33760 research outputs found
Sort by
GOVERNMENT BY CHARITY: THE EMERGENCE OF THE EARLY MODERN HOSPITAL IN THE HISPANIC WORLD
No full textThis dissertation examines the emergence of the early modern hospital as an institution in Spain and Spanish America, tracing its development through laws, intellectual debates, reform proposals, case studies, and its adaptation across both the Iberian Peninsula and the New World. While scholarship on poor relief has made valuable contributions, the institutional history of the hospital during this period remains largely overlooked. Existing studies have tended to focus instead on the regulation of mendicancy and vagrancy, the causes of poverty and epidemics, and the projection of modern welfare goals onto early institutions—often portraying them as tools for enforcing social discipline in response to the economic and ideological transformations of the sixteenth century. This study, however, demonstrates that hospitals did not arise from proto-modern economic or medical advancements that emerged in the short term, but rather from canon and civil law, evolving over time in response to enduring political and religious contexts.
Originating in monasteries and sustained by private donations, hospitals evolved into autonomous and distinctive institutions under the leadership of the Spanish monarchy and the Church. They played a key role in the territorial expansion of the Iberian kingdoms throughout the Middle Ages and the Early Modern Period. Always endowed in perpetuity, hospitals were often established with specific provisions that defined their purpose as secular spaces where individuals could carry out charitable acts grounded in biblical ordinances. This legal and religious foundation gave them remarkable stability and resilience, along with the flexibility to support the organization of local communities by channeling resources, power, and authority. While the punishment of vagrancy, the regulation of mendicancy, and colonialism spurred legal innovation and intellectual debate, hospitals followed their own trajectories of reform and redefinition, shaped by the ongoing interplay of the monarchy, the Church, and local interests
METHANE SOURCE APPORTIONMENT USING CLUMPED ISOTOPOLOGUES
No full textThe global atmospheric methane concentration continues to rise, exacerbating the greenhouse effect. This increase results from the rapid growth of methane source emissions that reflects an imbalance between sources and sinks. The development of methane clumped isotopologue (doubly-substituted methane molecules, 13CH3D and 12CH2D2) measurements provides new opportunities for methane emission source apportionment and for distinguishing methane sink intensities and pathways. However, significant uncertainties remain regarding the intensity and isotopic characteristics of various methane sources and sinks.
This dissertation provides a detailed introduction to various methane sources—categorized either by activity (fossil fuel-related, agriculture, waste, biomass burning, and wetlands) or by formation mechanisms (thermogenic, pyrogenic, and microbial)—and methane sinks (including OH, Cl, O(1D), and soil), along with their subcategories and spatial and temporal flux variations. It further discusses the bulk carbon and hydrogen isotope characteristics of these sources and the isotope effects associated with different methane formation and oxidation pathways. The discussion then extends from bulk isotopes to clumped isotopologues. My research aims to enhance our understanding of clumped isotopologue signatures for specific methane sources and sinks and to develop a global atmospheric methane budget model incorporating clumped isotopologues.
