@phdthesis{rieger2026_116940,
  title = {Integrating Domain Knowledge for Robust and Interpretable Deep Neural Networks in Facial Affect Recognition},
  author = {Rieger, Ines Sophia},
  year = {2026},
  publisher = {Otto-Friedrich-Universität},
  address = {Bamberg},
  doi = {10.20378/irb-116940},
  urn = {urn:nbn:de:bvb:473-irb-116940x},
  url = {https://fis.uni-bamberg.de/handle/uniba/116940},
  language = {eng},
  abstract = {Facial affect recognition is a key component of human-centered AI, enabling systems to respond appropriately to human nonverbal signals. The deployment of deep learning for facial affect recognition is critically hindered by vulnerability to data-driven biases and a lack of transparency. This thesis addresses these challenges through a multi-faceted framework, structured along the CRISP-ML(Q) machine learning lifecycle, integrating domain knowledge to build more robust and interpretable systems.

First, it quantifies demographic bias through extensive subject-level metadata annotation, uncovering performance gaps across demographic groups in leading facial expression recognition (FER) classifiers. To mitigate demographic and multi-label class imbalance, an optimization-based data balancing technique, MLOPT, is introduced to improve performance, fairness, and generalization. Second, the training of deep learning models is guided by injecting anatomical knowledge about facial Action Units (AUs), that is, localized facial muscle activations defined by the Facial Action Coding System (FACS). This is achieved through the CorrLoss regularization term and a Facial Movement Constraint Network (FMC-Net), which together improve model robustness and predictive performance. Third, interpretability is enhanced through a fuzzy-logic ensemble that combines expert rules with model predictions. This dual approach enables self-validation of the models by flagging uncertain outputs while providing inherently interpretable behavior through logical rules. Finally, a new evaluation framework, RelEx, is proposed to assess the physiological plausibility of trained models by quantifying the alignment between feature attributions and expert-defined facial regions.

By addressing these challenges with targeted solutions, this research demonstrates that systematically integrating domain knowledge moves FER beyond  performance optimization. The resulting framework provides a foundation for developing AI systems that are not only high-performing but also robust and interpretable.},
  keywords = {Facial Affect Recognition, Artificial Intelligence, Deep Neural Networks, Robustness, Interpretability, Domain Knowledge},
  school = {Otto-Friedrich-Universität Bamberg}
}

@phdthesis{wehner2026_115564,
  title = {Explainable and Interactive Link Prediction in Knowledge Graphs},
  author = {Wehner, Christoph},
  year = {2026},
  publisher = {Otto-Friedrich-Universität},
  address = {Bamberg},
  doi = {10.20378/irb-115564},
  urn = {urn:nbn:de:bvb:473-irb-115564x},
  url = {https://fis.uni-bamberg.de/handle/uniba/115564},
  language = {eng},
  abstract = {Machine learning models move from the safe walls of research labs to high-risk real-world applications. In these environments, users need to understand why a machine learning model behaves as it does and correct it if necessary. One critical application domain is Root Cause Analysis in manufacturing, where data-driven models identify fault causes in time-critical environments. Link prediction models in Knowledge Graphs can be used to tackle this task. However, Knowledge Graph Completion models, particularly embedding-based link predictors, often operate as black-boxes. This limits human oversight, complicates validation, and undermines trust.
This thesis addresses the challenge of making link prediction models correct for the correct reasons with the help of human feedback. Specifically, it proposes a technical and conceptual pipeline in which model explanations serve as interfaces for collecting and incorporating user feedback to iteratively align model behaviour with human intent.
Three research questions are addressed in this work:
(1) How can the decision-making process of a link prediction model be explained in ways that support downstream alignment?
(2) What types of user feedback on explanations are effective for guiding model updates?
(3) How can feedback on model explanations be integrated into the training and inference processes of link prediction models?
In response, three major contributions are made in this thesis:
First, KGEPrisma is introduced, a post-hoc explanation method for embedding-based link prediction models. KGEPrisma translates latent embeddings into instance-based, analogy-based, and rule-based explanations. This method provides a multi-perspective view on model decisions and establishes a foundation for the subsequent exploration of feedback modalities.
Second, RootFinder is presented, a system that treats Root Cause Analysis as a link prediction task in a manufacturing Knowledge Graph. RootFinder enables domain experts to assess model explanations and to provide simple, customised feedback for real-world settings. This application motivates an in-depth discussion of feedback modalities.
Finally, LiEr is introduced in the thesis. LiEr is a reinforcement learning approach that iteratively aligns link prediction with preference-based feedback over model explanations. LiEr reformulates Knowledge Graph Completion as a Markov Decision Process and incorporates preference-based user feedback into the reward function. This enables the model to learn policies that converge toward human-plausible reasoning. LiEr is applied to Root Cause Analysis in electric vehicle manufacturing. The case study confirms that preference-based feedback over explanations enables root cause prediction alignment with domain knowledge.
Together, these contributions provide a methodological and empirical foundation for using explanations to collect human feedback and align the behaviour of link prediction models with it. Explanations enable an interactive learning loop. This thesis presents the first work on interactive and explainable Knowledge Graph Completion.},
  keywords = {Knowledge Graphs, Explainable AI, Link Prediction, Human-in-the-Loop},
  school = {Otto-Friedrich-Universität Bamberg}
}

