Refine
Document Type
- Article (9)
- Doctoral Thesis (9)
- Postprint (3)
- Monograph/Edited Volume (1)
- Master's Thesis (1)
- Review (1)
Language
- English (24)
Is part of the Bibliography
- yes (24)
Keywords
- deep learning (24) (remove)
Institute
- Hasso-Plattner-Institut für Digital Engineering GmbH (6)
- Institut für Umweltwissenschaften und Geographie (4)
- Institut für Geowissenschaften (3)
- Department Linguistik (2)
- Institut für Physik und Astronomie (2)
- Strukturbereich Kognitionswissenschaften (2)
- Department Psychologie (1)
- Department Sport- und Gesundheitswissenschaften (1)
- Fachgruppe Betriebswirtschaftslehre (1)
- Hasso-Plattner-Institut für Digital Engineering gGmbH (1)
Rapidly growing seismic and macroseismic databases and simplified access to advanced machine learning methods have in recent years opened up vast opportunities to address challenges in engineering and strong motion seismology from novel, datacentric perspectives. In this thesis, I explore the opportunities of such perspectives for the tasks of ground motion modeling and rapid earthquake impact assessment, tasks with major implications for long-term earthquake disaster mitigation.
In my first study, I utilize the rich strong motion database from the Kanto basin, Japan, and apply the U-Net artificial neural network architecture to develop a deep learning based ground motion model. The operational prototype provides statistical estimates of expected ground shaking, given descriptions of a specific earthquake source, wave propagation paths, and geophysical site conditions. The U-Net interprets ground motion data in its spatial context, potentially taking into account, for example, the geological properties in the vicinity of observation sites. Predictions of ground motion intensity are thereby calibrated to individual observation sites and earthquake locations.
The second study addresses the explicit incorporation of rupture forward directivity into ground motion modeling. Incorporation of this phenomenon, causing strong, pulse like ground shaking in the vicinity of earthquake sources, is usually associated with an intolerable increase in computational demand during probabilistic seismic hazard analysis (PSHA) calculations. I suggest an approach in which I utilize an artificial neural network to efficiently approximate the average, directivity-related adjustment to ground motion predictions for earthquake ruptures from the 2022 New Zealand National Seismic Hazard Model. The practical implementation in an actual PSHA calculation demonstrates the efficiency and operational readiness of my model. In a follow-up study, I present a proof of concept for an alternative strategy in which I target the generalizing applicability to ruptures other than those from the New Zealand National Seismic Hazard Model.
In the third study, I address the usability of pseudo-intensity reports obtained from macroseismic observations by non-expert citizens for rapid impact assessment. I demonstrate that the statistical properties of pseudo-intensity collections describing the intensity of shaking are correlated with the societal impact of earthquakes. In a second step, I develop a probabilistic model that, within minutes of an event, quantifies the probability of an earthquake to cause considerable societal impact. Under certain conditions, such a quick and preliminary method might be useful to support decision makers in their efforts to organize auxiliary measures for earthquake disaster response while results from more elaborate impact assessment frameworks are not yet available.
The application of machine learning methods to datasets that only partially reveal characteristics of Big Data, qualify the majority of results obtained in this thesis as explorative insights rather than ready-to-use solutions to real world problems. The practical usefulness of this work will be better assessed in the future by applying the approaches developed to growing and increasingly complex data sets.
PyFin-sentiment
(2023)
Responding to the poor performance of generic automated sentiment analysis solutions on domain-specific texts, we collect a dataset of 10,000 tweets discussing the topics of finance and investing. We manually assign each tweet its market sentiment, i.e., the investor’s anticipation of a stock’s future return. Using this data, we show that all existing sentiment models trained on adjacent domains struggle with accurate market sentiment analysis due to the task’s specialized vocabulary. Consequently, we design, train, and deploy our own sentiment model. It outperforms all previous models (VADER, NTUSD-Fin, FinBERT, TwitterRoBERTa) when evaluated on Twitter posts. On posts from a different platform, our model performs on par with BERT-based large language models. We achieve this result at a fraction of the training and inference costs due to the model’s simple design. We publish the artifact as a python library to facilitate its use by future researchers and practitioners.
