@phdthesis{Kobs2024,
  author    = {Kobs, Konstantin},
  title     = {Think outside the Black Box: Model-Agnostic Deep Learning with Domain Knowledge},
  doi       = {10.25972/OPUS-34968},
  url       = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-349689},
  school      = {Universit{\"a}t W{\"u}rzburg},
  year      = {2024},
  abstract  = {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.},
  subject      = {Deep learning},
  language  = {en}
}
@phdthesis{Steininger2023,
  author    = {Steininger, Michael},
  title     = {Deep Learning for Geospatial Environmental Regression},
  doi       = {10.25972/OPUS-31312},
  url       = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-313121},
  school      = {Universit{\"a}t W{\"u}rzburg},
  year      = {2023},
  abstract  = {Environmental issues have emerged especially since humans burned fossil fuels, which led to air pollution and climate change that harm the environment. These issues' substantial consequences evoked strong efforts towards assessing the state of our environment. Various environmental machine learning (ML) tasks aid these efforts. These tasks concern environmental data but are common ML tasks otherwise, i.e., datasets are split (training, validatition, test), hyperparameters are optimized on validation data, and test set metrics measure a model's generalizability. This work focuses on the following environmental ML tasks: Regarding air pollution, land use regression (LUR) estimates air pollutant concentrations at locations where no measurements are available based on measured locations and each location's land use (e.g., industry, streets). For LUR, this work uses data from London (modeled) and Zurich (measured). Concerning climate change, a common ML task is model output statistics (MOS), where a climate model's output for a study area is altered to better fit Earth observations and provide more accurate climate data. This work uses the regional climate model (RCM) REMO and Earth observations from the E-OBS dataset for MOS. Another task regarding climate is grain size distribution interpolation where soil properties at locations without measurements are estimated based on the few measured locations. This can provide climate models with soil information, that is important for hydrology. For this task, data from Lower Franconia is used. Such environmental ML tasks commonly have a number of properties: (i) geospatiality, i.e., their data refers to locations relative to the Earth's surface. (ii) The environmental variables to estimate or predict are usually continuous. (iii) Data can be imbalanced due to relatively rare extreme events (e.g., extreme precipitation). (iv) Multiple related potential target variables can be available per location, since measurement devices often contain different sensors. (v) Labels are spatially often only sparsely available since conducting measurements at all locations of interest is usually infeasible. These properties present challenges but also opportunities when designing ML methods for such tasks. In the past, environmental ML tasks have been tackled with conventional ML methods, such as linear regression or random forests (RFs). However, the field of ML has made tremendous leaps beyond these classic models through deep learning (DL). In DL, models use multiple layers of neurons, producing increasingly higher-level feature representations with growing layer depth. DL has made previously infeasible ML tasks feasible, improved the performance for many tasks in comparison to existing ML models significantly, and eliminated the need for manual feature engineering in some domains due to its ability to learn features from raw data. To harness these advantages for environmental domains it is promising to develop novel DL methods for environmental ML tasks. This thesis presents methods for dealing with special challenges and exploiting opportunities inherent to environmental ML tasks in conjunction with DL. To this end, the proposed methods explore the following techniques: (i) Convolutions as in convolutional neural networks (CNNs) to exploit reoccurring spatial patterns in geospatial data. (ii) Posing the problems as regression tasks to estimate the continuous variables. (iii) Density-based weighting to improve estimation performance for rare and extreme events. (iv) Multi-task learning to make use of multiple related target variables. (v) Semi-supervised learning to cope with label sparsity. Using these techniques, this thesis considers four research questions: (i) Can air pollution be estimated without manual feature engineering? This is answered positively by the introduction of the CNN-based LUR model MapLUR as well as the off-the-shelf LUR solution OpenLUR. (ii) Can colocated pollution data improve spatial air pollution models? Multi-task learning for LUR is developed for this, showing potential for improvements with colocated data. (iii) Can DL models improve the quality of climate model outputs? The proposed DL climate MOS architecture ConvMOS demonstrates this. Additionally, semi-supervised training of multilayer perceptrons (MLPs) for grain size distribution interpolation is presented, which can provide improved input data. (iv) Can DL models be taught to better estimate climate extremes? To this end, density-based weighting for imbalanced regression (DenseLoss) is proposed and applied to the DL architecture ConvMOS, improving climate extremes estimation. These methods show how especially DL techniques can be developed for environmental ML tasks with their special characteristics in mind. This allows for better models than previously possible with conventional ML, leading to more accurate assessment and better understanding of the state of our environment.},