@phdthesis{wiing2026_114288,
  title = {Efficient machine learning for sensor systems with an application to computational spectrometers},
  author = {Wißing, Julio},
  year = {2026},
  publisher = {University of Bamberg Press},
  address = {Bamberg},
  series = {Schriften aus der Fakultät Wirtschaftsinformatik und Angewandte Informatik der Otto-Friedrich-Universität Bamberg},
  isbn = {978-3-98989-108-1},
  doi = {10.20378/irb-114288},
  urn = {urn:nbn:de:bvb:473-irb-114288x},
  url = {https://fis.uni-bamberg.de/handle/uniba/114288},
  language = {eng},
  abstract = {This thesis explores two distinct approaches to optimizing machine learning for resource-constrained environments: hierarchical machine learning (HiML) for energy-efficient classification and neural network-based spectral reconstruction for cost-effective computational spectrometers. Both topics address different challenges but share the common goal of improving computational efficiency for embedded systems. The first part investigates hierarchical machine learning as a means to reduce energy consumption in classification tasks. Unlike traditional flat classifiers, which apply equal computational effort to all inputs, hierarchical machine learning structures classification as a hierarchy of decisions, allowing for early exits when classification confidence is high. Experimental results demonstrate that this approach reduces energy consumption by up to 47.63\% while maintaining competitive accuracy. Additionally, reinforcement learning is introduced as an optimization strategy for classifier selection, significantly accelerating the search process compared to exhaustive methods. The second part focuses on spectral reconstruction, where neural networks recover full spectral information from complex sensor outputs. A physics-informed data augmentation method is introduced, leveraging the optical properties of transmission spectra to expand the dataset from 214 to over 10,000 samples. This augmentation improves model generalization without requiring additional physical measurements. Furthermore, layer-wise relevance propagation is applied to optimize the sensor hardware, enabling a 94\% reduction in input dimensions while preserving high reconstruction accuracy. Benchmarking results reveal that smaller, well-optimized models can match or even surpass the performance of larger networks, demonstrating the potential for efficient, compact spectral reconstruction systems. The findings of this thesis contribute to the development of energy-efficient and computationally lightweight machine learning solutions. The proposed approaches enable more effective classification pipelines and high-fidelity spectral reconstruction on embedded systems, paving the way for real-world applications in industrial monitoring, portable spectral sensing, and other edge AI scenarios.},
  keywords = {hierarchisches Maschinelles Lernen, Hierarchical Machine Learning, Spektrale Rekonstruktion, Eingebettete Systeme, Edge AI, Energieeffiziente Klassifikation, Computational Spectrometer, Spectral Reconstruction, Embedded Systems, Energy-Efficient Classification},
  school = {Otto-Friedrich-Universität Bamberg}
}

@phdthesis{slany2025_107376,
  title = {Enhancing Explanatory Interactive Machine Learning : A Generalization of the CAIPI Algorithm},
  author = {Slany, Emanuel},
  year = {2025},
  publisher = {Otto-Friedrich-Universität},
  address = {Bamberg},
  doi = {10.20378/irb-107376},
  urn = {urn:nbn:de:bvb:473-irb-1073760},
  url = {https://fis.uni-bamberg.de/handle/uniba/107376},
  language = {eng},
  abstract = {An inherent challenge of statistical machine learning models is the possibility of erroneous outcomes, which underscores the importance of corrigibility, particularly in sensitive domains. Advances in the field of explainable Artificial Intelligence have revealed that machine learning models might conduct correct decisions based on incorrect decision-making mechanisms. The possibility of detecting erroneous decision making does not prevent incorrect decision-making mechanisms, which is why the research field explanatory interactive machine learning has equipped researchers with the opportunity to revise explanations. Explanatory interactive machine learning procedures vary in their feedback injection mechanism: Some utilize model-specific internals, e.g., they penalize gradients in indecisive regions. Others are model-agnostic and induce additional training data, termed counterexamples, which only contain the relation of human explanation revisions and the target. CAIPI, the origin of the latter category, is an algorithmic framework to optimize machine learning models that iteratively combines components to predict, explain, and interact with instances. It generates counterexamples in iterations with correct prediction and erroneous decision-making mechanism. CAIPI, despite being claimed to be model-agnostic, has been proposed as a theoretical concept, which does not transfer well to practical application scenarios beyond image and text classification. Moreover, its formalization requires refinement, particularly when evaluating the impact of counterexamples on the optimization framework. This thesis aims to address CAIPI’s limitations through a formal analysis of how counterexamples influence the iterative optimization of machine learning models, together with algorithmic modifications of the original CAIPI framework.},
  keywords = {Explanatory Interactive Machine Learning, Explainable Artificial Intelligence, Interactive Learning, Machine Learning, CAIPI},
  school = {Otto-Friedrich-Universität Bamberg}
}

@phdthesis{rabold2024_104269,
  title = {Relational Explanations for Visual Domains : A Neural-symbolic Approach Combining ILP and CNNs},
  author = {Rabold, Johannes},
  year = {2024},
  publisher = {Otto-Friedrich-Universität},
  address = {Bamberg},
  doi = {10.20378/irb-104269},
  urn = {urn:nbn:de:bvb:473-irb-1042696},
  url = {https://fis.uni-bamberg.de/handle/uniba/104269},
  language = {eng},
  abstract = {With powerful machine learning methods leaving the lab, the need for transparency in automatic decision processes becomes apparent. Only if humans have the ability to scrutinize how a model behaves and what the rationale behind a decision was, they will gain trust in a system. However, current neural network architectures are mainly black-box, so it is not easily possible to comprehend model decisions. Symbolic machine learning approaches that are inherently interpretable already exist for a long time in artificial intelligence research. However, most of these are not actually suitable for real world applications, since they lack the ability to work with raw (image) data, and thus, an accuracy-interpretability trade-off emerges.

The research branch of eXplainable Artificial Intelligence (XAI) advises interpretable surrogate models as a possible solution: Instead of abandoning a powerful model, it is kept as is and an additional interpretable model is generated, mimicking the behavior of the black-box. The gold standard for explaining image processing approaches like convolutional neural networks are visual attribution methods. For an image instance, methods like LIME or Grad-CAM output heatmaps, indicating regions that were influential (in a positive or negative way) for a particular model decision. These methods can give a first idea, what constituents in an image are important, and can pinpoint flaws in the trained model that can stem from e.g. bias in the training data. However, they lack expressiveness and can obscure the importance of e.g. relations that hold between image parts. This becomes particularly important for relational domains, where classification depends not only on the presence of image parts, but also their (spatial) constellation.

