05 Fakultät Informatik, Elektrotechnik und Informationstechnik
Permanent URI for this collectionhttps://elib.uni-stuttgart.de/handle/11682/6
Browse
Search Results
Item Open Access Challenges of computational social science analysis with NLP methods(2022) Dayanik, Erenay; Padó, Sebastian (Prof. Dr.)Computational Social Science (CSS) is an emerging research area at the intersection of social science and computer science, where problems of societal relevance can be addressed by novel computational methods. With the recent advances in machine learning and natural language processing as well as the availability of textual data, CSS has opened up to new possibilities, but also methodological challenges. In this thesis, we present a line of work on developing methods and addressing challenges in terms of data annotation and modeling for computational political science and social media analysis, two highly popular and active research areas within CSS. In the first part of the thesis, we focus on a use case from computational political science, namely Discourse Network Analysis (DNA), a framework that aims at analyzing the structures behind complex societal discussions. We investigate how this style of analysis, which is traditionally performed manually, can be automated. We start by providing a requirement analysis outlining a roadmap to decompose the complex DNA task into several conceptually simpler sub-tasks. Then, we introduce NLP models with various configurations to automate two of the sub-tasks given by the requirement analysis, namely claim detection and classification, based on different neural network architectures ranging from unidirectional LSTMs to Transformer based architectures. In the second part of the thesis, we shift our focus to fairness, a central concern in CSS. Our goal in this part of the thesis is to analyze and improve the performances of NLP models used in CSS in terms of fairness and robustness while maintaining their overall performance. With that in mind, we first analyze the above-mentioned claim detection and classification models and propose techniques to improve model fairness and overall performance. After that, we broaden our focus to social media analysis, another highly active subdomain of CSS. Here, we study text classification of the correlated attributes, which pose an important but often overlooked challenge to model fairness. Our last contribution is to discuss the limitations of the current statistical methods applied for bias identification; to propose a multivariate regression based approach; and to show that, through experiments conducted on social media data, it can be used as a complementary method for bias identification and analysis tasks. Overall, our work takes a step towards increasing the understanding of challenges of computational social science. We hope that both political scientists and NLP scholars can make use of the insights from this thesis in their research.Item Open Access Stochastic neural networks : components, analysis, limitations(2022) Neugebauer, Florian; Polian, Ilia (Prof. Dr.)Stochastic computing (SC) promises an area and power-efficient alternative to conventional binary implementations of many important arithmetic functions. SC achieves this by employing a stream-based number format called Stochastic numbers (SNs), which enables bit-sequential computations, in contrast to conventional binary computations that are performed on entire words at once. An SN encodes a value probabilistically with equal weight for every bit in the stream. This encoding results in approximate computations, causing a trade-off between power consumption, area and computation accuracy. The prime example for efficient computation in SC is multiplication, which can be performed with only a single gate. SC is therefore an attractive alternative to conventional binary implementations in applications that contain a large number of basic arithmetic operations and are able to tolerate the approximate nature of SC. The most widely considered class of applications in this regard is neural networks (NNs), with convolutional neural networks (CNNs) as the prime target for SC. In recent years, steady advances have been made in the implementation of SC-based CNNs (SCNNs). At the same time however, a number of challenges have been identified as well: SCNNs need to handle large amounts of data, which has to be converted from conventional binary format into SNs. This conversion is hardware intensive and takes up a significant portion of a stochastic circuit's area, especially if the SNs have to be generated independently of each other. Furthermore, some commonly used functions in CNNs, such as max-pooling, have no exact corresponding SC implementation, which reduces the accuracy of SCNNs. The first part of this work proposes solutions to these challenges by introducing new stochastic components: A new stochastic number generator (SNG) that is able to generate a large number of SNs at the same time and a stochastic maximum circuit that enables an accurate implementation of max-pooling operations in SCNNs. In addition, the first part of this work presents a detailed investigation of the behaviour of an SCNN and its components under timing errors. The error tolerance of SC is often quoted as one of its advantages, stemming from the fact that any single bit of an SN contributes only very little to its value. In contrast, bits in conventional binary formats have