http://hdl.handle.net/10900/42126 2026-09-17T06:58:13Z http://hdl.handle.net/10900/183400 Archaeomalacology of the Iron Age Site of Muweilah: Taphonomy and Sustenance Assessment Müller García, Inés de la Fortuna Archaeomalacology, the study of mollusk remains from archaeological contexts, provides key insights into past human-environment interactions, including resource use, technology, and cultural practices. The present study examines over 40,000 mollusk remains from the Iron Age II site of Muweilah (northern United Arab Emirates), offering a rare perspective on marine resource use in an inland desert settlement. The assemblage was analyzed with respect to taxonomy, morphometry, and taphonomy to reconstruct patterns of collection, processing, and use, as well as environmental constraints. Intra-site variation was assessed using quantitative and diversity measures across excavation contexts and compared with other Southeast Arabian sites. Results show selective exploitation of a limited range of food species, with processing and cooking practices strongly influencing preservation alongside shell morphology. Non-dietary use was limited: most bivalves were likely collected post-mortem, while larger gastropods were occasionally used for ornamentation or as containers. Smaller species may have arrived incidentally. Spatial analyses reveal clear intra-site differentiation corresponding to architectural divisions. Notably, the columned hall (Building II) is marked by high oyster frequencies and low species diversity, whereas other areas show greater diversity dominated by Marcia cordata, indicating functionally distinct spaces. Regional comparisons highlight strong similarity with nearby coastal sites. In contrast, assemblages from Oman and the eastern Gulf reflect greater variability, reflecting environmental and cultural influences. Taken together, the findings demonstrate how inland communities integrated coastal resources to mitigate subsistence risk while participating in wider networks of mobility and exchange. At the same time, intra-site variation points to structured spatial organization and potentially socially differentiated activity areas. The study underscores the value of archaeomalacological evidence for understanding adaptation, subsistence, and regional connectivity in Southeast Arabia. 2026-09-16T00:00:00Z http://hdl.handle.net/10900/183399 Transformations, Equivalences and Comparisons of Learning Problems Iacovissi, Laura While artificial intelligence (AI) has achieved striking practical success across diverse applications, the theoretical foundations of many modern methods remain incomplete. In particular, a rigorous account of how the individual components of an AI system influence the outcome of its (machine) learning process is yet to be established. These components are numerous and heterogeneous, contributing to the high complexity of AI models; consequently, there is no reason to expect a single approach to fully resolve this gap. Thus, the present thesis abstracts away from the particular data points, optimization algorithms, and other technical aspects of AI systems. Paying a price in specificity, we gain a significant advantage: a framework for theoretically characterizing the interplay between the core components of a machine learning problem — loss function, model class, and data-generating probability distribution. In particular, we pay specific attention to the latter, as it is an often overlooked component in many modern works. We prove two main types of results, namely Bayes risk equalities and inequalities, obtaining equivalences and comparisons of learning problems, respectively. As a preliminary analysis, we provide an exhaustive taxonomy of stochastic transformations of a problem’s data distribution by leveraging properties of Markov kernels, i.e., the objects used to model the transformations. Studying equalities, we then learn how different Markov kernel types lead to distinct consequences. For example, applying label noise to the data distribution yields a learning problem equivalent to one that retains the original data-generating probability but with a modified loss function. By contrast, attribute noise will jointly affect both loss and model class in the equivalent problem. These findings suggest that classical loss correction techniques for noisy data are only effective in the case of label noise, not for general attribute noise. We then compute what a generalized loss correction approach would entail for attribute noise. However, focusing solely on equivalences only partially explains how tweaking one component affects the learning outcome. Studying the inequalities, we explore when one learning problem can be considered superior to another. This challenge can be addressed in multiple ways, as problems can be compared by varying the probability distribution, model class, or loss. Previous work explored comparing conditional probabilities for every loss and model class, arriving at interesting characterization results nowadays grouped under the Blackwell-Sherman-Stein theorem. This work has found application in economics as well as learning theory, and nicely relates to classical results in information theory. Our