Supervision

All machine learning system have to cope with uncertainty. And therefore, have to communicate their internal decision making and their internal uncertainty. In this thesis, small machine learning systems shall be implemented and methods to visualise their uncertainty about their predictions. The development of the methods will support machine learning engineers and end-users. It also provides a method to explain the decision making, thus leading to explainable AI.

Based on time series data recorded in industry, and domain knowledge about the underlying processes, test-cases shall be generated (semi) automatic. For this, large language models (LLMs) will be trained to transform time-series data into test-cases. This thesis will be jointly supervised with the IAV GmbH.

To explore the current state, or to make predictions into the future, a digital shadow can be utilised. In this thesis, a digital shadow of a port in Rostock shall be realised, which incorporates publicly available data. It shall visualise the current state and allow for predictions into the future.

The ultimate goal of this internship was to provide a 3D visualization of these digital twins. However, rendering them with traditional rendering software can be time-consuming. For instance, generating a visualization of just 5 objects for 10 seconds takes approximately twenty minutes using Unreal Engine. Therefore, the objective of my internship was to design a lightweight application capable of visualizing the same data with significantly reduced or even negligible computation time.

This bachelor thesis addresses the challenges of rule-based Question-Answering (QA) systems by proposing and evaluating an alternative QA pipeline tailored for the e-commerce industry in the German language. The proposed pipeline includes two key phases: retrieval and reading. In the retrieval phase, we implement and evaluate three approaches, including BM25, DPR (dual-encoder), and Embedding Retriever (single-encoder), with the latter proving to be the most successful for our specific context. In the reading phase, we explore extractive and generative reader approaches and select the Llama-2 model as the reader for our pipeline. Comparing our proposed pipeline to a rule-based system currently used by novomind, we discover that the pipeline excels in scalability and context comprehension. However, rule-based systems retain an advantage in cases requiring strict control or rapid responses. Additionally, we evaluate the performance of the Llama-2-7b model against the closed-source GPT-3.5 model, with the latter demonstrating superior performance, though this performance gap may narrow with larger Llama-2 models.

Alzheimer’s Disease is a serious neurodegenerative disease and the main cause of dementia in people over 65 years. To this day, the underlying pathological mechanisms have remained partly unclear, and currently available therapies cannot ultimately stop the disease’s progression. Artificial Neural Networks have been shown to achieve good performances on the task of detecting Alzheimer’s Disease on neuroimaging data and could provide assistance during the diagnosis. However, their black-box nature compromises their application in clinical settings, since decisions are hardly traceable, and a model error is associated with a high risk. This work investigates possibilities to explain the decision of a CNNs for the detection of Alzheimer’s Disease in structural magnetic resonance images. Special attention is paid to the feature at- tribution methods, which are also used in state-of-the-art research for the interpretation of neural networks. A selection of popular feature attribution methods is discussed in detail and has been evaluated on the well-known MNIST dataset. Finally, a CNN will be trained on the ADNI dataset. Using feature attribution, the brain regions with the greatest impact on the model for the detection of Alzheimer’s Disease have been identified.

In der vorgelegten Arbeit wurden folgenden Aspekte behandelt: Zunächst der Vergleich verschiedener datengetriebener und modellbasierter Ansätze zur Objekterkennung in tabellarischer Form, basierend auf Veröffentlichungen. Daraufhin die experimentelle Untersuchung ausgewählter Systeme hinsichtlich ihrer Performanz unter verschiedenen Gesichtspunkten. Diese sind die Komplexität des Objekts, der Einfluss verschiedener Störgrößen und die Variation der verfügbaren Informationskanäle. Abschließend eine hierfür entwickelte Pipeline zur Evaluation, welche verwendet werden kann, um weiterführende Untersu- chungen vorzunehmen.