This thesis will describe and analyze methods of generating expressive symbolic relational explanations in the form of first-order logic rules. These rules inherently highlight the importance of not only visual concepts, but also of the relations between them. They also can be easily interpreted by humans and can even be converted to natural language in a straight-forward fashion. This work gives approaches for generating explanations in a variety of problem settings: Not only is it possible to explain the decision for a single image, it is also possible to explain the model as a whole. Also, when access to model parameters is given, explanations can benefit from that.

In an attempt to holistically approach relational explanation generation, this work will look at methods on how to be as close to the behavior of the original model as possible and also on how to quantify this "fidelity". Additionally, this work will revisit the generated explanations and ask, what types of explanations are particularly useful for humans.},
  keywords = {Explainable Machine Learning, Artificial Intelligence, Convolutional Neural Networks, Inductive Logic Programming, Neural-symbolic},
  school = {Otto-Friedrich-Universität Bamberg}
}

@phdthesis{neufeld2024_95413,
  title = {Supporting Experts in Detecting and Interpreting Anomalies in Time Series : Exploring Data Science Approaches for the Monitoring of Hydraulic Test Benches},
  author = {Neufeld, Deniz},
  year = {2024},
  publisher = {Otto-Friedrich-Universität},
  address = {Bamberg},
  doi = {10.20378/irb-95413},
  urn = {urn:nbn:de:bvb:473-irb-954138},
  url = {https://fis.uni-bamberg.de/handle/uniba/95413},
  language = {eng},
  abstract = {Hydraulic systems are important in the functioning of everyday life, as well as critical infrastructure by driving various systems such as automotive, aircrafts, and construction machines. To assure their functionality according to specification over their complete lifetime, reliability tests are conducted in test benches to check a representative number of samples for long amounts of time. During the tests, physical system inputs and outputs such as pressures, electric currents and voltages are recorded for analysis. The goal of this thesis is to research anomaly detection methods based on the recorded data to support root cause analysis and reduce testing time. Since the first decision makers in case of an error are the responsible engineers and technicians, it is important to focus on methods that are both robust and easily understandable for non-experts of data science. 

Several challenges stem from the given problem. With the start of a test bench, there is often no prior measurement data available for model training, and samples are not repeated after a successful test run. Furthermore, there are various kinds of signals with different properties, from digital bus system signals to fluid flow or electrical currents. If a complete system is assessed, different failure modes can occur based on different sub-components. This leads to many possible failure constellations and therefore a large number of relevant features.

There are multiple fields of research which are related to this problem, ranging from the general field of time series analysis to more specific condition monitoring research in the domain of engineering. This means there exist similar problem types in state of the art, but also limited closely related work. The following work not only focuses on research on the anomaly detection on hydraulic test benches per se, but also investigates data visualization, anomaly detection in physical systems, multivariate time series anomaly detection, and vibration analysis.

The original contribution to knowledge of this thesis are advances in multiple aspects of anomaly detection with a focus on hydraulic test benches. First a collection of methods for the visualization of multivariate, repeating time series is provided. It supports engineers in viewing data and detecting anomalies visually by aligning the signals along the time and the amplitude axis. From this, this thesis examines a model-based method for a digital twin of the tested system using recurrent neural networks. Challenges with this approach are shown and their root cause described in depth. Additionally, data-based methods are examined: One, an unsupervised, statistical method is developed for anomaly detection on multivariate, periodical data in the time domain, as well as its robustness and limits of this method towards concept drift are investigated. It is further shown how the results of the method can be visualized for root-cause analysis. Second, a supervised approach is followed on vibrational data using convolutional neural networks (\acrshort\{cnn\}s). For this, preprocessing in the frequency-domain its influence on model performance is researched. Due to the black box nature of \acrshort\{cnn\}s, an explainable artificial intelligence method is developed to make the relevant features of the data in the frequency domain interpretable for engineers and system experts. The XAI method is verified in quantitatively using an accessible data set specifically designed for this task. 

Based on the shown research, this thesis presents possibilities for future work, specifically with a focus on the enormous amounts of data collected. During a test bench run, millions of time series can be collected, but the visualization and anomaly detection methods developed are not yet perfected towards this fact. Due to the size of the data, dealing with the data and the results of the anomaly detection methods must become more efficient as well, be it for displaying, labeling or for judging the output of the algorithms. While the methods shown can be used on subsets of the data, an important aspect of future work is the clustering of data to reduce the amount of data to support inspection and labeling. Additionally, the higher-level visualizations with a drill-down functionality need to be developed as well to guide users towards relevant information.},
  keywords = {Anomaly detection, Hydraulic systems, Condition Monitoring},
  school = {Otto-Friedrich-Universität Bamberg}
}

