize new transient sources. I use spectral and timing analysis methods and supervised machine learning algorithms to design an automated source classification pipeline. The test and training sample were based on the third XMM-Newton point source catalog (3XMM-DR6). The set of input features for the machine learning algorithm was derived from an automated spectral modeling of all sources in the 3XMM-DR6, summing up to 137200 individual detections. The spectral features were complemented by results of a basic timing analysis as well as multiwavelength information provided by catalog cross-matches. The training of the algorithm and application to a test sample showed that the definition of the training sample was crucial: Despite oversampling minority source types with synthetic data to balance out the training sample, the algorithm preferably predicted majority source types for unclassified objects. In general, the training process showed that the combination of spectral, timing and multiwavelength features performed best with the lowest misclassification rate of \\sim2.4\\%.
The methods of time-resolved spectroscopy was then used in two studies to investigate the properties of two individual AGN, Mrk 421 and PKS 2004-447, in detail. Both objects belong to the class of gamma-ray emitting AGN. A very elusive sub-class are gamma-ray emitting Narrow Line Seyfert 1 (gNLS1) galaxies. These sources have been discovered as gamma-ray sources only recently in 2010 and a connection to young radio galaxies especially compact steep spectrum (CSS) radio sources has been proposed. The only gNLS1 on the Southern Hemisphere so far is PKS2004-447 which lies at the lower end of the luminosity distribution of gNLS1. The source is part of the TANAMI VLBI program and is regularly monitored at radio frequencies. In this thesis, I presented and analyzed data from a dedicated multiwavelength campaign of PKS 2004-447 which I and my collaborators performed during 2012 and which was complemented by individual observations between 2013 and 2016. I focussed on the detailed analysis of the X-ray emission and a first analysis of its broadband spectrum from radio to gamma-rays. Thanks to the dynamic SED I could show that earlier studies misinterpreted the optical spectrum of the source which had led to an underestimation of the high-energy emission and had ignited a discussion on the source class. I show that the overall spectral properties are consistent with dominating jet emission comprised of synchrotron radiation and inverse Compton scattering from accelerated leptons. The broadband emission is very similar to typical examples of a certain type of blazars (flat-spectrum radio quasars) and does not present any unusual properties in comparison. Interestingly, the VLBI data showed a compact jet structure and a steep radio spectrum consistent with a compact steep spectrum source. This classified PKS 2004-447 as a young radio galaxy, in which the jet is still developing.
The investigation of Mrk 421 introduced the blazar monitoring program which I and collaborator have started in 2014. By observing a blazar simultaneously from optical, X-ray and gamma-ray bands during a VHE outbursts, the program aims at providing extraordinary data sets to allow for the generation of a series of dynamical SEDs of high spectral and temporal resolution. The program makes use of the dense VHE monitoring by the FACT telescope. So far, there are three sources in our sample that we have been monitoring since 2014. I presented the data and the first analysis of one of the brightest and most variable blazar, Mrk 421, which had a moderate outbreak in 2015 and triggered our program for the first time. With spectral timing analysis, I confirmed a tight correlation between the X-ray and TeV energy bands, which indicated that these jet emission components are causally connected. I discovered that the variations of the optical band were both correlated and anti-correlated with the high-energy emission, which suggested an independent emission component. Furthermore, the dynamic SEDs showed two different flaring behaviors, which differed in the presence or lack of a peak shift of the low-energy emission hump. These results further supported the hypothesis that more than one emission region contributed to the broadband emission of Mrk 421 during the observations.
Overall,the studies presented in this thesis demonstrated that time-resolved spectroscopy is a powerful tool to classify both source types and emission processes of astronomical objects, especially relativistic jets in AGN, and thus provide a deeper understanding and new insights of their physics and properties.
Deep learning enables enormous progress in many computer vision-related tasks. Artificial Intel- ligence (AI) steadily yields new state-of-the-art results in the field of detection and classification. Thereby AI performance equals or exceeds human performance. Those achievements impacted many domains, including medical applications.
One particular field of medical applications is gastroenterology. In gastroenterology, machine learning algorithms are used to assist examiners during interventions. One of the most critical concerns for gastroenterologists is the development of Colorectal Cancer (CRC), which is one of the leading causes of cancer-related deaths worldwide. Detecting polyps in screening colonoscopies is the essential procedure to prevent CRC. Thereby, the gastroenterologist uses an endoscope to screen the whole colon to find polyps during a colonoscopy. Polyps are mucosal growths that can vary in severity.
This thesis supports gastroenterologists in their examinations with automated detection and clas- sification systems for polyps. The main contribution is a real-time polyp detection system. This system is ready to be installed in any gastroenterology practice worldwide using open-source soft- ware. The system achieves state-of-the-art detection results and is currently evaluated in a clinical trial in four different centers in Germany.
The thesis presents two additional key contributions: One is a polyp detection system with ex- tended vision tested in an animal trial. Polyps often hide behind folds or in uninvestigated areas. Therefore, the polyp detection system with extended vision uses an endoscope assisted by two additional cameras to see behind those folds. If a polyp is detected, the endoscopist receives a vi- sual signal. While the detection system handles the additional two camera inputs, the endoscopist focuses on the main camera as usual.
The second one are two polyp classification models, one for the classification based on shape (Paris) and the other on surface and texture (NBI International Colorectal Endoscopic (NICE) classification). Both classifications help the endoscopist with the treatment of and the decisions about the detected polyp.
The key algorithms of the thesis achieve state-of-the-art performance. Outstandingly, the polyp detection system tested on a highly demanding video data set shows an F1 score of 90.25 % while working in real-time. The results exceed all real-time systems in the literature. Furthermore, the first preliminary results of the clinical trial of the polyp detection system suggest a high Adenoma Detection Rate (ADR). In the preliminary study, all polyps were detected by the polyp detection system, and the system achieved a high usability score of 96.3 (max 100). The Paris classification model achieved an F1 score of 89.35 % which is state-of-the-art. The NICE classification model achieved an F1 score of 81.13 %.
Furthermore, a large data set for polyp detection and classification was created during this thesis. Therefore a fast and robust annotation system called Fast Colonoscopy Annotation Tool (FastCAT) was developed. The system simplifies the annotation process for gastroenterologists. Thereby the
i
gastroenterologists only annotate key parts of the endoscopic video. Afterward, those video parts are pre-labeled by a polyp detection AI to speed up the process. After the AI has pre-labeled the frames, non-experts correct and finish the annotation. This annotation process is fast and ensures high quality. FastCAT reduces the overall workload of the gastroenterologist on average by a factor of 20 compared to an open-source state-of-art annotation tool.