Gutenberg Open Science (Univ. Mainz)
Not a member yet
11784 research outputs found
Sort by
Informationflow in Deep ReLU Networks
Get PDFDeep learning has proven its effectiveness in large parts of the scientific world. Even large-scale applications, especially text-to-image or text-to-text processors with billions of parameters, consist at their core of simple linear algebra, stacked and separated by non-linear functions. One such so-called activation function, Rectified Linear Unit (ReLU), is defined as the maximum of its argument with zero, effectively discretizing space into one of two cases: greater or smaller than zero. These mechanisms; a continuous basis (using linear algebra) and a discrete choice (using ReLU) seem sufficient to induce representations capable of tackling tasks such as Autonomous Driving or passing the Turing Test.
This thesis aims to explore the propagation of information in training deep ReLU networks, moving beyond the perspective of a solely continuous optimization process. By switching back and forth between these two ideas, continuous and discrete interpretation of the very same process, this work aims to explore different instances of the same underlying question: How does information flow from the dataset using the learning scheme through a deep network? One way to answer this question is to observe what discrete decisions a deep network implicitly makes during training and inference, leading to one of the key contributions of this work, which is to examine the activation patterns and their changes during training, enabling the analysis of architectural and optimization choices in a unified model of the training process. Using these insights, the thesis introduces ActCooLR, a proof-of-concept learning rate scheduler based on the introduced transition model of activation pattern changes. A second way to approach the question is to adaptively enhance the optimization process by incorporating additional discrete decisions using a stochastic number system during training, and monitoring optimization for this increasing difficulty.Deep Learning hat sich in vielen Bereichen der Wissenschaft als effektive Methode bewährt. Selbst groß angelegte Anwendungen, insbesondere Text-zu-Bild- oder Text-zu-Text-Modelle mit Milliarden von Parametern, bestehen im Kern aus einfachen linearen Abbildungen, gestapelt und getrennt durch nicht-lineare Funktionen. Eine solche nicht-lineare Funktion, die sogenannte Rectified Linear Unit (ReLU), ist definiert als das Maximum ihres Arguments mit Null, was effektiv den Raum in einen von zwei Fällen diskretisiert: größer oder kleiner als Null. Für Aufgaben wie autonomes Fahren oder das Bestehen des Turing-Tests scheinen diese beiden Mechanismen auszureichen: eine kontinuierliche Basis (unter Verwendung Linearer Algebra) und eine Diskretisierung des Raumes (unter Verwendung von ReLU).
Diese Arbeit beschäftigt sich mit der Propagation von Information beim Training von tiefen ReLU-Netzwerken und betrachtet diese Propagation auch jenseits des Blickwinkels eines rein kontinuierlichen Optimierungsprozesses. Durch Hin- und Herwechseln zwischen diesen beiden Ideen – kontinuierlicher und diskreter Interpretation desselben Prozesses – exploriert diese Arbeit an verschiedenen Mechanismen dieselbe zugrunde liegende Frage: Wie fließt Information vom Datensatz mithilfe der Optimierung in ein tiefes Netzwerk? Eine Möglichkeit, diese Frage zu beantworten, besteht darin zu beobachten, welche diskreten Entscheidungen ein tiefes Netzwerk implizit während des Trainings und der Inferenz trifft. Dies führt zu einem der Hauptbeiträge dieser Arbeit: der Untersuchung der Aktivierungsmuster und ihrer Veränderungen während des Trainings, was die Analyse von Architektur- und Optimierungsalgorithmen in einem vereinheitlichten Modell des Trainingsprozesses ermöglicht. Mithilfe der dadurch gewonnenen Einsichten, stellt die Arbeit “ActCooLR” vor, einen Proof-of-Concept Learning Rate Scheduler, der das vorgestellte theoretische Modell der Aktivierungsmusteränderungen nutzt. Als weitere Möglichkeit, die Frage der Informationspropagation zu beantworten, werden zusätzliche diskrete Entscheidungen mithilfe eines adaptiven stochastischen Zahlensystems in das Netzwerk eingebaut, um den Optimierungsprozess unter der resultierenden erhöhten Schwierigkeit zu überwachen.xix, 209 Seiten ; Illustrationen, Diagramm
Chemical retention and diffusional uptake of secondary organic aerosol precursors in hydrometeors