Chapter 2 introduces this novel measurement technique. Chapter 3 presents vehicle exhaust methane clumped isotopologue measurements, and provides representative signatures for abiotic and biomass burning methane. The isotope effects of abiotic catalyzed methane oxidation are also discussed. Chapter 4 presents the first clumped isotopologue measurements of atmospheric methane and demonstrates how these data refine previous estimates of the global total methane source compositions. Chapter 5 explores an attempt to use air measurements to infer methane sources and track source variations. Chapter 6 reports measurements from Arctic firn-trapped air samples, allowing the reconstruction of an atmospheric methane clumped isotopologue signal profile for the past 30 years. These time-resolved clumped isotopologue data are incorporated into a global model, providing strong constraints on the source-sink imbalance. Finally, Chapter 7 summarizes the findings and discusses future research directions
Interpreting Deep Learning Models and Unlocking New Applications With It
No full textIn recent years, modern deep learning has made significant strides across various domains, including natural language processing, computer vision, and speech recognition. These advancements have been driven by innovations in scaling pre-training data, developing new model architectures, integrating distinct modalities (e.g., vision and language, audio and language), and employing modern engineering practices. However, despite these innovations in building better models, progress in understanding these models to enhance their reliability has been relatively slow. In this thesis, we lay the groundwork for interpreting modern deep learning models—such as vision, text-to-image, and multimodal language models—by examining them through the perspectives of \textbf{data} and \textbf{internal model components}. We aim to unlock various capabilities, including model editing and model steering, to enhance their reliability. First, we build on the principles of robust statistics to interpret test-time predictions by identifying important training examples using higher-order influence functions. However, we find that influence functions can be fragile for large deep models, which limits their practical applications. To address this, we develop optimization-based data selection strategies to automatically generate stress-testing sets from large vision datasets, testing the reliability of vision models within a few-shot learning framework. Overall, our investigations show that while analyzing models through the lens of data provides valuable insights for potential improvements, it does not offer a direct method for controlling and enhancing the reliability of these models. To this end, we investigate deep models by focusing on their internal components. We develop causal mediation analysis methods to understand knowledge storage in text-to-image generative models like Stable Diffusion. Based on these insights, we create novel model editing techniques that can remove copyrighted styles and objects from text-to-image models with minimal weight updates. We scale these methods to edit large open-source models such as SD-XL and DeepFloyd.As a follow-up, we then introduce innovative causal mediation analysis methods and a richly annotated probe dataset to interpret multimodal large language models like LLaVa. Our approach allows us to understand how these models internally retrieve relevant knowledge for factual Visual Question Answering (VQA) tasks. Leveraging these insights, we develop a novel model editing method that can effectively introduce rare, long-tailed knowledge or correct specific failure modes in multimodal large language models. Using similar principles, we explore vision models (in particular the ViT architecture), developing methods to interpret image representations based on internal components such as attention heads, using text descriptions. We apply these interpretability insights to (i) mitigate spurious correlations, (ii) enable zero-shot segmentation, and (iii) facilitate text or image-conditioned image retrieval. We also extend our mechanistic interpretability techniques to understand and control language models for real-world tasks, such as context-augmented generation in question-answering systems (i.e., extractive QA). In particular, we find that insights from mechanistic circuits can be useful towards context-data attribution and model steering towards improved context faithfulness. Finally, we leverage interpretability insights from multimodal models to enhance their compositionality in image-conditioned text retrieval and text-guided image generation. For vision-language models (VLMs) like CLIP, we propose a distillation method that transfers compositional knowledge from diffusion models to CLIP. For diffusion models, we introduce a lightweight fine-tuning approach that learns a linear layer on the conditioning text encoder, improving compositional generation for attribute binding. Overall, our thesis designs and adapts interpretable methods and leverages interpretable insights to uncover various capabilities in pre-trained models
Essays on Public Defense, Juvenile Crime, and Education
No full textThis dissertation consists of three chapters in empirical microeconomics. The first chapter focuses on the public provision of defense counsel outsourced to private attorneys and explores the role of attorney quality and attorney pay affect case outcomes. I show that being quasi-randomly assigned a higher quality attorney increases the likelihood of more favorable case outcomes, and changes in attorney pay alter the performance and composition of private attorneys who provide public defense.
The second chapter focuses on the crime effects of a nationwide program targeted at youth that created digital spaces with high-speed Wi-Fi and internet-connected devices in over 1,250 community and recreation centers. My coauthor and I find the program led to meaningful decreases in juvenile crime. The third chapter evaluates the impact of non-traditional school calendars on student and teacher productivity. My coauthor and I find that while school schedules have little impact on younger children's learning, school schedules with longer and fewer school days have large negative effects on older students. The three chapters are described in further detail below.