different weights and can contribute as much as 50\% of a number's value. SC is therefore a candidate for extreme low-power systems, as it could potentially tolerate timing errors that appear in such environments. While the error tolerance of SC image processing systems has been demonstrated before, a detailed investigation into SCNNs in this regard has been missing so far. It will be shown that SC is not error tolerant in general, but rather that SC components behave differently even if they implement the same function, and that error tolerance of an SC system further depends on the error model. In the second part of this work, a theoretical analysis into the accuracy and limitations of SC systems is presented. An existing framework to analyse and manage the accuracy of combinational stochastic circuits is extended to cover sequential circuits. This framework enables a designer to predict the effect of small design changes on the accuracy of a circuit and determine important parameters such as SN length without extensive simulations. It will further be shown that the functions that are possible to implement in SC are limited. Due to the probabilistic nature of SC, some arithmetic functions suffer from a small bias when implemented as a stochastic circuit, including the max-pooling function in SCNNs.Item Open Access Improved usability of differential privacy in machine learning : techniques for quantifying the privacy-accuracy trade-off(2022) Bernau, Daniel; Küsters, Ralf (Prof.)Differential privacy allows bounding the influence that training data records have on a neural network. To use differential privacy in machine learning with neural networks, data scientists must choose privacy parameter epsilon. Choosing meaningful privacy parameters is key since differentially private neural networks that have been trained with weak privacy parameters might result in excessive privacy leakage, while strong privacy parameters might overly degrade model utility. However, privacy parameter values are difficult to choose for two main reasons. First, the theoretical upper bound on privacy loss epsilon might be loose, depending on the chosen sensitivity and data distribution of practical datasets. Second, legal requirements and societal norms for anonymization often refer to individual identifiability, to which epsilon is only indirectly related. Within this thesis, we address the problem of choosing epsilon from two angles. First, we quantify the empirical lower bound on the privacy loss under empirical membership inference attacks to allow data scientists to compare the empirical privacy-accuracy trade-off between local and central differential privacy. Specifically, we consider federated and non-federated discriminative models, as well as generative models. Second, we transform the privacy loss under differential privacy into an analytical bound on identifiability map legal and societal expectations w.r.t. identifiability to corresponding privacy parameters. The thesis contributes techniques for quantifying the trade-off between accuracy and privacy over epsilon. The techniques provide information for interpreting differentially private training datasets or models trained with the differentially private stochastic gradient descent to improve usability of differential privacy in machine learning. In particular, we identify preferable ranges for privacy parameter epsilon and compare local and central differential privacy mechanisms for training differentially private neural networks under membership inference adversaries. Furthermore, we contribute an implementable instance of the differential privacy adversary that can be used to audit trained models w.r.t. identifiability.Item Open Access Prediction and similarity models for visual analysis of spatiotemporal data(2022) Tkachev, Gleb; Ertl, Thomas (Prof. Dr.)Ever since the early days of computers, their usage have become essential in natural sciences. Whether through simulation, computer-aided observation or data processing, the progress in computer technology have been mirrored by the constant growth in the size of scientific data. Unfortunately, as the data sizes grow, and human capabilities remains constant, it becomes increasingly difficult to analyze and understand the data. Over the last decades, visualization experts have proposed many approaches to address the challenge, but even these methods have their limitations. Luckily, recent advances in the field of Machine Learning can provide the tools needed to overcome the obstacle. Machine learning models are a particularly good fit as they can both benefit from the large amount of data present in the scientific context and allow the visualization system to adapt to the problem at hand. This thesis presents research into how machine learning techniques can be adapted and extended to enable visualization of scientific data. It introduces a diverse set of techniques for analysis of spatiotemporal data, including detection of irregular behavior, self-supervised similarity metrics, automatic selection of visual representations and more. It also discusses the general challenges of applying Machine Learning to Scientific Visualization and how to address them.Item Open Access Improving automotive radar spectra object classification using deep learning and multi-class uncertainty calibration(2022) Patel, Kanil; Yang, Bin (Prof. Dr.