Get PDFThe phase changes of hydrometeors can cause chemical exchange that may lead to transport of certain water-soluble organic compounds (WSOC) between the lower troposphere and the upper troposphere lower stratosphere (UTLS). Specifically, during droplet freezing WSOC dissolved in the supercooled cloud droplets can be released into the gas phase. Alternately, many volatile organic compounds can be removed from the UTLS via ice phase deposition scavenging. These microphysical and multiphase chemical processes may lead to the vertical redistribution of substances that become available for atmospheric processes in the upper troposphere. Motivations for understanding the processes of chemical retention and diffusional ice uptake as well as descriptions of them are detailed in this dissertation as justification for their investigation through several publications. These publications detail the measurement of retention coefficients and ice-gas partitioning coefficients through the use of wind tunnel studies, acoustic levitator experiments, and flowtube experiments. These measurements complicate the current understanding of the relationship between chemical retention and Henry’s law solubilities as well as reveal previously unknown entropy-enthalpy compensation behavior resulting from diffusional ice uptake.viii, 160, 13 Seiten ; Illustrationen, Diagramm
Relocate to compete : a critical view on the diaspora of Russian athletes
Get PDFAmong the many areas experiencing the growing interplay between geopolitical dynamics and sport is athlete migration and sporting nationality change. These often occur as a reaction to factors such as economic incentives, career prospects, political stability, or personal safety. The current scholarship on this topic is, however, largely underdeveloped. This brief research report critically assesses data on the migration of Russian athletes following the 2022 conflict in Ukraine and related sanctions on Russia. The report reveals that the number of Russian athletes who changed their sporting nationality in order to compete continues to rise, with key sports affected at the moment being chess, figure skating, wrestling, equestrian, auto racing, and soccer. This wave of migration includes both top-level and mid-level athletes, notably featuring six medalists from the 2024 Olympics. The destinations chosen by Russian athletes post-2022 largely align with broader Russian migration patterns, while some also relocated to sporting powerhouses in the Global North. Perspectives on this migration vary depending on stakeholders: state officials stress the investments made in athlete development and call for compensation or loyalty, whereas fellow athletes frame migration as an individual necessity dictated by career prospects and longevity. Athlete migration highlights unique challenges in governance, ethics, and policy-making within sports. Based on our case study, future research directions are outlined to empirically examine the situation through the lens of sports ethics and integrity
Screening for transposon regulators in mouse development
Get PDFTransposable elements (TEs) are mobile genetic elements that contribute to genome evolution but pose potential threats to genomic stability. Their activity is tightly regulated by host defense mechanisms, including DNA methylation, histone modifications, and transcription factor (TF)-mediated control. This thesis investigates the regulation of TEs in two model systems: mouse embryonic stem cells (mESCs) and the mouse male germline.
In the first part of this thesis, we aimed to establish a CRISPR/Cas9-based screening using an innovative endogenous readout to identify TE regulators in mESCs. By generating KAP1-degron and Suv39h-dKO cell lines we aimed to monitor transcriptional and translational activation of endogenous ERV and LINE-1 elements. Despite significant efforts, technical challenges prevented the successful implementation of the screening approach. However, this work highlights the complexities of studying endogenous TE regulation and provides a foundation for future optimization.
The second part of this thesis investigates TE reactivation during DNA methylation loss in mouse spermatogenesis. Using Dnmt3CKO/KO mice, we found that young ERVs remain active throughout postnatal germ cell stages in the absence of DNA methylation, and in contrast to other studies, we observed LINE-1 reactivation before meiosis, which further increased during meiosis. DNA pulldown assays identified NRF1, a DNA methylation-sensitive TF, as a potential regulator of unmethylated TEs in the male germline and chromatin profiling analysis confirmed NRF1 binding to unmethylated TEs. Conditional knockout of Nrf1 in Dnmt3CKO/KO mice resulted in significant downregulation of IAP protein expression prior to meiosis, demonstrating NRF1 as a trans-activator of these elements. Additionally, H3K27me3 was identified as a potential barrier to NRF1 binding, suggesting a complex interplay between histone modifications and TF-mediated regulation.