contribution frames the existing results into the larger perspective of comparing learning problems instead of conditional probabilities. We then focus on comparing couples of model classes and losses for every probability, introducing a theory complementary to Blackwell’s work, revolving around establishing when a decision maker is better than another. We study this question by means of Bayes risk orderings, which are proved to be equivalent to the inclusion ordering on certain sets, called superprediction sets. These sets are defined starting from fixing a loss and model class, and therefore provide geometrical understanding to the question of comparing decision makers. We additionally prove sufficient conditions for the ordering to hold on certain types of loss and model classes, which are decided by the type of noisy data considered. Hence, these results additionally show the central role of knowing what are the data at hand so to be able to understand the related learning problem. This work advances theoretical understanding of machine learning systems as highly entangled entities. Its first set of results underscores the critical importance of choosing your data carefully, as they demonstrate that changing the distribution is equivalent to changing the loss and model class. This can cause strong mismatches between what one expects the machine to learn and what the machine actually learns. In particular, the equalities proved offer guidance for designing corrections that aim to achieve accurate learning in the presence of noisy data. The second set of contributions introduces a novel theory for comparison of models and losses, and in particular for comparing them before and after a transformation is applied to the learning problem. This is relevant in the field of model complexity, as it introduces a decision-theoretic inspired framework to assess whether a certain decision maker is more reliable (in terms of risk) than another under a certain type of noise. 2026-09-16T00:00:00Z http://hdl.handle.net/10900/183385 Nanoscale Investigation on Linear and Nonlinear Optical Properties in Layered and Nanostructured Materials Zhao, Yang This dissertation presents a systematic nanoscale investigation of linear and nonlinear optical properties in layered and nanostructured materials, with particular emphasis on how local structural variations determine optical responses under far-field and polarization-resolved excitation conditions. Twodimensional transition metal dichalcogenides (2D-TMDCs) are widely studied due to their strong excitonic effects, symmetry-dependent nonlinear optical responses, and promising applications in optoelectronic and valleytronic devices. In addition, plasmonic nanostructures provide strong light confinement and local electromagnetic field enhancement through localized surface plasmon resonances, making them important platforms for nanoscale photonics, sensing, and nonlinear optical enhancement. Therefore, these two material systems are chosen as representative platforms to investigate geometry-dependent optical responses. By combining confocal scanning optical microscopy, ultrafast nonlinear spectroscopy, Raman spectroscopy and back focal plane imaging, this work establishes direct correlations between morphology, symmetry, dipole orientation, and optical emission behavior.; Die Dissertation ist gesperrt bis zum 30. Juni 2028 ! 2028-06-30T00:00:00Z http://hdl.handle.net/10900/183382 Der Einfluss von Schlaf auf das Erlernen mehrdimensionaler Stimulus-Response-Assoziationen Müller, Carolin Der Mensch ist im Alltag zahlreichen Reizen ausgesetzt, die mit passenden motorischen Reaktionen verknüpft werden müssen. Die vorliegende Studie untersuchte, ob die Verknüpfung visueller Reize mit einem bestimmten Aufgabenkontext und einer spezifischen motorischen Reaktion durch eine anschließende Schlafphase gefestigt wird. Hierzu wurden 47 gesunde Erwachsene in einem Between-Subjects-Design untersucht, wobei eine nächtliche Schlafphase mit einer vergleichbaren Wachphase verglichen wurde. Mithilfe eines Priming-Paradigmas wurden Stimulus-Aktions- und Stimulus-Klassifikations-Assoziationen getrennt erfasst. Die Gedächtnisleistung wurde anhand von Reaktionszeiten, Switch costs und Korrektheit bewertet. Zusätzlich wurde in der Schlafgruppe die Schlafspindelaktivität während des Non-REM-Schlafs analysiert. Für Stimulus-Aktions-Assoziationen zeigten sich schlafabhängige Veränderungen: Die Action-Switch costs nahmen nach Schlaf zu, während sie nach Wachsein abnahmen. Zudem blieb die Korrektheit nach Schlaf besser erhalten. Eine höhere Schlafspindelaktivität war außerdem mit einer stärkeren Konsolidierung von Stimulus-Klassifikations-Assoziationen verbunden. Insgesamt liefern die Ergebnisse Hinweise darauf, dass Schlaf zur Konsolidierung von Stimulus-Response-Assoziationen beiträgt, wobei insbesondere Stimulus-Aktions-Assoziationen von schlafabhängigen Konsolidierungsprozessen profitieren könnten. Die Befunde sprechen zudem für eine Beteiligung der Non-REM-Schlafspindelaktivität an der Konsolidierung dieser Gedächtnisinhalte, wobei deren Einfluss vom jeweiligen Assoziationstyp und den Aufgabenanforderungen abzuhängen scheint. 2026-09-16T00:00:00Z