This Master's thesis presents a novel approach to action modeling and probabilistic inference by factor representations in Computational Causal Behavior Models (CCBMs) for human activity recognition. CCBM employs a symbolic model for changes of discrete states as the effects of actions. Previously, the progress of actions was modeled by durations, and inference performed by means of sampling. The research question addressed in this work is how to perform analytical inference with explicit modeling of action progress in continuous space. To investigate the feasibility of non sampling-based inference with a combination of discrete and continuous Random Variables (RVs), a comprehensive literature and tool review on hybrid Bayesian Networks (BNs) was conducted. The resulting limitation of a discrete child node depending on a continuous parent node was addressed by extending the canonical representation for Conditional Linear Gaussian (CLG) distributions. This connects the symbolic model in CCBM to truncated CLGs for the progress of actions. Using operations on representations of hybrid probabilities, a factor-based filtering algorithm for the probabilistic CCBM structure was then implemented. The main contributions of this work are the description of a CCBM with a continuous subsymbolic model for the progress of actions, the formulation of factor operations on representations of truncated CLGs, and the proof-of-work within a filter algorithm. An exemplary CCBM demonstrates how actions progress as bivariate Gaussian distributions. Discrete transitions can be modeled precisely in dependence of the continuous state, so that actions have distinct effects between regions. The thesis also highlights the limitations of existing BN tools and algorithms in hybrid probabilistic inference w.r.t. requirements of the CCBM domain. Implications of this research include the novel representation that allows factor-based inference algorithms involving truncated Gaussians and the potential for modeling the progress of actions in continuous space to improve human activity recognition. Overall, this thesis provides new insights into hybrid inference by operations on representations of probability distributions and offers a promising direction for future research in CCBMs with a subsymbolic state space.

Um mögliche Veränderungen eines Schiffsmotors frühzeitig erkennen zu können, ist eine kontinuierliche Überwachung notwendig. Insbesondere akustische Sensoren sind hierfür geeignet, da Klangveränderungen häufig den Ausfall einer Maschinenkomponente vorausgehen. Das Ziel der vorliegenden Arbeit ist es, verschiedene Methoden zur Anomaliedetektion in Tondaten auf ihre Eignung für Schiffsmotoren zu evaluieren. Aufgrund verschiedener externer Faktoren ist es schwierig im laufenden Schiffsbetrieb eine ausreichend große Menge an hoch-qualitativen Daten zu generieren. Aus diesem Grund soll untersucht werden, ob durch Generierungsverfahren neue Daten für das Training von Verfahren zur Anomaliedetektion und die Evaluation von Verfahren zur Anomaliedetektion erzeugt werden können. Verschiedene Experimente zeigten, dass insbesondere Autoencoder eine geeignete Methode zur Anomaliedetektion in Tondaten darstellen. Das untersuchte Autoencoder-Netzwerk ist insgesamt besser in der Lage gewesen, Anomalien zu identifizieren als es das Neuronale Netz VGG19 oder ein One-Class Support Vector Machines waren. Zur Generierung synthetischer Daten wurde ein SpecGAN verwendet. Damit ist die durchgeführte Arbeit eine Grundlage zur weiteren Forschung im Bereich der Anomaliedetektion in Tondaten von Schiffsmotoren und zeigt Bedürfnisse in der Generierung synthetischer Daten auf.

Diese Arbeit präsentiert die Konzeption und Realisierung eines webbasierten Systems für die Unterstützung der Schlaganfalltherapie. Dieses System zielt darauf ab, die Effizienz der Therapie zu steigern und die Zugänglichkeit für Patienten zu verbessern. Mit diesem System können die Therapeuten parallel mehrere Pateinten in einer Videokonferenz betreuen, wobei das System selbst die Patienten während der Videokonferenz verfolgt und ihre Leistungen bewertet. Damit kann man eine effiziente Alternative für die Rehabilitationssitzung schaffen, die Rehabilitationsprozess unterstützten kann.