Together, these studies enhance our understanding of the multilayered regulatory networks controlling TE activity in the mouse genome. While challenges remain in developing high-throughput approaches for studying TE regulation, our findings emphasize the importance of both epigenetic modifications and TFs in balancing TE repression and activation during development.132 Seiten ; Illustrationen, Diagramm
Unsupervised identification of metastable molecular conformations with deep learning methods
Get PDFThe rise of compute power over the last decades, best described by Moore’s empirical
law, has made it possible to establish simulation as the third pillar of science in
between the longstanding pillars of theory and experiment. Investigating systems ’in-
silico’ has since then become a wide-spread approach to research, enabling numerical
insights on scales not accessible to theory and experiment. In recent years, artificial
intelligence and machine learning, specifically deep learning, has emerged as one
of the key technologies of the information age, fueled by the abundant availability
of computation and data.
In this thesis, we show in two case studies that a deep learning approach to dimen-
sionality reduction, called EncoderMap, is able to find better, more descriptive col-
lective variables in the same amount of dimensions than established linear methods.
In the main chapter, we concern ourselves with improving the analysis of simulation
data by incorporating this deep learning method. Simulation can be considered as
an experiment conducted on a computer that creates a lot of raw data from which
insights can only be extracted in a nontrivial manner. This analysis follows an elab-
orate modeling pipeline, which consists of multiple steps and algorithms. One of
these crucial steps is dimensionality reduction, in which high-dimensional data is
mapped into a lower-dimensional space, retaining as much of the important informa-
tion as possible and aiming to find descriptive collective variables fit for modeling.
We show with a well-studied small peptide, deca-alanine, that the aforementioned
deep autoencoder architecture with an additional distance metric - EncoderMap -
allows to find collective variables that are at least as good as an established linear
method - TICA - in the same amount of dimensions. Connecting results, obtained
by simulation, back to experiment is done by identifying metastable states, long-
lived structural conformations that are accessible to experiment. We compare these
dimensionality reduction methods in their capabilities to find expressive collective
variables that allow to find these metastable states. Lastly, as EncoderMap does
not make use of the time-series character of the data and works on structure alone,
our results hint towards potential applications in combination with algorithms that
allow to harvest unordered data fast, e.g. Monte Carlo simulations.ix, 142 Seiten ; Illustrationen, Diagramm
The Trendermarsch sunken in the Wadden Sea (North Frisia, Germany) : reconstructing a drowned medieval cultural landscape with geoarchaeological and geophysical investigations
Get PDFLocated in the North Frisian Wadden Sea, today's tidal flat sediments cover a formerly cultivated area outside the present-day sea dike of Nordstrand that had been occupied by settlers since the Middle Ages. The intensive cultivation increased the coast's vulnerability to storm floods. Especially the medieval storm surges of the first Grote Mandränke (St Marcellus flood) in 1362 CE and the second Grote Mandränke (Burchardi flood) in 1634 CE destroyed large parts of the diked marshland.
This study focusses on the investigation of the drowned part of the medieval Trendermarsch outside the modern sea dike of Nordstrand.
We combine different geophysical and geoarchaeological methods to estimate how and to what extent anthropogenic impact has shaped the coastline of Nordstrand since the Middle Ages. Guided by the evaluation and georeferencing of historical sources and embankment plans, the geophysical prospection was initially carried out in equidistant search grids and then intensified in areas of detected anomalies. The coring locations were selected accordingly to these results.
Sedimentary, geochemical and microfaunal analyses of sediment samples of 22 sediment cores were carried out to calibrate the geophysical results and to establish a local stratigraphy of landscape evolution. Radiocarbon dating together with historical reports provides a geochronological framing.