@phdthesis{kiefer2023_89771,
  title = {Human-centered Interactions with Text Classifiers : Fusing Concept-based Knowledge with Local Surrogate Explanation Models},
  author = {Kiefer, Sebastian},
  year = {2023},
  publisher = {Otto-Friedrich-Universität},
  address = {Bamberg},
  doi = {10.20378/irb-89771},
  urn = {urn:nbn:de:bvb:473-irb-897715},
  url = {https://fis.uni-bamberg.de/handle/uniba/89771},
  language = {eng},
  abstract = {Human interactions often take place with the aim of exchanging understanding between individuals. Thus, there is a human need to develop and communicate explanations on the one hand and to receive explanations on the other hand in order to further develop one’s own understanding. Explanations are an attempt to trace events back to their causes, for example, to provide answers to the question of "why".
The concept of bidirectional explanations can also be applied to the interaction between humans and machines, for example, in the context of machine learning (ML), which is a sub-area of artificial intelligence (AI). Whenever machines are used to support human decision-making or even to provide recommendations for action, there is a need to be able to understand how the results were obtained and to influence them if necessary. Human decision-makers should develop trust in machine learners, especially in high-stakes application domains, like in medical diagnostics, but also in application domains that are strongly influenced by regulations, like the domain of financial auditing. The comprehension of concrete decisions and the incorporation of expert knowledge can contribute to the development of trust.
Modern machine learning methods have recently been improved in terms of their predictive accuracy, which has led to them already surpassing human performance in some tasks. On the other hand, such powerful methods are usually so-called black-box-methods, what makes it difficult or even impossible for humans to develop an understanding of the general decision logic or the specific model behavior. Lack of comprehensibility not only impairs the formation of trust in machine companions, but also impairs the ability to interact in the sense of correctability. In comparison to global explanations that explain how a system works, local explanations that provide information about correlations of specific inputs and outputs and thus justify why a particular output was produced are often considered more reliable and purposeful. This fact especially holds true for non-expert AI users.
Researchers in the field of explainable artificial intelligence have developed approaches that enable post hoc model-agnostic and mostly local explanations of supervised machine learners. The goal is to explain individual ML results retrospectively, i.e., after predictions have been made for previously unobserved instances, and independently of the ML model used, i.e., without knowledge about its inner workings. Especially in the case of surrogate explanation models that locally approximate the model to be explained, there is a risk that, when contextual information, such as dependencies between features, is neglected, this will result in data points that are described as "out-of-distribution". Explanations based on such unrealistic data points can easily be misinterpreted. Furthermore, explanation approaches are often developed exclusively from a technical point of view. Insights from psychology or the social sciences, according to which explanations that are understandable for humans should, for example, have a coherent structure, often are not taken into account.
In this dissertation, a new approach is proposed and validated that enables human-centered interactivity with text classifiers by using bidirectional model-agnostic explanations. To this end, a framework is introduced that defines comprehensible and interactive artificial intelligence in an interdisciplinary manner using cognitive concepts such as explainability, interpretability, transparency, and interactivity. It elaborates that semantic and contextual information available within a certain textual application domain should be taken into account during the generation and representation of explanations such that contextual explanations result. In order to fill this identified research gap and obtain coherent explanations, a new technique is presented that generates model-agnostic concept-based explanations whose explanatory features consist of semantically related words. The consideration of context enables more realistic and meaningful local perturbation distributions, which form the basis of many model-agnostic and local explanation approaches, such as LIME. A technical evaluation of the new explanation methodology called topicLIME, in particular of the underlying surrogate models and the resulting explanations, analyzes its local fidelity related to the text classifier to be explained. Besides a technical evaluation, the obtained results are empirically analyzed by means of two user studies. It is investigated how concept-based contextual explanations in comparison to contextless explanations are perceived by humans interacting with a text classifier that predicts to which content category a presented document belongs.
As another contribution of this dissertation, a method is presented that extends state-of-the-art for explanatory interactive learning, especially for text classification. The method intends to enable humans to engage in broader interactions, such as correcting predictions and explanations in a constructive and contextual manner. Previous model-agnostic methods for explanatory interactive learning, such as the CAIPI approach, only allow correcting a classifier regarding its functioning via explanations in case a correct prediction has been made for the wrong reasons. The newly developed Semantic Push approach qualifies humans to perform corrections via concept-based explanations and integrate them into the learner using non-extrapolating training documents, across different error types of a classifier. The evaluation of the approach shows that the newly developed method outperforms the baseline CAIPI method in terms of learning performance and local explanation quality.
In summary, the fusion of concept-based knowledge with local surrogate explanation models is a promising research direction. The results of this dissertation further show that more interdisciplinary research is needed to address the challenges in the field of human-centered machine learning.},
  keywords = {Human-centered Machine Learning, Explainable Artificial Intelligence, Local Surrogate Explanation Models, Model-agnostic Explanations, Explanatory Interactive Machine Learning, Explanatory Understanding, Human-like Explanations, Topic Modeling},
  school = {Otto-Friedrich-Universität Bamberg}
}

@phdthesis{eirich2023_58339,
  title = {The Creation, Formalization, and Transfer of Expert Knowledge with Visual Analytics in Industrial Manufacturing Processes of Electrical Vehicles},
  author = {Eirich, Joscha},
  year = {2023},
  publisher = {Otto-Friedrich-Universität},
  address = {Bamberg},
  doi = {10.20378/irb-58339},
  urn = {urn:nbn:de:bvb:473-irb-583392},
  url = {https://fis.uni-bamberg.de/handle/uniba/58339},
  language = {eng},
  abstract = {The 21st century will be heavily impacted by the capability to create value from recorded data of all kinds. In this regard, industrial manufacturing intuitions increasingly rely on new data-driven technologies, such as the Internet of Things or Machine Learning. In terms of data collection, manufacturing processes increasingly include sophisticated sensor equipment, which results in interconnected networks of manufacturing all assembling parts and producing data. However, manufacturing institutions currently face two challenges. First, large amounts of parts and hence data are produced during fully automated manufacturing processes. Second, due to the overwhelming amount of recorded data, it is particularly challenging to efficiently analyze manufacturing data. Hence, it is important to efficiently store and  share gained knowledge from performed data analyses. Information visualization and Visual Analytics are two prolific branches of data analysis exploiting sophisticated visualization techniques to support the execution of analytical tasks and to store gained knowledge. This thesis looks at how data visualization approaches can help industrial manufacturing organizations to create value from large amounts of manufacturing data, as well as how to efficiently store and share knowledge from manufacturing data analyses. The goal is to understand how Visual Analytics can improve manufacturing processes in the context of knowledge management. Five Visual Analytics systems were designed, developed, and evaluated to tackle different domain problems that emerged from manufacturing setups. Findings from each system were used to carry out additional studies to enhance established theories of knowledge management. As a result, five success stories are provided of how Visual Analytics can significantly improve manufacturing processes and how knowledge can be efficiently created, formalized, and shared in an organization with Visual Analytics.},
  keywords = {Visual Analytics, Design Study, Human Computer Interaction, Knowledge Manamgenet},
  school = {Otto-Friedrich-Universität Bamberg}
}