Patent document collections are an immense source of knowledge for research and innovation communities worldwide. The rapid growth of the number of patent documents poses an enormous challenge for retrieving and analyzing information from this source in an effective manner. Based on deep learning methods for natural language processing, novel approaches have been developed in the field of patent analysis. The goal of these approaches is to reduce costs by automating tasks that previously only domain experts could solve. In this article, we provide a comprehensive survey of the application of deep learning for patent analysis. We summarize the state-of-the-art techniques and describe how they are applied to various tasks in the patent domain. In a detailed discussion, we categorize 40 papers based on the dataset, the representation, and the deep learning architecture that were used, as well as the patent analysis task that was targeted. With our survey, we aim to foster future research at the intersection of patent analysis and deep learning and we conclude by listing promising paths for future work.
Casualties and damages from urban pluvial flooding are increasing. Triggered by short, localized, and intensive rainfall events, urban pluvial floods can occur anywhere, even in areas without a history of flooding. Urban pluvial floods have relatively small temporal and spatial scales. Although cumulative losses from urban pluvial floods are comparable, most flood risk management and mitigation strategies focus on fluvial and coastal flooding. Numerical-physical-hydrodynamic models are considered the best tool to represent the complex nature of urban pluvial floods; however, they are computationally expensive and time-consuming. These sophisticated models make large-scale analysis and operational forecasting prohibitive. Therefore, it is crucial to evaluate and benchmark the performance of other alternative methods.
The findings of this cumulative thesis are represented in three research articles. The first study evaluates two topographic-based methods to map urban pluvial flooding, fill–spill–merge (FSM) and topographic wetness index (TWI), by comparing them against a sophisticated hydrodynamic model. The FSM method identifies flood-prone areas within topographic depressions while the TWI method employs maximum likelihood estimation to calibrate a TWI threshold (τ) based on inundation maps from the 2D hydrodynamic model. The results point out that the FSM method outperforms the TWI method. The study highlights then the advantage and limitations of both methods.
Data-driven models provide a promising alternative to computationally expensive hydrodynamic models. However, the literature lacks benchmarking studies to evaluate the different models' performance, advantages and limitations. Model transferability in space is a crucial problem. Most studies focus on river flooding, likely due to the relative availability of flow and rain gauge records for training and validation. Furthermore, they consider these models as black boxes. The second study uses a flood inventory for the city of Berlin and 11 predictive features which potentially indicate an increased pluvial flooding hazard to map urban pluvial flood susceptibility using a convolutional neural network (CNN), an artificial neural network (ANN) and the benchmarking machine learning models random forest (RF) and support vector machine (SVM). I investigate the influence of spatial resolution on the implemented models, the models' transferability in space and the importance of the predictive features. The results show that all models perform well and the RF models are superior to the other models within and outside the training domain. The models developed using fine spatial resolution (2 and 5 m) could better identify flood-prone areas. Finally, the results point out that aspect is the most important predictive feature for the CNN models, and altitude is for the other models.
While flood susceptibility maps identify flood-prone areas, they do not represent flood variables such as velocity and depth which are necessary for effective flood risk management. To address this, the third study investigates data-driven models' transferability to predict urban pluvial floodwater depth and the models' ability to enhance their predictions using transfer learning techniques. It compares the performance of RF (the best-performing model in the previous study) and CNN models using 12 predictive features and output from a hydrodynamic model. The findings in the third study suggest that while CNN models tend to generalise and smooth the target function on the training dataset, RF models suffer from overfitting. Hence, RF models are superior for predictions inside the training domains but fail outside them while CNN models could control the relative loss in performance outside the training domains. Finally, the CNN models benefit more from transfer learning techniques than RF models, boosting their performance outside training domains.
In conclusion, this thesis has evaluated both topographic-based methods and data-driven models to map urban pluvial flooding. However, further studies are crucial to have methods that completely overcome the limitation of 2D hydrodynamic models.
Stripe rust (Pst) is a major disease of wheat crops leading untreated to severe yield losses. The use of fungicides is often essential to control Pst when sudden outbreaks are imminent. Sensors capable of detecting Pst in wheat crops could optimize the use of fungicides and improve disease monitoring in high-throughput field phenotyping. Now, deep learning provides new tools for image recognition and may pave the way for new camera based sensors that can identify symptoms in early stages of a disease outbreak within the field. The aim of this study was to teach an image classifier to detect Pst symptoms in winter wheat canopies based on a deep residual neural network (ResNet). For this purpose, a large annotation database was created from images taken by a standard RGB camera that was mounted on a platform at a height of 2 m. Images were acquired while the platform was moved over a randomized field experiment with Pst-inoculated and Pst-free plots of winter wheat. The image classifier was trained with 224 x 224 px patches tiled from the original, unprocessed camera images. The image classifier was tested on different stages of the disease outbreak. At patch level the image classifier reached a total accuracy of 90%. To test the image classifier on image level, the image classifier was evaluated with a sliding window using a large striding length of 224 px allowing for fast test performance. At image level, the image classifier reached a total accuracy of 77%. Even in a stage with very low disease spreading (0.5%) at the very beginning of the Pst outbreak, a detection accuracy of 57% was obtained. Still in the initial phase of the Pst outbreak with 2 to 4% of Pst disease spreading, detection accuracy with 76% could be attained. With further optimizations, the image classifier could be implemented in embedded systems and deployed on drones, vehicles or scanning systems for fast mapping of Pst outbreaks.