The application possibilities of methods for 3D object detection, by means of which the position, size, and orientation of objects in three-dimensional space can be determined, are manifold. Among them are in the aerial domain, the application in self-driving vehicles [ZEZ+22] or in robots for grasping objects [CK17]. Depending on the environment, different sensors are used for this purpose, including LiDAR, Radar, and RGB-D cameras. In addition to the above-mentioned applications, there are others in the underwater domain, where the use of Sonars is irreplaceable due to the properties of water and the associated difficulty of using LiDAR, Radar, and cameras. These include, for example, according to [HRG19], maintenance work on offshore installations, the exploration of inaccessible or dangerous maritime objects, and the identification of marine wildlife. However, research in the field of neural network-based 3D object detection is far less advanced compared to its 2D counterpart. This is even more true for the underwater 3D object detection domain. Accordingly, this thesis deals with the identification and evaluation of a neural network-based approach for the detection of underwater objects in point clouds. In addition, the existing literature is surveyed for annotated underwater point cloud datasets and a tool for the randomized generation of underwater point cloud scenes is developed. The requirements for the model as well as the dataset are noted in a requirements analysis. Furthermore, the identified model is adapted for use on an underwater point cloud dataset and the combination of model and dataset is evaluated using hypothesis-driven experiments. For the execution of these, a framework is developed that allows automated execution of the experiments.

In dieser Arbeit soll untersucht werden, wie moderne Ansätze zur Gebäuderegelung aus dem Bereich der künstlichen Intelligenz für die Regelung eines trägen Systems wie der Betonkernaktivierung angewendet werden können. Dabei stehen vor allem der reale Nutzen und die Anwendbarkeit der Regelsysteme im Vordergrund. Mit Hilfe von Zeitreihenanalysen wird zunächst ein real existierendes und funktionierendes Heizungssystem bzw. Gebäude nachgebildet. Anschließend soll wiederum auf Basis von Zeitreihen eine Optimierung der Regelung für das modellierte Gebäude gefunden werden. Dabei werden die erstellten Modelle des Gebäudes verwendet. Letztlich ist es das Ziel, die Regelung des modellierten Gebäudes auf das reale Gebäude zu übertragen und gegebenenfalls anzupassen.

In dieser Arbeit wird untersucht, ob Aktuatorstrukturen, welche sowohl ein- als auch mehrgelenkigen Muskeln modellieren, im Vergleich zu Aktuatorstrukturen, welche ausschließlich eingelenkige Muskeln modellieren, einen signifikanten Vorteil bei dem Lernprozess des Balancierens eines humanoiden Reinforcement Learning Agenten bieten. Um diese Frage zu beantworten, werden zwei verschiedene Agentenmodelle konzeptioniert und implementiert. Diese Modelle werden in zwei verschiedenen Umgebungen, mit derselben Aufgabe des Balancierens, trainiert. Die bei den Versuchsausführungen gesammelten Daten werden anschließend ausgewertet. Die Auswertung erfolgt anhand von sechs im Vorhinein aufgestellten Hypothesen, welche teils die Systemvalidität überprüfen und teils verschiedene Teilaspekte der Forschungsfrage behandeln. Am Ende der Arbeit werden die verwendeten Methodiken und technischen Mittel sowie in der Konzeption getroffene Entscheidungen kritisch bewertet. Aus dieser Bewertung werden Verbesserungsmöglichkeiten für die in dieser Arbeit entwickelten Modelle und Hinweise für die Entwicklung neuer Modelle in zukünftige Forschungsarbeiten, welche diese Grundthematik behandeln, erarbeitet.

Magneto-mechanical resonators are a new class of sensors that can be used for location tracking or to measure physical quantities like temperature and pressure. They are excited using a transmit coil and induce a signal in a receive coil or a coil array. The location of the resonator is coupled to the signal amplitude while the resonance frequency of the oscillation is influenced by the physical quantity that is to be measured. If the sensor is operated in an unshielded environment, a variety of background signals such as harmonics of the grid frequency and signals of switched-mode power supplies or monitors disturb the sensor signal and decrease the signal-to-noise ratio. Since these signals are not random, prior knowledge about them can be used for modeling and removal. In cooperation with Philips Innovative Technologies Research Laboratories, methods for the removal of noise signals in an unshielded multi-coil receive system are to be developed. As a first approach, different background signals are to be identified and statistically modeled. These models are then used for the removal of the background signals. As a second step, artificial neural networks (ANN) shall be used to facilitate the process and remove more complex signals which cannot easily be modeled by hand. Both approaches – the statistical modeling and the ANN approach – are evaluated using measurement data with and without a sensor signal.

