71446 research outputs found
Sort by
Du Bois complexes and singularity theory
No full textIn this dissertation, we study the Du Bois complexes of varieties over complex
numbers, which are generalizations of K¨ahler differentials for singular varieties,
and explore their applications in singularity theory.
In the first part, we introduce new notions of m-Du Bois and m-rational
singularities, extending the existing definitions in the case of local complete intersections
(LCI), to include natural examples beyond this setting. These notions
provide a gradual refinement of notions associated with singularities, offering a
measure of how far a variety is from being smooth, ranging from Du Bois singularities
to smoothness. We show that varieties with m-rational singularities are
m-Du Bois, extending previous results of Mustat¸ă-Popa and Friedman-Laza in
the LCI and the isolated singularities cases. We also study other properties of
these singularities, such as partial Hodge symmetries.
The second part of the dissertation develops generic vanishing theory in the
singular setting. Using Du Bois complexes, we establish appropriate generic
vanishing theorems for singular varieties, generalizing the well-known generic
vanishing theorem by Green and Lazarsfeld and the generic vanishing theorem
of Nakano type by Popa and Schnell. Our results clarify the limitations of the
naive generalizations, which were pointed out to be false by Hacon and Kovács.Mathematic
From Understanding to Improving Artificial Intelligence: New Frontiers in Machine Learning Explanations
No full textAs machine learning systems increasingly shape outcomes in high-stakes domains, the need to understand, trust, and effectively guide their decision-making grows urgent. This dissertation advances the field of machine learning explainability, offering a cohesive framework for enabling AI systems whose underlying reasoning is transparent, resilient, and actionable. By examining three critical frontiers—explainability amidst adversarial robustness, scalable rationale generation for large language models (LLMs), and decoding LLM behavior under iterative prompting—this work illuminates how explanations can inform, protect, and empower stakeholders.
The first part reveals how adversarial training, while bolstering model security, can inadvertently undermine the provision of meaningful, low-cost algorithmic recourse. This tension exposes trade-offs between securing decision boundaries and preserving explanations that help individuals improve their predicted outcomes. The second part introduces a novel approach to scaling explanations without human annotation, integrating post hoc attributions from smaller, more interpretable proxy models directly into LLM prompting. This not only reduces the need for manual rationales but also demonstrates that automatically generated explanations can actively guide complex models toward more coherent and well-founded reasoning.
The final part focuses on decoding LLM behavior through iterative prompting. While one might expect repeated user-model interactions to improve understanding and truthfulness, naïve iterative prompting can paradoxically degrade factual alignment and confidence calibration. By carefully analyzing how LLMs respond to iterative queries, the dissertation uncovers new insights into model tendencies, including over-apologizing and sycophantic patterns, and develops strategies to mitigate these issues. This examination shows that how we interact with models—how we request, refine, and interpret explanations—fundamentally shapes model reliability and clarity.