Chapter 1. Governments increasingly outsource public services to private actors, but these contracted workers often face a "public-work pay disparity," earning less for their public service work than for work in the private market. This paper focuses on the public provision of defense counsel outsourced to private attorneys and explores two main questions: How does the quality of publicly contracted defense attorneys impact case outcomes? And, how do changes in pay influence the composition and performance of these attorneys? Using administrative data from North Carolina and the quasi-random assignment of attorneys to cases, I estimate the quality of private attorneys providing public defense. The results reveal significant variation in attorney quality as measured by the likelihood of case dismissal, incarceration, probation, and pleading guilty. I find that a one SD increase in attorney quality raises the likelihood of case dismissal by 5.5% and reduces the likelihood of incarceration by 6.5% for felony cases. A statewide hourly pay reduction led to adverse outcomes for defendants, comparable to a 1 SD decrease in attorney quality. The adverse effects are driven by lower attorney performance and a shift in the pool of contracted attorneys toward lower-quality attorneys. My results raise concerns about equity for low-income defendants and that the lack of financial resources may exacerbate broader social and economic inequalities.
Chapter 2. Many juvenile crime reduction strategies rely on policing or intensive youth interventions. We analyze the crime impacts of a nationwide program targeted at youth that created digital spaces with high-speed Wi-Fi and internet-connected devices in over 1,250 community and recreation centers. The program significantly decreased juvenile offending and victimization in high-exposure cities. We find no evidence of impacts on adult crime, nor do we find that crime was spatially displaced. This model of juvenile crime prevention appears highly cost-effective: the startup costs needed to attract youth to modernized, supervised community spaces are small compared to their potentially outsized social benefits.
Chapter 3. Firms and schools strive to increase productivity by optimally structuring the schedules of their employees and students. We analyze the impact of non-traditional school calendars on student and teacher productivity. These calendars differentially allocate mandated instructional time by choosing 1) the number of hours in the school day, 2) the number of school days each year, and 3) the distribution of school days throughout the year. To do this, we use administrative data on over 2 million students and exploit the staggered elimination of non-traditional school calendars that vary on these three dimensions. We find that while school schedules have little impact on younger children's learning, school schedules with longer and fewer school days have large negative effects on older students that are equivalent to decreasing teacher quality by nearly one standard deviation. Our results appear to be driven by changes in at-home study behavior and school start times rather than how school days are distributed throughout the year. In addition, school schedules with longer and fewer school days increase teacher turnover. Our results reveal that daily school schedules appear to impact school productivity more than yearly school calendars
AN ANALYSIS OF THE DISPROPORTIONATE DISCIPLINE REFERRAL RATES OF AFRICAN AMERICAN MIDDLE SCHOOL STUDENTS
No full textABSTRACT
Title of Dissertation: AN ANALYSIS OF THE DISPROPORTIONATE DISCIPLINE REFERRAL RATES OF AFRICAN AMERICAN MIDDLE SCHOOL STUDENTS
Tshela H. Dennis, Doctor of Education, 2025
Dissertation Directed By: Dr. Christine M. Neumerski, Associate Director of the EdD, College of Education, University of Maryland
African American students disproportionately receive discipline referrals for disrespect and disruption in comparison to White students. Even though this problem has been researched and documented for more than 50 years, (Children’s Defense Fund report, 1975; Sheets, 1996; Brooks & Johnson, 2006; Raja, 2019) and national, state and local initiatives have been developed to eradicate it, disproportionate discipline practices continue to plague our school systems. This study investigated the strategies teachers use to address subjective student behaviors in the classroom and the supports teachers need to increase their use of pre-referral interventions prior to writing an office discipline referral. During this study, the researcher focused on the DRAAS in one middle school in Shoreline School district, a school system in a mid-Atlantic state. The researcher created the acronym DRAAS to indicate the equity gap in discipline between African American students and other student subgroups being studied. DRAAS was calculated by comparing the percentage of African American students referred for subjective offenses at Shoreline Middle School to other student groups at Shoreline.Quantitative data from surveys and qualitative insights from interviews with teachers were synthesized to address four research questions: (1) What types of student behaviors do middle school teachers report result in them issuing disciplinary referrals? (2) What, if any, pre-referral interventions do teachers report using prior to writing discipline referrals for subjective behaviors? (3) What reason, if any, do teachers report for non-use of pre-referral interventions and or management strategies prior to writing discipline referrals for subjective behaviors? and (4) What supports do teachers report needing to increase their use of pre-referral interventions?