@phdthesis{schwalbe2022_57172,
  title = {Concept Embedding Analysis Based Methods for the Safety Assurance of Deep Neural Networks : towards safe automotive computer vision applications},
  author = {Schwalbe, Gesina},
  year = {2022},
  publisher = {Otto-Friedrich-Universität},
  address = {Bamberg},
  doi = {10.20378/irb-57172},
  urn = {urn:nbn:de:bvb:473-irb-571725},
  url = {https://fis.uni-bamberg.de/handle/uniba/57172},
  language = {eng},
  abstract = {Deep neural networks (DNNs)  are handled as a key technology for  computer vision (CV) in automated driving and similar safety critical domains. For admission to market, the safe operation of such systems has to be ensured. However,  DNNs  come along with properties like complexity and opaqueness. These make new approaches towards safety argumentation and evidence acquisition inevitable. This thesis both identifies and tackles key practical issues associated with safety assurance of  convolutional deep neural networks (CNNs)  in automotive computer vision applications.
The key practical issue handled here is that many safety requirements origin from symbolic domain knowledge which relate semantic concepts from natural language, such as “eyes usually belong to an obstacle”. That is a problem since outputs of typical object detection CNNs are restricted to few semantic label classes, and both input and intermediate outputs are non-symbolic. This thesis bridges this gap using methods from  concept embedding analysis (CA).  CA  research tries to associate semantic concepts with items in the DNN  intermediate outputs, thus providing access to symbolic knowledge encoded in the model. A baseline  CA  method is chosen according to a broad literature analysis, and, in the course of several experimental studies, substantially improved regarding efficiency and performance. This allowed for the first time to successfully apply  CA  to state-of-the-art CNNs  for object detection.
Based on the improved  CA  method, diverse approaches to provide evidence for different types of symbolic safety requirements are developed. Concretely,  CA  is used (1) to verify correct encoding of semantic relations, (2) to build global, interpretable, and inspectable proxy models, and it is used (3) as part of a framework for inspection and verification of compliance with symbolic fuzzy logic rules. Applicability of the approaches is shown and evaluated on several state-of-the-art object detector  CNNs  and backends of such.
Lastly, the contribution of  CA -based evidence generation to safety assurance is highlighted. A template for a safety argumentation structure is developed, and a broad review of existing  DNN  specific evidence generation methods is conducted. This reveals the need for the  CA -based methods developed in this thesis, and properly positions them in the overall safety argument.
Altogether, the works in this thesis constitute a solid step towards safer usage of  DNNs in automotive  CV applications.},
  keywords = {Deep neural network, Computer Vision, Safety, Automated Driving, Concept Embedding Analysis, Explainable Artificial Intelligence, XAI, Safety Argument, Verification, Object Detection},
  school = {Otto-Friedrich-Universität Bamberg}
}

@mastersthesis{rabold2022_46527,
  title = {Enriching LIME with Inductive Logic Programming: Explaining Deep Learning Classifiers with Logic Rules in a Companion System Framework},
  author = {Rabold, Johannes},
  year = {2022},
  publisher = {Otto-Friedrich-Universität},
  address = {Bamberg},
  doi = {10.20378/irb-46527},
  urn = {urn:nbn:de:bvb:473-irb-465273},
  url = {https://fis.uni-bamberg.de/handle/uniba/46527},
  language = {eng},
  abstract = {With the rise of black-box classifiers like Deep Learning networks, the need for interpretable and complete explanations for them becomes apparent. Users need to have the possibility to ask why a classifier inferred a particular result. Logic clauses induced by Inductive Logic Programming systems are superior in expressibility over visual explanations alone.
This thesis uses the ideas of LIME, a visual explanation framework, and enriches it with an ILP component to get comprehensible and powerful explanations for the inference results of Deep Learning Networks for images. The background knowledge for the predicates is obtained both automatically and by an annotation system that lets humans annotate labels and relations.   The  human  labeling  system  and  the  explanation component form a Companion System where not only AI helps the user but also the other way round.},
  keywords = {Explainable AI, Deep Learning, Inductive Logic Programming, LIME, Companion System, Local Interpretable Model-Agnostic Explanations},
  school = {Otto-Friedrich-Universität Bamberg}
}

@phdthesis{schimanke2022_55317,
  title = {The Effect of using Spaced Repetition in Mobile Learning Games on the Learning Success},
  author = {Schimanke, Florian},
  year = {2022},
  publisher = {Otto-Friedrich-Universität},
  address = {Bamberg},
  doi = {10.20378/irb-55317},
  urn = {urn:nbn:de:bvb:473-irb-553175},
  url = {https://fis.uni-bamberg.de/handle/uniba/55317},
  language = {eng},
  abstract = {The field of learning has been researched in multiple ways and directions. The different learning strategies of individual learners represent one of those fields. Especially in traditional learning environments, such as school or university, learners tend to result-driven learning strategies like massing and cramming, which foremost aim at using the short-term memory to have the necessary information available during a test or an exam. An approach that focuses more on the long-term is called “spacing” and uses sophistically determined intervals between learning sessions that are unique to each learner and are proven to have a positive effect on the sustainable retention of knowledge. However, since spaced repetition is still a repetition-based learning strategy, it is prone to bore the learners by requiring them to repeat the same action over and over again.

One approach to take the boredom out of repetition-based learning is gamification or game-based learning. Using learning games can lead to a higher motivation and thus to a higher probability that the learners continue to use the game and keep on learning through it. Current technologies like smartphones and tablets are capable of handling sophisticated learning games, offer the possibility to use them ubiquitously and therefore act as a facilitator for a learning on demand.

While spaced repetition is not a new field of research, the effect of this learning approach on the learning success has not yet been researched when being used in combination with a mobile learning game. Combining the motivating effect of ubiquitous game-based learning with spaced repetition seems to be a promising combination for an improved learning success. The aim of the presented research is to investigate if using spaced repetition in a mobile learning game leads indeed to the desired success.