-Ing.)Being a prerequisite for successful automated driving, ameliorating the perception capabilities of vehicles is of paramount importance for reliable and robust scene understanding. Required for decision-making in autonomous vehicles, scene understanding becomes particularly challenging in adverse weather and lighting conditions; situations also often posing challenges for human drivers. Automotive radars can greatly assist sensors currently deployed on vehicles for robust measurements, especially in challenging conditions where other sensors often fail to operate reliably. However, classification using radar sensors is often limited to a few classes (e.g. cars, humans, and stop signs), controlled laboratory settings, and/or simulations. Already offering reliable distance, azimuth and velocity estimates of the objects in the scene, improving radar-based classification greatly expands the usage of radar sensors for tackling multiple driving-related tasks which are often performed by other less robust sensors. This thesis investigates how automated driving perception can be improved using multi-class radar classification by using deep learning algorithms for exploiting object class characteristics captured in the radar spectra. Despite the highly-accurate predictions of deep learning models, such classifiers exhibit severe over-confidence which can lead decision-making systems to false conclusions, with possibly catastrophic consequences - often a matter of life and death for automated driving. Consequently, high-quality, robust, and interpretable uncertainty estimates are indispensable characteristics of any unassailable automated driving system. With the goal of uncertainty estimates for real-time predictive systems, this thesis also aims at tackling the prominent over-confidence of deep learning classification models, which persists for all data modalities. Being an important measure for the quality of uncertainty estimates, this work focuses on the accurate estimation of the calibration of trained classifiers, as well as present novel techniques for improving their calibration. The presented solutions offer high-quality real-time confidence estimates for classification models of all data modalities (e.g. non-radar applications), as well as classifiers which are already trained and used in practise and new training strategies for learning new classifiers. Furthermore, the presented uncertainty calibration algorithms could also be extended to tasks other than classification, for example, regression and segmentation. On a challenging new realistic automated driving radar dataset, the solutions proposed in this thesis show that radar classifiers are able to generalize to novel driving environments, driving patterns, and object instances in realistic static driving scenes. To further replicate realistic encounters of autonomous vehicles, we study the behaviour of the classifiers to spectra corruptions and outlier detection of unknown objects, showing significant performance improvements in safely handling these prevalent encounters through accurate uncertainty estimates. With the proposed generalization and requisite accurate uncertainty estimation techniques, the radar classifiers in this study greatly improve radar-based perception for scene understanding and lay a solid foundation for current sensor fusion techniques to leverage radar measurements for object classification.Item Open Access Scatter and beam hardening correction for high-resolution CT in near real-time based on a fast Monte Carlo photon transport model(2022) Alsaffar, Ammar; Simon, Sven (Prof. Dr.-Ing.)Computed tomography (CT) is a powerful non-destructive testing (NDT) technique. It provides inception about the inner of the scanned object and is widely used for industrial and medical applications. However, this technique suffers from severe quality degradation artifacts. Among these artifacts, the scatter and the beam hardening (BH) causes severe quality degradation of the reconstructed CT images. The scatter results from the change in the direction, or the direction and the energy of the photon penetrating the object, while the beam hardening results from the polychromatic nature of the X-ray source. When photons of different energies penetrate through the object, low-energy photons are more easily absorbed than high-energy photons. This results in the hardening of the X-ray beam which causes the non-linear relation between the propagation path length and the attenuation of the beam. These kinds of artifacts are the major source of the cupping and the streak artifacts that highly degrades the quality of the computed tomography imaging. The presence of the cupping and the streak artifacts reduce the contrast of this image and the contrast-to-noise and cause distortion of the grey values. As a consequence important analysis of the results from the computed tomography technique is affected, e.g., the detectability of voids and cracks is reduced by the reduction of the contrast and affects the dimensional measurement. Monte Carlo (MC) simulation is considered the most accurate approach for scatter estimation. However, the existing MC estimators are computationally expensive, especially for the considered high-resolution flat-panel CT. In this work, a muli-GPU photon forward projection model and an iterative scatter correction algorithm were implemented. The Monte Carlo model has been highly accelerated and extensively verified using several experimental and simulated examples. The implemented