Decubitus is one of the most relevant diseases in nursing and the most expensive to treat. It is caused by sustained pressure on tissue, so it particularly affects bed-bound patients. This work lays a foundation for pressure mattress-based decubitus prophylaxis by implementing a solution to the single-frame 2D Human Pose Estimation problem.
For this, methods of Deep Learning are employed. Two approaches are examined, a coarse-to-fine Convolutional Neural Network for direct regression of joint coordinates and a U-Net for the derivation of probability distribution heatmaps.
We conclude that training our models on a combined dataset of the publicly available Bodies at Rest and SLP data yields the best results. Furthermore, various preprocessing techniques are investigated, and a hyperparameter optimization is performed to discover an improved model architecture.
Another finding indicates that the heatmap-based approach outperforms direct regression.
This model achieves a mean per-joint position error of 9.11 cm for the Bodies at Rest data and 7.43 cm for the SLP data.
We find that it generalizes well on data from mattresses other than those seen during training but has difficulties detecting the arms correctly.
Additionally, we give a brief overview of the medical data annotation tool annoto we developed in the bachelor project and furthermore conclude that the Scrum framework and agile practices enhanced our development workflow.
The quantitative analysis of microstructural features is a key to understanding the micromechanical behavior of metal matrix composites (MMCs), which is a premise for their use in practice. Herein, a 3D microstructural characterization of a five-phase MMC is performed by synchrotron X-ray computed tomography (SXCT). A workflow for advanced deep learning-based segmentation of all individual phases in SXCT data is shown using a fully convolutional neural network with U-net architecture. High segmentation accuracy is achieved with a small amount of training data. This enables extracting unprecedently precise microstructural parameters (e.g., volume fractions and particle shapes) to be input, e.g., in micromechanical models.
Search for light primordial black holes with VERITAS using gamma γ-ray and optical observations
(2023)
The Very Energetic Radiation Imaging Telescope Array System (VERITAS) is an array of four imaging atmospheric Cherenkov telescopes (IACTs). VERITAS is sensitive to very-high-energy gamma-rays in the range of 100 GeV to >30 TeV. Hypothesized primordial black holes (PBHs) are attractive targets for IACTs. If they exist, their potential cosmological impact reaches beyond the candidacy for constituents of dark matter. The sublunar mass window is the largest unconstrained range of PBH masses. This thesis aims to develop novel concepts searching for light PBHs with VERITAS. PBHs below the sublunar window lose mass due to Hawking radiation. They would evaporate at the end of their lifetime, leading to a short burst of gamma-rays. If PBHs formed at about 10^15 g, the evaporation would occur nowadays. Detecting these signals might not only confirm the existence of PBHs but also prove the theory of Hawking radiation. This thesis probes archival VERITAS data recorded between 2012 and 2021 for possible PBH signals. This work presents a new automatic approach to assess the quality of the VERITAS data. The array-trigger rate and far infrared temperature are well suited to identify periods with poor data quality. These are masked by time cuts to obtain a consistent and clean dataset which contains about 4222 hours. The PBH evaporations could occur at any location in the field of view or time within this data. Only a blind search can be performed to identify these short signals. This thesis implements a data-driven deep learning based method to search for short transient signals with VERITAS. It does not depend on the modelling of the effective area and radial acceptance. This work presents the first application of this method to actual observational IACT data. This thesis develops new concepts dealing with the specifics of the data and the transient detection method. These are reflected in the developed data preparation pipeline and search strategies. After correction for trial factors, no candidate PBH evaporation is found in the data. Thus, new constraints of the local rate of PBH evaporations are derived. At the 99% confidence limit it is below <1.07 * 10^5 pc^-3 yr^-1. This constraint with the new, independent analysis approach is in the range of existing limits for the evaporation rate.