Using magnetic gradiometry, we found imprints of at least three settlement sites and a dike segment preserved in the recent tidal flats. Reconstruction of the natural local depositional environment, based on 13 stratigraphic units, initially shows sedimentation under lagoon-like brackish–marine shallow water conditions (unit A), followed by salt marsh formation (units C.1, C.2) and subsequent development of fenlands at around ca. 800 BCE (unit D).
A hiatus between a younger marsh formation (unit C.3) at around ca. 650 CE and recent tidal flat deposition (unit G) reveals that evidence of colonisation in the High Middle Ages is mostly preserved as a “footprint”. Still, we found distinct evidence of terp enlargement that indicates different phases of settlement, presumably in response to the threats induced by storm flood events. Using brick-like ashes (unit F.1), we provide the first evidence of some type of (salt) peat processing on a local scale after the second Grote Mandränke and thus after the drowning of parts of the Trendermarsch in 1634 CE.Im nordfriesischen Wattenmeer bedecken Wattsedimente eine ehemals kultivierte Landschaft außerhalb des heutigen Seedeichs von Nordstrand, die seit dem Mittelalter von Siedlern bewohnt wurde. Die intensive Bewirtschaftung erhöhte die Anfälligkeit der Küste gegenüber Sturmfluten. Insbesondere die mittelalterlichen Sturmfluten der Ersten Großen Mandränke in 1362 n. Chr. (St. Marcellus-Flut) und der Zweiten Großen Mandränke in 1634 n. Chr. (Burchardi-Flut) zerstörten große Teile des eingedeichten Marschlandes.
Diese Untersuchung konzentriert sich auf die Erforschung der untergegangenen Gebiete der mittelalterlichen Trendermarsch, welche sich heute unmittelbar vor dem Deich Nordstrands befinden. Wir kombinieren dabei verschiedene geophysikalische und geoarchäologische Methoden, um abzuschätzen, wie und in welchem Umfang anthropogene Einflüsse die Küstenlinie von Nordstrand seit dem Mittelalter geprägt haben. Auf Grundlage von Auswertungen historischer Quellen und Bedeichungsplänen wurde die magnetische Prospektion zunächst in gleichmäßigen Suchrastern angelegt und anschließend in Bereichen mit Anomalien verdichtet. Basierend auf diesen Ergebnissen wurde die Standortauswahl der Bohrungen durchgeführt.
Zur Kalibrierung der geophysikalischen Ergebnisse und zur Erstellung einer lokalen Stratigraphie der Landschaftsentwicklung wurden sedimentologische, geochemische und mikrofaunistische Analysen von Sedimentproben aus 22 Sedimentkernen durchgeführt. Radiokohlenstoffdatierungen liefern gemeinsam mit der Auswertung historischer Berichte einen geochronologischen Rahmen.
Mithilfe der magnetischen Gradiometrie konnten wir die Abdrücke von mindestens drei Siedlungsplätzen und einem Deichabschnitt im heutigen Gebiet des Wattenmeeres auffinden. Die Auswertung des lokale Ablagerungsmilieus ergibt eine Entwicklung von lagunären Brackwasserbedingungen (Einheit A), hin zu einer Ausbildung von Salzmarschen (Einheiten C.1, C.2) und anschließenden Entstehung von Niedermooren um ca. 800 v. Chr. (Einheit D).
Der Hiatus zwischen der Ausbildung einer jüngeren Marsch (Einheit C.3) um ca. 650 n. Chr. und den rezenten Gezeitensedimenten (Einheit G) zeigt, dass die Spuren der Besiedlung im Hochmittelalter lediglich als “Abdrücke” erhalten geblieben sind. Es zeigen sich dennoch deutliche Anzeichen für Warfterweiterungen, welche auf verschiedene Phasen der Besiedlung hinweisen, vermutlich als Reaktion auf die Gefahren durch Sturmfluten.