Collectively, these contributions emphasize that robust, scalable, and iteratively refined explanations are both feasible and vital. By reconciling adversarial defenses with user-friendly recourse, automating rationales for complex models, and decoding LLM behaviors through iterative engagement, the dissertation provides a principled path toward AI systems whose inner workings can be understood, trusted, and responsibly guided by human stakeholders.Engineering and Applied Sciences - Computer Scienc
Pharmacy Networks, Prescription Drug Access, and Patient Outcomes
No full textThis thesis comprises three essays on pharmacy networks, examining their roles in health outcomes, consumer welfare, and patient selection. The first chapter analyzes the effects of losing pharmacy access due to periodic network rearrangements by Medicare Part D prescription drug plans, whereas the last two chapters focus on preferred pharmacies---a selective tier of pharmacies that provide reduced cost-sharing for consumers. Overall, my results reveal that the design of both general and preferred pharmacy networks significantly affects various patient outcomes, including drug utilization and mortality. In the first chapter, I leverage a matched difference-in-differences research design to demonstrate that disrupting in-network access to pharmacies leads to an immediate 5% decline in medication adherence. Although adherence rates for affected patients partially recover over time, their mortality rates increase by 23% after one year. The second chapter investigates a different type of network shock---the initial adoption of a preferred pharmacy tier. I show that one year after a pharmacy gains preferred status, its patients experience a 15% reduction in out-of-pocket costs, a 1.3% increase in days' supply of prescriptions, and a 11.6% decline in mortality. I then estimate a model of plan and pharmacy demand to derive a measure of consumer willingness to pay for pharmacy networks based on revealed preferences. Counterfactual analyses indicate that consumers vastly undervalue the cost-sharing and mortality benefits of preferred pharmacies. Given their significant impact on healthcare provision, the third chapter studies whether preferred pharmacy networks are employed by insurance plans as a tool for consumer screening. Using an instrumental variable approach, I find that variation in local enrollees' exposure to Medicare's annual risk adjustment updates incentivizes plans to contract with pharmacies that serve more profitable patients. This underscores the need for regulatory reforms to ensure equitable access to the advantages offered by preferred pharmacy networks.Economic
Compulsory Voting: A Mathematical Analysis
No full textTo address the policy question of whether a country should adopt compulsory voting, this thesis develops a novel mathematical framework for evaluating the trade-offs between compulsory and voluntary voting. We use a heterogeneous preference model with costly voting to quantify the social-welfare implications of both systems. Our analysis reveals that, although compulsory voting incurs higher participation costs, it can elect a more representative candidate---particularly when socioeconomic disparities cause voting costs to be correlated with candidate preference. By deriving social-welfare equations, we identify the conditions under which a switch to compulsory voting is beneficial; namely, when the benefit of electing a more representative candidate exceeds the additional voting costs.
Furthermore, we extend the framework to capture the dynamics of close elections and polarization, enhancing the model's realism. In doing so, we account for how ideological divides affect voter turnout and election outcomes. Finally, we apply our model to analyze U.S. presidential elections, providing an empirical example that demonstrates its real-world applicability. Overall, this thesis provides a quantitative foundation for policymakers considering electoral reform.Computer Scienc
Unpacking Length of Stay in Youth Homeless Shelters: A Demographic and Outcome-Based Analysis of Best Practices
No full textThis thesis examines the factors influencing the length of stay for homeless youth in shelters and evaluates the impact of shelter policies and case management practices on guest outcomes. Using anonymized quantitative data from Y2Y, a student-run youth shelter in Cambridge, Massachusetts, the study employs regression analysis and correlation tests to identify key demographic and experiential predictors of shelter duration. In parallel, qualitative interviews with case managers provide contextual insight into best practices and the support systems for homeless youth. Findings suggest that factors such as criminal justice involvement, shelter disciplinary actions, substance use, and employment status significantly affect length of stay, while qualitative data underscore the importance of external referrals, goal setting, and creating an inclusive environment. The study contributes to the limited research on youth homelessness by offering data-driven recommendations to optimize shelter practices and support pathways out of homelessness.Applied Mathematic
New Order in the Air: On Aggregation and Dispersion of an Urban Body in Nature
No full textThere is a disconnection between urban centers and the larger context of nature even for a town in a largely rural state, such as Vermont and New Hampshire. Dartmouth College, which has an identity unmistakably based on nature, and the adjacent center of the town of Hanover are experiencing this type of disconnection behind the current site of the Hopkins Center for the Arts, which connects to the town center. This condition forms an opportunity to not only expand the current performing arts facilities, which is what Dartmouth asked for, but also instill nature both physically and ideologically into the center of this urban body. Since the intervention needs to aggregate programs and resources to form a downtown experience, an oxymoronic combination of aggregation and dispersion is required in the design of such a town center. This inquiry into such paradoxical nature of an urban body in nature adds to the trend of interest in built spaces in the countryside along with its social and ideological implications, evident in OMA’s exhibition “Countryside: The Future” and Pedro Gadanho’s option studio on the rural metropolis in Portugal at the GSD in the spring of 2024. This thesis proposes additions to the Hopkins Center at Dartmouth College while densifying the adjacent town center; meanwhile, it takes in and projects out to its natural context. Eventually, it proposes an ideal of what it is like to be in an urban body in a larger context of nature.Department of Architectur