Diese Bachelorarbeit verfolgte die folgenden Ziele:

1. Es sollen Ansätze identifiziert werden, mit denen die Wahrscheinlichkeitsverteilungen eines Datensatzes gelernt und die Wahrscheinlichkeitsdichtefunktion approximieren werden kann.
2. Es soll eine Auswahl aus den gefundenen Ansätzen implementiert werden.
3. Die implementierten Ansätze sollen experimentell evaluiert werden.

Auf Basis einer systematischen Literaturanalyse wurden drei Ansätze identifiziert (VAE, Student-t VAE und CycleGan) und hinsichtlich der Darstellung einfacher 1-dimensionaler Wahrscheinlichkeitsverteilungen getestet und evaluiert.

Deep convolutional neural networks (CNNs) are considered to be the quasi standard for semantic segmentation in computer vision. To generalize geospatial insights on remote sensing (RS) imagery, semantic segmentation is commonly used to detect a wide variety of objects. However, the quality of predictions generated by CNNs relies heavily on two input factors: the parameters of the models architecture and the input data that consists of images together with a set of labels containing the true classification or ground truth (GT). In order to adequately train a CNN, GT labels need to be procured beforehand which is often associated with intensive costs and time as well as bottlenecks in availability. In order to reduce the costs of generating geospatial insights as well as to overcome the bottlenecks in the availability of data, the most promising alternative to increase data volume might be to shift the focus towards increasing the quality of the class labels. In order to examine the impact of the quality of geospatial labels on the quality of predictions generated by CNNs in the case of building detection on RS images, a U-ResNet-50 network is trained with various label datasets of different quality levels. It is shown that missing and noisy labels have a significant impact on the quality of the predictions generated by the network. Moreover, it is shown for a specific example that it is possible to achieve just as precise predictions with one fortieth of the GT dataset as with a 40 times larger dataset lacking 20 percent of the labels.

In the modern workplace, rising numbers of stress related health problems of employees are responsible for the growing importance of a good occupational healthcare system. Here, the concept of human energy can be of help. It summarises sensations like feeling vital, awake and concentrated. The human energy can be individually rated on a designated battery scale, which is already approved and well-tested. However, the options for obtaining information about the human energy of a person are limited to highly sub- jective methods such as self-assessment questionnaires. Consequently, there is a need for objective methods to determine the level of human energy. The heart rate variability (HRV) is a physiological parameter that measures the variation of time in between heart beats. It is already known to be an indicator of cardiovascular and psychological health in a person and could serve as the above-mentioned objective method to detect human energy. An experiment with 8 participants was conducted, which combined the recording of the heart’s activity with an Empatica E4 wristband and the human energy through a self assessment. HRV was computed from recorded heart activity. In an analysis of correlations, three hypotheses were examined: 1) Self reported scales of vitality and fatigue correlate with the self reported human energy; 2) HRV parameters reflecting parasympathetic activity (RMSSD and HF) are believed to rise with higher human energy; 3) The frequency-based HRV parameter LF will lower with increased levels of human energy. Only the first hypothesis could be verified on the basis of the present data. Out of five evaluated HRV parameters, only the LF/HF ratio expressed a significant positive correlation to human energy. The findings suggest that individuals feel more energised when they are in an alert state better known as the ‘fight-or-flight’ mode. However, his interpretation is to be treated with caution as the validity of LF/HF ratio is disputed.

In this thesis topic, we explored synthetic data generation using conditional variational autoencoder, a new data augmentation technique which is useful when you have small amount of data. To test whether whether adequate data samples can be generated, we used three publicly available image based datasets called MNIST, Fashion MNIST and Cifar10. We generated data samples from different sizes of training data and generated different amount of synthetic data. We used two different cVAE architectures, one with convolutional layers and one without. The results were generated with original dataset being set as benchmark before performing the same experiments with augmented synthetic samples. All experiments were performed multiples times to get an average score, this was done due to stochastic nature of neural networks. We used three different classifiers to calculate the accuracy performances of the experiments. We also briefly explored a different data augmentation technique and observed its outcome, which was done using Keras preprocessing library.