To do so, several aspects were considered, including how the SM2 spaced repetition algorithm can be integrated into a mobile learning game, the negative effect that too much motivation can have, how to deal with early and late repetitions, how to conclude automated tests of the implementation, how to collect and filter real-world data to analyze the effectiveness, and how learning analytics can be applied to this analysis. For the data collection, an algorithmic framework, containing the SM2 and the self-developed FS algorithm was integrated into an established geography learning game. The collected data was then filtered and analyzed in accordance with an established definition for learning success. For this, a test group with players who played at least one category from the learning game strictly according to the spaced repetition approach and a control group with players who played the game at more or less random times were built from the collected data. Their progress throughout the observation period was compared and a conclusion was drawn from this.

The two main indicators for the effect on the learning success were defined as the “Average Base Knowledge” (ABK) and the “Average Retention Score” (ARS). The ABK is the average number of correct answers during the first game in both groups, while the ARS is the average correctness during the so-called "leaerning-games", which are the games from the 3rd to the nth game. While both groups started with a comparable ABK of 0.5, the test group already achieved an almost perfect ARS for the fourth game (i.e. the third repetition) of 0.95. The control group on the other hand reached an ARS of 0.63 for the same repetition without the interval calculation by the SM2 algorithm. This effect continues during the following repetitions, during which the ARS for the test group stayed between 0.98 and 1.0, while the control group reaches a maximum ARS of 0.87 after 14 games, which marked the end of the observation period. Further structure exploratory analyses, such as a cluster analysis and a logistic regression confirmed the findings from the comparison of the development of the ABK and the ARS.

Summarized, the analyses showed that there is a considerable difference between the two groups. Foremost, while the retention improved in both groups, the test group achieved a quicker improvement and a lasting retention of knowledge over time, which is matches exactly the definition that was established for learning success in the presented research.},
  keywords = {spaced repetition, game-based learning, learning strategies, learning analytics},
  school = {Otto-Friedrich-Universität Bamberg}
}

@phdthesis{seu2021_49932,
  title = {Exploiting Domain-specific Knowledge for Classifier Learning : AU-based Facial Expression Analysis and Emotion Recognition},
  author = {Seuß, Dominik},
  year = {2021},
  publisher = {Otto-Friedrich-Universität},
  address = {Bamberg},
  doi = {10.20378/irb-49932},
  urn = {urn:nbn:de:bvb:473-irb-499326},
  url = {https://fis.uni-bamberg.de/handle/uniba/49932},
  language = {eng},
  abstract = {Gesichtsausdrücke sind einer der wichtigsten Kanäle menschlicher nonverbaler Kommunikation. Sie lassen Rückschlüsse auf unseren mentalen Zustand zu und gelten als universell verständlich über Ethnien hinweg. Das Facial Action Coding System (FACS) wird in der Psychologie verwendet, um diese Mimiken beschreiben zu können. Dabei werden den verschiedenen möglichen Gesichtsmuskelbewegungen Bezeichner zugeordnet, sogenannte Action Units (AUs). Mit Hilfe der AUs wird die Analyse der Semantik von Gesichtsausdrücken durchgeführt. In dieser Dissertation werden mehrere Herausforderungen, insbesondere der Mangel an Trainingsdaten, die Nachvollziehbarkeit von Systementscheidungen für den Menschen in der Emotions- und Schmerzforschung und die Einbeziehung von domänespezifischen Wissen im Kontext von Systemen zur automatischen Erkennung von Gesichtsausdrücken adressiert.

Für die Entwicklung von Systemen zur automatischen Mimikerkennung werden meist Trainingsdaten benötigt. Es gibt bereits öffentliche Datenbanken zum Training von solchen Systemen, die aber alle Defizite aufweisen. Einer der wichtigsten Faktoren ist die Qualität der Annotationen, da diese maßgebend für die spätere Leistung des Systems sind. Für eine robuste Erkennung von Gesichtsausdrücken, auch bei leicht rotierten Gesichtern, müssen Personen aus verschiedenen Winkeln aufgenommen werden. Diese Robustheit ist wichtig, da in vielen Domänen nicht garantiert werden kann, dass Menschen immer frontal in die Kamera blicken. Das Bildmaterial muss hochwertig sein, damit auch subtile Änderungen im Gesicht erkennbar sind. Im Rahmen dieser Doktorarbeit wurde eine neue Datenbank, das sogenannte Actor Study Dataset, kuratiert, evaluiert und veröffentlicht. Die Datenbank beinhaltet neben qualitativ hochwertigem Bildmaterial Annotationen der dargestellten Gesichtsausdrücke, von Appraisal-Dimensionen und Emotionen bis hin zu AUs. Hierfür wurden 21 Schauspieler aus fünf Perspektiven mit synchronisierten Industrie und high-speed Kameras gefilmt und das daraus gewonnene Filmmaterial durch FACS-Experten annotiert. Zwei aktuelle Systeme zur AU-Detektion wurden zur Erstellung von Benchmark Ergebnissen für die verschiedenen Kamerawinkel genutzt. Sowohl die Aufnahmen als auch die Annotationen wurden kuratiert, aufbereitet und für die nicht-kommerzielle Nutzung zur Verfügung gestellt. 

Viele Experten im Bereich der künstlichen Intelligenz (KI), respektive des maschinellen Lernens und der Computer Vision haben Systeme entwickelt, die automatisiert Mimiken erkennen können. Es bestehen aber immer noch viele Herausforderungen: Wenige Ansätze nutzen AUs als Möglichkeit zur Nachvollziehbarkeit von Systementscheidungen und verwenden stattdessen rein datengetriebene Ansätze, deren Validierung aus Sicht von Domäneexperten oftmals schwierig ist. In dieser Arbeit werden zwei zweistufige Ansätze zur Klassifikation von Gesichtsausdrücken auf Basis von AUs vorgestellt. In die Entwicklung beider Ansätze ist Expertenwissen aus den Domänen der Emotionsforschung und der Schmerzforschung eingeflossen, um die Erkennungsleistung zu verbessern und die Komplexität zu verringern. Ein zweistufiges Verfahren ermöglicht Nachvollziehbarkeit und Validierung der Entscheidungen und Ergebnisse beider Systeme und liefert einen wichtigen Beitrag zum Forschungsfeld Erklärbarkeit von KI Entscheidungen. 