The study found that teachers believe that insufficient training, unclear schoolwide expectations, and inadequate administrative support are barriers to their effective and consistent use of pre-referral interventions for subjective student behaviors
UTOPIA
No full textThis dissertation, "Utopia," explores the enduring influence of Soviet-era visual culture on post-Soviet identity, using my personal experience growing up in Omsk, Siberia, as a lens. My artistic practice reinterprets Soviet imagery, from brutalist architecture to television broadcasts and "Raw Capitalism" advertising, revealing a shared visual code that shaped my generation. Through installations like "Utopia 4" and "Dialogue," I examine the manufactured utopia of Soviet media, contrasting it with the lived reality of post-Soviet Russia. Works such as "Smile" utilize AI to generate images based on Soviet-era visuals, highlighting the homogenizing tendencies of both propaganda and contemporary technology. By juxtaposing personal memories with cultural artifacts, I aim to create a dialogue between past ideologies and present technological advancements, reflecting on how the pursuit of idealized systems can erase individuality. This work seeks to understand and reinterpret the visual codes that shaped my upbringing, offering a critical reflection on the interplay between memory, AI, and identity
This Story
No full textThis Story follows unhoused and aggrieved child narrators who are either trafficked, abused, married off to idiotic old men or are brainwashed into actors of mayhem. It ferments at some point into the frantic protest and resistance of heinous political schemes and affliction, as it turns to the personal grief of a woman who losses her husband and her only child to state violence—and who seeks the ghost of the latter for closure. At its core the body of work comprising five long stories—three of which are interconnected and preface a larger narrative—uses the canvas of childhood within distinct sociopolitical contexts to raise questions on the living conditions of the voiceless, the have-nots and the politically oppressed in contemporary Nigerian societies: for instance, what does it mean to be human, especially when the social contract between a people and their government is fractured? How do people cope when they are denied human dignity or rights? How do they resist control, oppression, and poverty? How do they create happiness for themselves and survive
Beyond Property Graphs: Development and Applications of a Hybrid Graph Analytics System
No full textOver the past two decades, graph analytics has grown from a fairly niche discipline into a ubiquitous field touching nearly every aspect of modern computing. This growth began with the internet, which led to the standardization of OWL and RDF, and was accelerated by social media and large-scale e-commerce, which produced a slew of content that could easily be represented as a graph. Along the way, graph databases emerged as an alternative to relational databases, and graph processing systems were conceived to accelerate graph algorithms and process graph queries. Today, graphs are used to solve a huge variety of problems across domains, from recommendation systems that serve targeted advertisements, to bank and insurance claim fraud detection, to customer churn prediction, to molecular and genomic analysis.
We now stand at a new frontier of graph analytics, at the intersection of traditional graph analytics, the rapidly growing technology of graph neural networks (GNNs), and large language models. These technologies are the key to unlocking the power of generative artificial intelligence, delivering more accurate, complete, up-to-date, and hallucination-free models through the process of retrieval augmented generation (RAG). However, the sheer size and complexity of enterprise-scale graphs makes deploying a graph-based RAG difficult. Much of the existing technology in these areas was not designed to work together, or to work at the scale needed by a large enterprise.