Human activity recognition is a wide-ranging field in artificial intelligence and is used in a wide variety of domains. Often, activity recognition is modeled purely data-driven from annotated sensor data. However, additional expert models are given for certain scenarios, thus incorporating these models into existing methods can yield better recognition. Also, discovering additional knowledge in the data can improve the classification of human activity. The types of such additional information can be broad. For example, they may describe relations between objects and humans, or they may describe the current tasks that need to be accomplished. The task of this thesis is to improve the recognition of challenging and agitated behaviors of patients with dementia. Through observations of the annotated behavior of dementia patients, it is noticeable that people behave differently at different segments of the day. In this work, we want to find behavioral predispositions that define the tendency to behave in certain ways. With the aim that the recognition of the dementia patients’ behaviors can be improved with complete knowledge of the behavioral predisposition, this work proposes a model that exploits the different prior distributions of the behaviors at the given behavioral predisposition. The evaluation showed a doubling of the F1 measure from 0.23 to 0.47 using this model with optimal parameters.

In dieser Arbeit wird ein System zur schnellen und einfachen Ausrichtung von Schwenke-Neige-Zoom Kameras vorgestellt. Für die Realisierung des Systems wurden zunächst Handgesten in Betracht gezogen. Nachdem sich für das Handgesten-basierte System verschiedene Probleme ergaben, wurde ein Markerbasiertes System entwickelt. Für dieses Markerbasierte System wird auf die Problematik des Markerdesigns eingegangen und anschließend ein Verfahren zur Identifikation von Markern ausführlich beschrieben. Zum Schluss wir die Bedienung des entwickelten Systems erklärt.

Ziel der in dieser Arbeit entwickelten Kommunikations-Infrastruktur ist eine hohe Kompatibilität, eine einfache Handhabbarkeit und die Realisierung einer globalen, persistenten Plattform für Informationen jeglicher Art.

Statistical methods are often used to built inference on sensor based data. A Hidden Markov Models is such a statistical method. They are used in many scientific branches to determine specific processes. One application area is activity recognition, where activities are inferred from sensor data. Often the creation of these models is very complex. Hence this work shows a concept how to simplify the process utilising stochastical grammars. Due to their logical structure it is more intuitive to specify the distributions.

In dieser Arbeit geht es um eine Verlaufsfunktion für Whiteboards. Zunächst wird das Problem beschrieben und durch Beispiele das Lösen dieser Problematik motiviert. Im folgenden Kapitel wird gezeigt, wie mit Hilfe eines Stitchingverfahrens hochauflösende und rektifizierte Bilder eines Whiteboards aufgenommen werden können. Das dritte Kapitel beschätigt sich mit der Umsetzung des Verfahrens, der Software „Whiteboard History Viewer“. Zum Abschluss folgt eine kleine Zusammenfassung und es werden Möglichkeiten zur Weiterentwickung aufgezeigt.

In the recent years, providing on-demand Quality of Services (QoS) in mobile networks is a key issue and also a challenge for mobile companies. The prediction of QoS parameters is a difficult job, due to the fact that the parameters vary with time, and mutually influential. The unsatisfactory solution for network performance prediction leads to the limit of developing high-end software on cellphones. This paper is focused on comparison of Artificial Neural Networks (ANN) to statistical method for time-series prediction. For statistical technique, we choose Auto-Regressive Integrated Moving Average (ARIMA) for model fitting and Kalman Filter for prediction. For neural networks, we use three layered ANN with Back-Propagation (BP). The intercomparison was done on three data sets, namely, Delay, Jitter, and Throughput. The result of analysis shows that ARIMA + Kalman Filter slightly outperforms ANN.

Combining artificial neural networks and logic programming for machine learning tasks is the main objective of neural symbolic integration. One important step towards practical applications in this field is the development of techniques for extracting symbolic knowledge from neural networks.
In this thesis a new extraction method is proposed and thoroughly investigated. It translates the class of feedforward networks with binary threshold functions into propositional logic programs by means of a decompositional approach.