  subject      = {Deep learning},
  language  = {en}
}
@phdthesis{Krenzer2023,
  author    = {Krenzer, Adrian},
  title     = {Machine learning to support physicians in endoscopic examinations with a focus on automatic polyp detection in images and videos},
  doi       = {10.25972/OPUS-31911},
  url       = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-319119},
  school      = {Universit{\"a}t W{\"u}rzburg},
  year      = {2023},
  abstract  = {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.},
  subject      = {Deep Learning},
  language  = {en}
}
@phdthesis{Grohmann2022,
  author    = {Grohmann, Johannes Sebastian},
  title     = {Model Learning for Performance Prediction of Cloud-native Microservice Applications},
  doi       = {10.25972/OPUS-26160},
  url       = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-261608},
  school      = {Universit{\"a}t W{\"u}rzburg},
  year      = {2022},
  abstract  = {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.},
  subject      = {Cloud Computing},
  language  = {en}
}
@phdthesis{Niebler2019,
  author    = {Niebler, Thomas},
  title     = {Extracting and Learning Semantics from Social Web Data},
  doi       = {10.25972/OPUS-17866},
  url       = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-178666},
  school      = {Universit{\"a}t W{\"u}rzburg},
  year      = {2019},
  abstract  = {Making machines understand natural language is a dream of mankind that existed since a very long time. Early attempts at programming machines to converse with humans in a supposedly intelligent way with humans relied on phrase lists and simple keyword matching. However, such approaches cannot provide semantically adequate answers, as they do not consider the specific meaning of the conversation. Thus, if we want to enable machines to actually understand language, we need to be able to access semantically relevant background knowledge. For this, it is possible to query so-called ontologies, which are large networks containing knowledge about real-world entities and their semantic relations. However, creating such ontologies is a tedious task, as often extensive expert knowledge is required. Thus, we need to find ways to automatically construct and update ontologies that fit human intuition of semantics and semantic relations. More specifically, we need to determine semantic entities and find relations between them. While this is usually done on large corpora of unstructured text, previous work has shown that we can at least facilitate the first issue of extracting entities by considering special data such as tagging data or human navigational paths. Here, we do not need to detect the actual semantic entities, as they are already provided because of the way those data are collected. Thus we can mainly focus on the problem of assessing the degree of semantic relatedness between tags or web pages. However, there exist several issues which need to be overcome, if we want to approximate human intuition of semantic relatedness. For this, it is necessary to represent words and concepts in a way that allows easy and highly precise semantic characterization. This also largely depends on the quality of data from which these representations are constructed. In this thesis, we extract semantic information from both tagging data created by users of social tagging systems and human navigation data in different semantic-driven social web systems. Our main goal is to construct high quality and robust vector representations of words which can the be used to measure the relatedness of semantic concepts. First, we show that navigation in the social media systems Wikipedia and BibSonomy is driven by a semantic component. After this, we discuss and extend methods to model the semantic information in tagging data as low-dimensional vectors. Furthermore, we show that tagging pragmatics influences different facets of tagging semantics. We then investigate the usefulness of human navigational paths in several different settings on Wikipedia and BibSonomy for measuring semantic relatedness. Finally, we propose a metric-learning based algorithm in adapt pre-trained word embeddings to datasets containing human judgment of semantic relatedness. This work contributes to the field of studying semantic relatedness between words by proposing methods to extract semantic relatedness from web navigation, learn highquality and low-dimensional word representations from tagging data, and to learn semantic relatedness from any kind of vector representation by exploiting human feedback. Applications first and foremest lie in ontology learning for the Semantic Web, but also semantic search or query expansion.},
  subject      = {Semantik},
  language  = {en}
}
@phdthesis{Kluegl2015,
  author    = {Kl{\"u}gl, Peter},
  title     = {Context-specific Consistencies in Information Extraction: Rule-based and Probabilistic Approaches},
  publisher = {W{\"u}rzburg University Press},
  address   = {W{\"u}rzburg},
  isbn      = {978-3-95826-018-4 (print)},
  doi       = {10.25972/WUP-978-3-95826-019-1},
  url       = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-108352},
  school      = {W{\"u}rzburg University Press},
  year      = {2015},
  abstract  = {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.},
  subject      = {Information Extraction},
  language  = {en}
}