Es gibt wenige Ansätze, die Gesichtsausdrücke in sogenannte Appraisal-Dimensionen einordnen. Diese Dimensionen sind kontinuierlich und wichtig für die Klassifikation von subtilen Gesichtsausdrücken und für Rückschlüsse auf mentale Zustände, die keiner Basisemotionen zugeordnet werden können. Bestehende Ansätze zur Klassifikation in die Appraisal-Dimensionen verwenden höchstens zwei Dimensionen und bilden daher nur ein relativ kleines Spektrum ab. Zur Klassifikation von Gesichtsausdrücken in die drei Appraisal-Dimensionen Valence, Control und Novelty wird in dieser Doktorarbeit ein zweistufiger Ansatz vorgestellt. In diesem Rahmen wird als erster Schritt ein eigens entwickelter Ansatz zur automatischen Detektion von AU-Intensitäten eingeführt. Dieser berücksichtigt die zeitlichen Zusammenhänge von AUs und kann einfach um neue Informationsquellen erweitert werden, ohne ein Neutraining des Systems durchführen zu müssen und unterscheidet sich damit maßgeblich von bereits bestehenden Systemen. Durch domänespezifisches Wissen aus der Emotionsforschung konnte die Entwicklung auf die Erkennung von 22 relevante AUs konzentriert werden. Aufbauend auf den erkannten AU-Intensitäten, wird im zweiten Schritt die Einordnung in die Appraisal Dimensionen mit Hilfe einer Kleinst-Quadrate-Regression vorgenommen. Der zweistufige Ansatz ermöglicht eine Nachvollziehbarkeit der Systementscheidung für den Menschen, da die Gewichtungen der Regression direkte Rückschlüsse auf den Beitrag jeder einzelnen AU und damit der verwendeten Gesichtsmuskeln erlauben Dies ermöglicht eine zusätzliche Validierung des Systems durch Experten.

Vorhandene Ansätze zur Schmerzerkennung versuchen meist Schmerz direkt aus Bildern zu lernen oder verwenden oftmals statistische Merkmale verschiedener wichtiger Punkte im Gesicht. Diese Systeme erreichen oft eine gute Klassifikationsleistung, lassen aber meist keine Rückschlüsse auf die Entscheidungsfindung zu. Der zweite vorgestellte Ansatz befasst sich mit der Erkennung von Schmerz auf Basis eines Regelwerks. Dieses Regelwerk, eine sogenannte Grammatik, wird aus AU-Sequenzen eines Trainingsdatensatzes inferiert. Mit Hilfe der extrahierten Regeln können neue Sequenzen in "Schmerz" und "nicht Schmerz" klassifiziert werden. Wenn eine neue Sequenz durch Ableitung verschiedener Regeln generierbar ist, handelt es sich um eine Schmerzsequenz. Bei der Entwicklung des Ansatzes wurde domänespezifisches Wissen aus der Schmerzforschung verwendet, um eine Optimierung des Regelextraktionsverfahrens zu ermöglichen. Die Vorteile des gewählten Ansatzes sind die Nachvollziehbarkeit der Systementscheidung für den Menschen durch die Nachverfolgung der verwendeten Regeln der Grammatik und der Möglichkeit der Validierung durch Experten.

Den Abschluss dieser Arbeit bildet ein Aufruf zur verstärkten gemeinsamen Forschung an Ansätzen zur Nachvollziehbarkeit von KI-Systemen durch Kombination der Forschungszweige "Erklärbare KI" und "Quantifizierung von Unsicherheit in KI Entscheidungen"},
  keywords = {emotion recognition, pain detection, machine learning, explainable AI},
  school = {Otto-Friedrich-Universität Bamberg}
}

@phdthesis{hassan2020_48641,
  title = {Towards Robust and Interpretable Practical Applications of Automatic Mental State Analysis Using a Dynamic and Hybrid Facial Action Estimation Approach},
  author = {Hassan, Teena},
  year = {2020},
  publisher = {Otto-Friedrich-Universität},
  address = {Bamberg},
  doi = {10.20378/irb-48641},
  urn = {urn:nbn:de:bvb:473-irb-486414},
  url = {https://fis.uni-bamberg.de/handle/uniba/48641},
  language = {eng},
  abstract = {This dissertation presents a probabilistic state estimation framework for integrating data-driven machine learning models and a deformable facial shape model in order to estimate continuous-valued intensities of 22 different facial muscle movements, known as Action Units (AU), defined in the Facial Action Coding System (FACS). A practical approach is proposed and validated for integrating class-wise probability scores from machine learning models within a Gaussian state estimation framework. Furthermore, driven mass-spring-damper models are applied for modelling the dynamics of facial muscle movements. Both facial shape and appearance information are used for estimating AU intensities, making it a hybrid approach. Several features are designed and explored to help the probabilistic framework to deal with multiple challenges involved in automatic AU detection. The proposed AU intensity estimation method and its features are evaluated quantitatively and qualitatively using three different datasets containing either spontaneous or acted facial expressions with AU annotations. The proposed method produced temporally smoother estimates that facilitate a fine-grained analysis of facial expressions. It also performed reasonably well, even though it simultaneously estimates intensities of 22 AUs, some of which are subtle in expression or resemble each other closely. The estimated AU intensities tended to the lower range of values, and were often accompanied by a small delay in onset. This shows that the proposed method is conservative. In order to further improve performance, state-of-the-art machine learning approaches for AU detection could be integrated within the proposed probabilistic AU intensity estimation framework.

In addition to AU intensity estimation, this dissertation explores the applicability of the estimated AU intensities for automatic analysis of mental states such as pain and distraction. A survey of automatic pain detection approaches, conducted as part of this dissertation, highlights the progress and deficits in this field. Several AU-based rules are designed for pain intensity estimation based on psychological evidence, and their performance is evaluated empirically. The potential of these AU-based rules to automatically generate explanations for pain detections is also illustrated. Furthermore, a preliminary analysis of the estimated AU intensities shows differences in facial actions between various distraction scenarios during simulated driving.