model describes the physics within the 1 keV to 1 MeV range using multiple controllable key parameters. Based on this model, scatter computation for a single projection can be completed within a range of a few seconds under well-defined model parameters. Smoothing and interpolation are performed on the estimated scatter to accelerate the scatter calculation without compromising accuracy too much compared to measured near scatter-free projection images. Combining the scatter estimation with the filtered backprojection (FBP), scatter correction is performed effectively in an iterative manner. In order to evaluate the proposed MC model, extensive experiments have been conducted on the simulated data and real-world high-resolution flat-panel CT. Compared to the state-of-the-art MC simulators, the proposed MC model achieved a 15× acceleration on a single GPU in comparison to the GPU implementation of the Penelope simulator (MCGPU) utilizing several acceleration techniques, and a 202× speed-up on a multi-GPU system comparing to the multi-threaded state-of-the-art EGSnrc MC simulator. Furthermore, it is shown that for high-resolution images, scatter correction with sufficient accuracy is accomplished within one to three iterations using a FBP and the proposed fast MC photon transport model. Moreover, a fast and accurate BH correction method that requires no prior knowledge of the materials and corrects first and higher-order BH artifacts has been implemented. In the first step, a wide sweep of the material is performed based on an experimentally measured look-up table to obtain the closest estimate of the material. Then the non-linearity effect of the BH is corrected by adding the difference between the estimated monochromatic and the polychromatic simulated projections of the segmented image. The estimated monochromatic projection is simulated by selecting the energy from the polychromatic spectrum which produces the lowest mean square error (MSE) with the BH-corrupted projection from the scanner. While the polychromatic projection is accurately estimated using the least square estimation (LSE) method by minimizing the difference between the experimental projection and the linear combination of simulated polychromatic projections using different spectra of different filtration. As a result, an accurate non-linearity correction term is derived that leads to an accurate BH correction result. To evaluate the proposed BH correction method, extensive experiments have been conducted on real-world CT data. Compared to the state-of-the-art empirical BH correction method, the experiments show that the proposed method can highly reduce the BH artifacts without prior knowledge of the materials. In summary, the lack of the availability of fast and computationally efficient methods to correct the major artifacts in CT images, i.e., scatter and beam hardening, has motivated this work in which efficient and fast algorithms have been implemented to correct these artifacts. The correction of these artifacts has led to better visualization of the CT images, a higher contrast-to-noise ratio, and improved contrast. Supported by multiple experimental examples, it is shown that the scatter corrected images, using the proposed method, resample the near artifacts-free reference images acquired experimentally within a reasonable time. On the other hand, the application of the proposed BH correction method after the correction of the scatter artifacts results in the complete removal of the rest cupping and streak artifacts that were degrading the scatter-corrected images and improved the contrast-to-noise (CNR) ratio of the scatter-corrected images. Moreover, assessments of the correction quality of the CT images have been performed using the software Volume Graphics VGSTUDIO MAX. Better surface determination can be derived from the artifacts-corrected images. In addition, enhancing the contrast by correcting these artifacts results in an improved detectability of voids and cracks in several concrete examples. This supports the efficiency of the implemented artifacts correction methods in this work.Item Open Access Neural network full-body human predictive models and their use for coordinated robot motion planning(2022) Kratzer, Philipp; Toussaint, Marc (Prof. Dr.)Item Open Access Deep learning based prediction and visual analytics for temporal environmental data(2022) Harbola, Shubhi; Coors, Volker (Prof. Dr.)The objective of this thesis is to focus on developing Machine Learning methods and their visualisation for environmental data. The presented approaches primarily focus on devising an accurate Machine Learning framework that supports the user in understanding and comparing the model accuracy in relation to essential aspects of the respective parameter selection, trends, time frame, and correlating together with considered meteorological and pollution parameters. Later, this thesis develops approaches for the interactive visualisation of environmental data that are wrapped over the time series prediction as an application. Moreover, these approaches provide an interactive application that supports: 1. a Visual Analytics platform to interact with the sensors data and enhance the representation of the environmental data visually by identifying patterns that mostly go unnoticed in large temporal datasets, 2. a seasonality deduction platform presenting analyses of the results that clearly demonstrate the