This thesis also investigates an alternative novel approach to searching for PBHs with IACTs. Above the sublunar window, the PBH abundance is constrained by optical microlensing studies. The sampling speed, which is of order of minutes to hours for traditional optical telescopes, is a limiting factor in expanding the limits to lower masses. IACTs are also powerful instruments for fast transient optical astronomy with up to O(ns) sampling. This thesis investigates whether IACTs might constrain the sublunar window with optical microlensing observations. This study confirms that, in principle, the fast sampling speed might allow extending microlensing searches into the sublunar mass window. However, the limiting factor for IACTs is the modest sensitivity to detect changes in optical fluxes. This thesis presents the expected rate of detectable events for VERITAS as well as prospects of possible future next-generation IACTs. For VERITAS, the rate of detectable microlensing events in the sublunar range is ~10^-6 per year of observation time. The future prospects for a 100 times more sensitive instrument are at ~0.05 events per year.
For around a decade, deep learning - the sub-field of machine learning that refers to artificial neural networks comprised of many computational layers - modifies the landscape of statistical model development in many research areas, such as image classification, machine translation, and speech recognition. Geoscientific disciplines in general and the field of hydrology in particular, also do not stand aside from this movement. Recently, the proliferation of modern deep learning-based techniques and methods has been actively gaining popularity for solving a wide range of hydrological problems: modeling and forecasting of river runoff, hydrological model parameters regionalization, assessment of available water resources. identification of the main drivers of the recent change in water balance components. This growing popularity of deep neural networks is primarily due to their high universality and efficiency. The presented qualities, together with the rapidly growing amount of accumulated environmental information, as well as increasing availability of computing facilities and resources, allow us to speak about deep neural networks as a new generation of mathematical models designed to, if not to replace existing solutions, but significantly enrich the field of geophysical processes modeling. This paper provides a brief overview of the current state of the field of development and application of deep neural networks in hydrology. Also in the following study, the qualitative long-term forecast regarding the development of deep learning technology for managing the corresponding hydrological modeling challenges is provided based on the use of "Gartner Hype Curve", which in the general details describes a life cycle of modern technologies.
Neural conversation models aim to predict appropriate contributions to a (given) conversation by using neural networks trained on dialogue data. A specific strand focuses on non-goal driven dialogues, first proposed by Ritter et al. (2011): They investigated the task of transforming an utterance into an appropriate reply. Then, this strand evolved into dialogue system approaches using long dialogue histories and additional background context. Contributing meaningful and appropriate to a conversation is a complex task, and therefore research in this area has been very diverse: Serban et al. (2016), for example, looked into utilizing variable length dialogue histories, Zhang et al. (2018) added additional context to the dialogue history, Wolf et al. (2019) proposed a model based on pre-trained Self-Attention neural networks (Vasvani et al., 2017), and Dinan et al. (2021) investigated safety issues of these approaches. This trend can be seen as a transformation from trying to somehow carry on a conversation to generating appropriate replies in a controlled and reliable way.
In this thesis, we first elaborate the meaning of appropriateness in the context of neural conversation models by drawing inspiration from the Cooperative Principle (Grice, 1975). We first define what an appropriate contribution has to be by operationalizing these maxims as demands on conversation models: being fluent, informative, consistent towards given context, coherent and following a social norm. Then, we identify different targets (or intervention points) to achieve the conversational appropriateness by investigating recent research in that field.
In this thesis, we investigate the aspect of consistency towards context in greater detail, being one aspect of our interpretation of appropriateness.
During the research, we developed a new context-based dialogue dataset (KOMODIS) that combines factual and opinionated context to dialogues. The KOMODIS
dataset is publicly available and we use the data in this thesis to gather new insights in context-augmented dialogue generation.
We further introduced a new way of encoding context within Self-Attention based neural networks. For that, we elaborate the issue of space complexity from knowledge graphs,
and propose a concise encoding strategy for structured context inspired from graph neural networks (Gilmer et al., 2017) to reduce the space complexity of the additional context. We discuss limitations of context-augmentation for neural conversation models, explore the characteristics of knowledge graphs, and explain how we create and augment knowledge graphs for our experiments.
Lastly, we analyzed the potential of reinforcement and transfer learning to improve context-consistency for neural conversation models. We find that current reward functions need to be more precise to enable the potential of reinforcement learning, and that sequential transfer learning can improve the subjective quality of generated dialogues.