Mindfulness Intervention and its Effect on Conspiracy Theory, including Climate Denialism
No full textThis study examined the impact of a mindfulness intervention on beliefs regarding whether climate change is human caused or naturally occurring. A total of 152 participants were recruited and were randomly assigned to either the experimental group (n= 68) or the control group, (n = 84). The experimental group received a mindfulness exercise, and the control group received an mindless exercise requiring very little effort. Participants completed five additional assessments to measure their mindfulness levels, both pre and post intervention, as well as three additional assessments measuring personality traits, internal vs. external locus of control attributes, and confirmation bias. The mindfulness intervention did not influence climate change causality beliefs as hypothesized but there were certain personality traits that exhibited a significant correlation between those traits and climate change causality beliefs and mindfulness measures. A regression analysis was also performed to determine if the personality assessments would provide a predictive element of climate causal belief, and they were found to be not significant. Further studies should be conducted to strengthen the mindfulness intervention as well as increasing the participant pool. Including more participants would improve the ability to detect an effect, if there is oneExtension Studie
Faithful Dissent: The Feminist Counterpublic on the Margins of Evangelicalism
No full textThis dissertation is a historical and ethnographic study of digital post-evangelical feminist communities in twenty-first century United States. From 2004-2024, politically progressive Christian women authors and their actively engaged readers co-created religious communities in and through blogs, books, social media, podcasts, digital newsletters, and platforms. Feminists who disidentified with conservative white evangelicalism because of their inclusive theologies and progressive political ideologies created communities of resistance and contestation with like-minded co-religionists. Rejecting dominant views of gender, race, and sexuality that permeated the political Religious Right and white evangelical institutions, this group held to a progressive form of Christianity. Drawing on textual examination of digital media, participant-observation at digital and in-person events, and seventy-five semi-structured interviews, I argue that post-evangelical feminists used digital media to construct a counterpublic that resisted conservative evangelicalism by employing evangelical practices. Twenty-first century women authors, including women marginalized by race, leveraged digital and social media to become religious leaders in online spaces. These digital communities were both a haven away from the dominant evangelical public and spaces to formulate resistance to it. At the same time, digital religious communities did not replace in-person religious communities but rather served as off-ramps from evangelicalism and bridges to more progressive communities. This study reveals the promises of new media for theological minorities within religious traditions, as well as their limitations.Religion, Committee on the Study o
Smell and survive: computational prediction of olfactory circuit activity and parasite-induced behavior in Drosophila
No full textWhile the fruit fly, Drosophila melanogaster, has been a celebrated model organism for over a century,
this thesis leverages two advancements from the past decade: 1) the release of the “connectome”
– a comprehensive wiring diagram between the ≈100,000 neurons of a full adult fly brain, and 2)
the successful introduction into the laboratory of Entomophthora muscae, a parasitic fungus that manipulates
the behavior of (and ultimately kills) flies. This work employs computational tools to make
predictions at both the neural circuit level (Chapters 1 and 2, modelling the Drosophila olfactory circuit)
and the behavioral level (Chapter 3, detecting flies exhibiting signs of parasitic infection).
Chapter 1 explores how variation in neural circuit activity among individual flies underlies variation
in behavioral responses to odors. We identify a site in the periphery of the fly odor-processing
circuit in which calcium activity is predictive of an individual’s preference between a pair of odorants.
By developing a connectome-based model of the olfactory circuit, I identify wiring variation strategies
that result in patterns of simulated neural activity resembling patterns in empirical calcium. In
Chapter 2, I introduce a Bayesian workflow for performing statistical inference using connectome-based
biophysical models. Focusing again on odor processing, we study models of neuron dynamics
with few fitting parameters and bring to bear the wealth of firing rate data for different components
of the olfactory circuit. We compare classes of models with differing levels of biophysical detail, and
assess their predictive power. Finally, in Chapter 3, I present a machine learning-based classifier that
predicts with high precision whether a fly is exhibiting signs of infection by E. muscae. This tool enables
real-time identification, enabling experiments that compare infected individuals with matched
uninfected controls.