The field of neural-symbolic integration has received much attention recently. While with propositional paradigms, the integration of symbolic knowledge and connectionist systems (also called artificial neural networks) has already resulted in applicable systems, the theoretical foundations for the first-order case are currently being laid and first perspectives for real implementations are emerging. Two important components of the neural-symbolic learning cycle [BH05] are representation, i.e. encoding symbolic knowledge into connectionist systems, and training, i.e. adjusting these connectionist systems according to information observed in other ways. These components are the focus of this thesis. Extending results from [Wit05, BHW05], a practically feasible way is presented to approximate and embed the semantic operator of covered logic programs in a real-valued domain, and connectionist architectures suitable for representing this particular form of symbolic knowledge are developed and evaluated along with appropriate training methods.

This document is essentially divided in two parts, where different methods are presented for extracting knowledge from an aritificial neural network representing an immediate consequence operator.
In the first part we investigate the relationship between neurosymbolic integration (in particular the extraction of a logic program from a neural network) and inductive logic programming from a practical point of view. After a general introduction to the foundations of ILP, the task of extraction of a neural network is reformulated to fit the problem setting of ILP. We then practically test a variety of different programs and evaluate them.
The second part of the document builds up a theoretical foundation for the special case of extracting propositional logic programs. We give algorithms for definite as well as normal propositional logic programs. Several theoretical results are presented, difficulties and possible solutions are observed.

Significant advances have recently been made concerning the integration of symbolic knowledge representation with connectionist systems (also called artificial neural networks). However, while the integration with propositional paradigms has resulted in applicable systems, the case of first-order knowledge representation has so far hardly proceeded beyond theoretical studies which prove the existence of connectionist systems for approximating first-order logic programs up to any chosen precision. Advances were hindered severely by the lack of concrete algorithms for obtaining the approximating networks which were known to exist: the corresponding proofs are not constructive in that they do not yield concrete methods for building the systems. In this paper, we will make the required advance and show how to obtain the structure and the parameters for different kinds of connectionist systems approximating covered logic programs.

Diese Belegarbeit soll einen Einblick in die Welt der iterierten Funktionssysteme geben und die Grundlagen zum Verständis der Mathematik dahinter veranschaulichen. Neben den theoretischen Aspekten iterierter Funktionssystem wird auch die Implementation angesprochen. Dies wird anhand von Beispielen und der Programmiersprache Java erklärt. Weiterhin wird in diesem Beleg eine Software vorgestellt, die fähig ist iterierte Funktionssysteme darzustellen.

Die vorliegende Arbeit entstand im Rahmen der Besonderen Lernleistung / eines wissenschaftlichen Projekts der Klasse 11 am Martin - Andersen - Nexö Gymnasium Dresden. Das Projekt wurde am Institut für Künstliche Intelligenz der Fakultät Informatik der Technischen Universität Dresden unter der Betreuung von Herrn Sebastian Bader bearbeitet. Ich habe mich mit einem Teilgebiet der Neuronalen Netze befasst, den RBF-Netzen (RBF = Radial Basis Functions – zu deutsch radial-symmetrische Funktionen). Die Neuronalen Netze sind wiederum ein Teilgebiet der Künstlichen Intelligenz und somit der Informatik. Mit der Programmiersprache JAVA habe ich ein Programm entworfen, das dazu dient Funktionen zu interpolieren. Der Benutzer gibt dabei mehrere Punkte in einem Koordinatensystem an und dann kann er mittels des Neuronalen Netzwerks eine Funktion errechnen lassen, die diese Punkte möglichst optimal annähert. Um dies zu realisieren, muss das Netzwerk gelernt werden. Das Lernen des Netzwerks ist dessen wichtigste Eigenschaft, da es ohne einen geeigneten Lernalgorithmus unbenutzbar bleibt. Um die vom Nutzer vorgegebenen Werte nun zu approximieren muss er die verschiedenen Variablen zum Lernen und das Netzwerk selbst einstellen.