relationship between these parameters in a combined temporal activities frame, and 3. air quality analyses that successfully discovers spatio-temporal relationships among complex air quality data interactively in different time frames by harnessing the user’s knowledge of factors influencing the past, present, and future behaviour with Machine Learning models' aid. Some of the above pieces of work contribute to the field of Explainable Artificial Intelligence which is an area concerned with the development of methods that help understand, explain and interpret Machine Learning algorithms. In summary, this thesis describes Machine Learning prediction algorithms together with several visualisation approaches for visually analysing the temporal relationships among complex environmental data in different time frames interactively in a robust web platform. The developed interactive visualisation system for environmental data assimilates visual prediction, sensors’ spatial locations, measurements of the parameters, detailed patterns analyses, and change in conditions over time. This provides a new combined approach to the existing visual analytics research. The algorithms developed in this thesis can be used to infer spatio-temporal environmental data, enabling the interactive exploration processes, thus helping manage the cities smartly.Item Open Access Evaluation and control of the value provision of complex IoT service systems(2022) Niedermaier, Sina; Wagner, Stefan (Prof. Dr.)The Internet of Things (IoT) represents an opportunity for companies to create additional consumer value through merging connected products with software-based services. The quality of the IoT service can determine whether an IoT service is consumed in the long-term and whether it delivers the expected value for a consumer. Since IoT services are usually provided by distributed systems and their operations are becoming increasingly complex and dynamic, continuous monitoring and control of the value provision is necessary. The individual components of IoT service systems are usually developed and operated by specialized teams in a division of labor. With the increasing specialization of the teams, practitioners struggle to derive quality requirements based on consumer needs. Consequently, the teams often observe the behavior of “their” components isolated without relation to value provision to a consumer. Inadequate monitoring and control of the value provision across the different components of an IoT system can result in quality deficiencies and a loss of value for the consumer. The goal of this dissertation is to support organizations with concepts and methods in the development and operations of IoT service systems to ensure the quality of the value provision to a consumer. By applying empirical methods, we first analyzed the challenges and applied practices in the industry as well as the state of the art. Based on the results, we refined existing concepts and approaches. To evaluate their quality in use, we conducted action research projects in collaboration with industry partners. Based on an interview study with industry experts, we have analyzed the current challenges, requirements, and applied solutions for the operations and monitoring of distributed systems in more detail. The findings of this study form the basis for further contributions of this thesis. To support and improve communication between the specialized teams in handling quality deficiencies, we have developed a classification for system anomalies. We have applied and evaluated this classification in an action research project in industry. It allows organizations to differentiate and adapt their actions according to different classes of anomalies. Thus, quick and effective actions to ensure the value provision or minimize the loss of value can be optimized separately from actions in the context of long-term and sustainable correction of the IoT system. Moreover, the classification for system anomalies enables the organization to create feedback loops for quality improvement of the system, the IoT service, and the organization. To evaluate the delivered value of an IoT service, we decompose it into discrete workflows, so-called IoT transactions. Applying distributed tracing, the dynamic behavior of an IoT transaction can be reconstructed in a further activity and can be made “observable”. Consequently, the successful completion of a transaction and its quality can be determined by applying indicators. We have developed an approach for the systematic derivation of quality indicators. By comparing actual values determined in operations with previously defined target values, the organization is able to detect anomalies in the temporal behavior of the value provision. As a result, the value provision can be controlled with appropriate actions. The quality in use of the approach is confirmed in another action research project with an industry partner. In summary, this thesis supports organizations in quantifying the delivered value of an IoT service and controlling the value provision with effective actions. Furthermore, the trust of a consumer in the IoT service provided by an IoT system and in the organization can be maintained and further increased by applying appropriate feedback loops.Item Open Access Large-scale analysis of textual and multivariate data combining machine learning and visualization(2022) Knittel, Johannes; Ertl, Thomas (Prof. Dr.)