On the whole, this thesis introduces computational techniques that provide insights into how individual
brains differ and how to integrate data to constrain neural circuit models, and a method for
behavioral phenotyping in a novel parasite-host laboratory system.Systems Biolog
Statistical methods for transcription factor footprinting in 3D genome assays
No full textChromatin loops are drivers of gene regulation, bringing distal enhancers into close proximity with their target genes. While a subset of these long-range contacts are mediated by the architectural factors CTCF and cohesin, the mechanisms underlying the majority of these regulatory contacts remains unclear. While transcription factors and other proteins are known to be contributors to DNA looping, no methods currently exist for simultaneously profiling 3D genome structure and the DNA-binding proteins involved. This presents a significant limitation for evaluating transcription factor impact on 3D contacts and constructing high-resolution protein occupancy maps of regulatory regions. This dissertation seeks to bridge this gap by (1) developing statistical methods that quantitatively assess DNA-binding protein occupancy in 3D genome assays and (2) leveraging these new methods to (i) investigate the relationship between protein-binding and genome architecture and (ii) provide high-resolution maps of DNA-binding proteins at enhancers and promoters.
Chapter 1 proposes a statistical method that determines CTCF binding in CTCF MNase HiChIP assays. MNase, beyond its use in profiling genome structure, has also been used to determine transcription factor and nucleosome occupancy at high-resolution in assays like CUT&RUN and MNase-seq due to its endo-exonuclease activity, where it cuts regions unprotected by proteins and chews back the fragments until it reaches protein-protected DNA. This enables inference of both protein size and location, since nucleosomes protect more than twice the DNA of a transcription factor, thus yielding DNA fragments of significantly longer length. We leverage short, transcription factor (TF) protected fragments to pinpoint locations of CTCF binding at base-pair resolution. We then use TF-protected fragments at CTCF binding sites to implement a novel fragment-level view of CTCF-mediated chromatin looping dynamics. With this approach, we determine that fully extruded chromatin loops between convergent CTCF-bound sites are rare genome-wide and that, in addition to CTCF, active regulatory elements hinder cohesin-mediated loop extrusion. This supports a model by which the partially extruded chromatin loop can enable distal enhancer-promoter contacts.
Chapter 2 expands upon the method developed in Chapter 1 by broadly mapping locations of DNA-binding transcription factors in Micro-C. Unlike Chapter 1, Chapter 2 does not rely on a ChIP step to profile one protein’s occupancy. Not being limited to just investigating one protein, CTCF, facilitates additional novel insights into protein occupancy and their relation to 3D contacts and transcription. We find that expression level is tightly linked with the size of the nucleosome depleted region at the TSS and the presence of a large TF complex at the promoter, immediately upstream of the TSS, with unexpressed genes exhibiting a TSS obstructed by a nucleosome and a lack of TF binding. Furthermore, the TF-sized proteins upstream of the TSS at expressed genes appear to facilitate long-range, cohesin-independent looping contacts, which may partially explain why gene expression is largely maintained when cohesin is degraded. Further investigation into whether specific TFs may enable these long-range cohesin-independent contacts identified transcription factor motif families such as the KLF/SP and NF-Y motif families as likely candidates for cohesin-independent looping factors.
Chapter 3 applies the approach developed in Chapter 2 to gain a high-resolution view of the MYC oncogene and its distal cell-type specific enhancers. This analysis identifies MYC distal enhancers as regions highly occupied by TF-enhancer assemblies, which depend on RNA for their coalescence. This chapter reveals the cooperation between TF binding and RNA required for gene regulation and genome structure, and presents a framework for developing high-resolution maps of regulatory architecture.Biostatistic