The results of this dissertation show that more interdisciplinary research is needed to address the open challenges in the field of automatic mental state analysis, particularly to build reference datasets, to model interpersonal differences, and to generate human-comprehensible explanations of predictions.},
  keywords = {Facial expression analysis, facial action units, pain detection, state estimation, machine learning, interpretability, information fusion, mental state analysis, distraction detection},
  school = {Otto-Friedrich-Universität Bamberg}
}

@phdthesis{wernsdorfer2019_45233,
  title = {Symbol Grounding as the Generation of Mental Representations},
  author = {Wernsdorfer, Mark},
  year = {2019},
  publisher = {Akademische Verlagsgesellschaft AKA GmbH},
  address = {Berlin},
  series = {Dissertations in artificial intelligence},
  isbn = {978-1-61499-962-1},
  url = {https://fis.uni-bamberg.de/handle/uniba/45233},
  language = {eng},
  school = {Otto-Friedrich-Universität Bamberg}
}

@phdthesis{hofmann2010_264,
  title = {Schema-Guided Inductive Functional Programming through Automatic Detection of Type Morphisms},
  author = {Hofmann, Martin},
  year = {2010},
  publisher = {opus},
  address = {Bamberg},
  urn = {urn:nbn:de:bvb:473-opus-2981},
  url = {https://fis.uni-bamberg.de/handle/uniba/264},
  language = {eng},
  abstract = {Inductive functional programming systems can be characterised by two diametric approaches: Either they apply exhaustive program enumeration which uses input/output examples (IO) as test cases, or they perform an analytical, data-driven structural generalisation of the IO examples. Enumerative approaches ignore the structural information provided with the IO examples, but use type information to guide and restrict the search. They use higher-order functions which capture recursion schemes during their enumeration, but apply them randomly in a uninformed manner. Analytical approaches on the other side heavily exploit this structural information, but have ignored the benefits of a strong type system so far and use only recursion schemes either fixed and built in, or selected by an expert user. In category theory universal constructs, such as natural transformations or type morphisms, describe recursion schemes which can be defined on any inductively defined data type. They can be characterised by specific universal properties. Those type morphisms and related concepts provide a categorical approach to functional programming, which is often called categorical programming. This work shows how categorical programming can be applied to Inductive Programming and how universal constructs, such as catamorphisms, paramorphisms, and type functors, can be used as recursive program schemes for inductive functional programming. The use of program schemes for Inductive Programming is not new. The special appeal and novelty of this work is that, contrary to previous approaches, the program schemes are neither fixed, nor selected by an expert user: The applicability of those recursion schemes can be automatically detected in the given IO examples of a target function by checking the universal properties of the corresponding type morphisms. Applying this to the analytical system Igor2, both the capabilities and the expressiveness can be extended without a decrease in efficiency. An extension of the analytical functional inductive programming system Igor2 is proposed and its algorithms described. An empirical evaluation demonstrates the improvements with respect to efficiency and effectiveness that can be achieved by the use of type morphisms for Igor2 due to a reduction in search space complexity.},
  keywords = {Induktive Programmsynthese, Künstliche Intelligenz, Maschinelles Lernen, Funktionale Programmierung, Kategorientheorie, Typmorphismen, Catamorphismus, Inductive Program Synthesis, Artificial Intelligence, Machine Learning, Functional Programming, Category Theory, Type Morphisms, Catamorphismus, Induktive Programmsynthese, Künstliche Intelligenz, Maschinelles Lernen, Funktionale Programmierung, Kategorientheorie, Typmorphismen, Catamorphismus, Inductive Program Synthesis, Artificial Intelligence, Machine Learning, Functional Programming, Category Theory, Type Morphisms, Catamorphismus},
  school = {Otto-Friedrich-Universität Bamberg}
}

@phdthesis{kitzelmann2010_250,
  title = {A Combined Analytical and Search-Based Approach to the Inductive Synthesis of Functional Programs+H16709},
  author = {Kitzelmann, Emanuel},
  year = {2010},
  publisher = {opus},
  address = {Bamberg},
  urn = {urn:nbn:de:bvb:473-opus-2801},
  url = {https://fis.uni-bamberg.de/handle/uniba/250},
  language = {eng},
  abstract = {This thesis is concerned with the inductive synthesis of recursive declarative programs and in particular with the analytical inductive synthesis of functional programs. Program synthesis addresses the problem of (semi-)automatically generating computer programs from specifications. In inductive program synthesis, recursive programs are constructed by generalizing over incomplete specifications such as finite sets of input/output examples (I/O examples). Classical methods for induction of functional programs are analytical, that is, a recursive function definition is derived by detecting and generalizing recurrent patterns between the given I/O examples. Most recent methods, on the other side, are generate-and-test based, that is, they repeatedly generate programs independently from the provided I/O examples until a program is found that correctly computes the examples. Analytical methods are much faster than generate-and-test methods, because they do not rely on search in a program space. Therefore, however, the schemas that generatable programs conform to, must be much more restricted. This thesis at first provides a comprehensive overview of current approaches and methods to inductive program synthesis. Then we present a new algorithm to the inductive synthesis of functional programs that generalizes the analytical approach and combines it with search in a program space. Thereby, the strong restrictions of analytical methods can be resolved for the most part. At the same time, applying analytical techniques allows for pruning large parts of the problem space so that solutions can often be found faster than with generate-and-test methods. By means of several experiments with an implementation of the described algorithm, we demonstrate its capabilities.},
  keywords = {Induktive Programmsynthese, Künstliche Intelligenz, Maschinelles Lernen, Funktionale-Induktive logische Programmierung, Artificial Intelligence, Machine Learning, Functional Programming, Program Synthesis, Inductive Logic Programming},
  school = {Otto-Friedrich-Universität Bamberg}
}