Durch ziegelbruchartige Torfaschen (Einheit F.1) können wir für den untergegangenen Teil der Trendermarsch erstmals Spuren einer (Salz-)Torfverarbeitung auf lokaler Ebene nach der Zweiten Grote Mandränke 1634 n. Chr. nachweisen
Towards detection of molecular parity violation via chiral co-sensing : the 1H/31P model system
Get PDFFundamental weak interactions have been shown to violate parity in both nuclear and atomic systems. However, observation of parity violation in a molecular system has proven an elusive target. Nuclear spin dependent contributions of the weak interaction are expected to result in energetic differences between enantiomers manifesting in nuclear magnetic resonance (NMR) spectra as chemical shift differences in the order of parts-per-trillion to parts-per-billion (μHz to mHz) for high-Z nuclei. This method uses simultaneous measurements of diastereomeric splittings for a light and a heavy nucleus in solution-state NMR to resolve chemical shift differences persisting in a non-chiral environment between enantiomers of chiral compounds smaller than the typical high-field NMR linewidth. Sources of error must be identified and minimized to verify that the observed effect is, in fact, due to parity violation and not systematic effects. This paper presents a detailed analysis of a system incorporating 31P and 1H NMR to elucidate the systematic effects and to guide experiments with higher-Z nuclei where molecular parity violation may be resolved
Einsatz künstlicher Intelligenz als Diagnostikum in der Hautkrebsversorgung aus Sicht der Patienten und der behandelnden Ärzte
Get PDFV, 98 Seiten ; Illustrationen, Diagramm
Einfluss von körperlicher Aktivität auf Lebensqualität, Fatigue, Sarkopenie und Mortalität bei Patienten mit gastrointestinalen Tumoren
Get PDFDie vorliegende Arbeit untersucht die Rolle körperlicher Aktivität auf Lebensqualität, Fatigue,
Sarkopenie/Kachexie und Mortalität bei Patienten mit gastrointestinalen Tumoren. Eine systematische
Literaturrecherche ergab 49 Studien, die in die Analyse miteinflossen. Diese Studien
umfassen randomisiert-kontrollierte Studien (RCTs), Reviews und Meta-Analysen.
Die Interventionen, teils kombiniert mit Ernährungs- und psychologischer Beratung, fanden in
verschiedenen Therapiephasen statt. Die Ergebnisse zeigen, dass körperliche Aktivität, bei
ausreichender Dauer und Intensität (150–300 Minuten/Woche), die Lebensqualität und Fatigue
verbessert, vor allem bei Ösophagus-, Pankreas- und Kolonkarzinomen. Frühzeitige, multimodale
Programme mit Ausdauer- und Krafttraining wirken präventiv gegen Sarkopenie/
Kachexie und senken das postoperative Komplikationsrisiko sowie die krebsspezifische
Mortalität. Weitere Studien sind erforderlich, um spezifischere Empfehlungen für alle Tumorarten
zu geben.V, 151 Seiten ; Illustrationen, Diagramm
Crossing the barrier or how regulation of ovastacin controls fertilization and translates into clinical phenotypes
Get PDFThe zona pellucida, a glycoprotein matrix enveloping the mammalian egg, exerts essential functions during fertilization and early embryonic development. Its safeguard property regulates sperm entry and thus indirectly controls fertility. Limited proteolysis by the metalloproteinase ovastacin, released from the egg during fertilization, induces hardening of the zona pellucida. This precludes sperm entry and protects the embryo until implantation. However, ovastacin leakage before fertilization causes premature hardening and infertility if activity is not inhibited. This highlights the importance of ovastacin regulation by its endogenous inhibitor, fetuin-B. Accordingly, both loss and excessive ovastacin activity are linked to infertility. Here, we review recent discoveries on how ovastacin is precisely controlled to preserve zona pellucida permeability prior to fertilization and prevent penetration afterward. Based on these molecular mechanisms, we propose explanations for clinical phenotypes of recently discovered genetic mutations in ovastacin and discuss how modulation of ovastacin activity might be employed to regulate fertilization