In this dissertation, I will discuss the state of graph analytics, how it can be applied to RAG, and the challenges currently preventing universal adoption of graph-based RAG. I will then introduce my work, the BitGraph framework, which provides the keys to unlocking GPU-accelerated graph-based RAG.
In the first component of my dissertation, I will start by describing the core construction of the BitGraph framework and its subcomponents, which include the Gremlin++ query language and Maelstrom, a lightweight backend for accelerating vector operations on both the CPU and GPU. I will show how the BitGraph framework is constructed in layers, with Maelstrom as the bottom layer, Gremlin++ as the middle layer, and BitGraph as the top layer, and how this construction makes the framework extremely versatile and enables acceleration of up to 35x over an equivalent naive CPU implementation.
In the second component of my dissertation, I will analyze and discuss handling and processing graph queries at scale, focusing on the theory behind query acceleration and how it applies to the paradigms and data structures used in the BitGraph framework. I will then show how specific types of query optimizations, called traversal strategies, work together to perform end-to-end just-in-time optimization to deliver 100\% speedup on a simple query, and 40\% speedup on a more advanced query.
In the third and final component of my dissertation, I will show how to use BitGraph and its query language, Gremlin++, to solve a large-scale RAG problem. I will also discuss how I further extended the BitGraph framework to include support for vertex embeddings, which were critical in developing what I call prize-aware graph traversal, a type of graph traversal that starts from a set of initial vertices and greedily selects additional vertices in the neighborhood through embedding comparison
Numerical Study of High-Mach Water Droplet Aerobreakup and Impingement
No full textThe deformation, breakup, and shock dynamics of liquid droplets in extreme compressible environments are investigated using a combination of high-fidelity experiments and advanced numerical simulations. Studied configurations include aerobreakup in the stagnation region of high-Mach (3–5) projectile flows, and high-speed droplet impingement on solid surfaces.
Experiments conducted at Stevens Institute of Technology use acoustically levitated water droplets and high-speed shadowgraphy or Schlieren imaging to capture transient flow structures and shock interactions. Numerical simulations employ the Allaire five-equation model with high-order schemes and a dense–dilute seven-equation model incorporating velocity non-equilibrium effects, implemented in the in-house CHAMPS solver.
Results indicate that early-time aerobreakup is governed by inertial dynamics, with minimal influence from viscosity or surface tension. In the impingement case, long-time shock morphology is primarily shaped by the reflection of the droplet’s own shock wave rather than jet-induced effects.
Strong agreement is observed between simulations and experimental data. Additional insights are offered into aerobreakup mass loss mechanisms, Mach number insensitivity, and the importance of non-equilibrium modeling to capture post-impingement jetting dynamics. These findings enhance understanding of multiphase flow interactions in high-speed regimes and support the development of validated computational tools for aerospace and defense applications
Accelerating material discovery and design through integrated robotics and machine intelligence
No full textConventional material discovery and design is labor-, time-, and cost-intensive. To address these challenges, this work presents a robotics- and machine learning-integrated workflow designed to accelerate the material design process, maximize experimental throughput, minimize time and labor expenses, and lead to improved outcomes. This multi-stage framework replaces labor-intensive wet lab experiments with collaborative robots that can autonomously perform sample fabrication and testing, utilizes a feasibility constrained design space to minimize experimental failures, replaces inefficient trial-and-error cycles with active learning loops to efficiently navigate complex design spaces, and incorporates virtual synthesis and screening to effectively manage multi-objective optimization tasks, enabling tunable design requirements. This framework is enabled by the development and integration of autonomous robotic platforms for high-throughput sample preparation and characterization. The efficacy of this integrated approach is demonstrated through two distinct projects titled the predictive and generative modeling of mixed-dimensional aerogels with programmable properties, and the predictive design of sustainable biobased packaging for improved postharvest preparation. This research highlights the transformative potential of combining robotics and machine intelligence to significantly accelerate the discovery and design of advanced materials with tailored property requirements