Alliance One Tobacco (Malawi)
Academic Research Repository at the Institute of Developing EconomiesNot a member yet
52305 research outputs found
Sort by
FL* theory of the pseudogap in the cuprates II. Critical quantum spin liquids and d-wave superconductivity
No full textPhysicsAuthor's Origina
Learning in Neural Networks: Lazy training, Feature Learning, and Fine-Tuning
No full textNeural networks trained on large amounts of data have found groundbreaking applications in language modeling, vision, and many other fields. The modern machine learning pipeline usually involves pre-training a model on a large, diverse dataset, and post-trained (e.g. fine tuned) on specialized downstream tasks. Models are able to learn good representations of the data in the pre-training stage, which is later tuned in the post-training stage. Despite the vast success of this pipeline, the exact mechanisms by which models are able to adapt their features to downstream tasks remains poorly understood.
In this thesis, we initially explore existing theoretical work on understanding questions related to over parametrization, generalization, and representation learning. To that end, we survey the literature on various mathematical techniques to answer these questions ranging from the neural tangent kernel and mean-field method to the drift martingale analysis; We do this while presenting original insights using self-contained examples and proofs.
Finally, we present original work on low-rank fine-tuning, which establishes a separation between the other learning regimes in the literature. In particular, we show that while fine-tuning is different than lazy training, it has a significantly lower sample and iteration complexity than full feature learning.Computer Scienc
Data-Driven Methods for Modeling Emissions and Atmospheric Composition
No full textThis dissertation investigates how large remote sensing datasets of atmospheric composition can be combined with traditional chemical transport models to advance understanding of pollutant budgets. Chemical transport models can simulate past and future pollutant burdens by representing atmospheric transport, reactive processes, and pollutant sources and sinks, but are subject to error in any of these components. Recent advances in chemical data assimilation and machine learning offer novel methods to combine the strengths of chemical transport models with information encoded in large measurement libraries from satellites and other instruments. I further develop and apply these methods in two core areas: fine particulate matter concentrations (focusing in East Asia) and pollutant emissions quantification (focusing on methane). Specific topics addressed in my dissertation include the following:
Quantifying surface fine particulate matter in East Asia using machine learning (Chapters 1 and 2). Inhalation of outdoor fine particulate matter (PM2.5) is a major public health burden. Surface instruments allow PM2.5 monitoring but cannot cover all areas, so satellite-based aerosol optical depth (AOD) measurements can be used in combination with machine learning to estimate gap-free surface PM2.5. Here I developed and applied a machine learning model to produce daily, high resolution maps of PM2.5 in East Asia.
Pendergrass et al. (2022) Atmos. Meas. Tech., Pendergrass et al. (2025) Atmos. Env.
Interpreting fine particulate matter trends in South Korea, 2011-2022 (Chapter 3). Despite steady reductions in precursor emissions, winter PM2.5 in South Korea has shown fluctuating trends. Here I apply results from Chapters 1 and 2 along with surface data and remote sensing products to analyze the drivers of PM2.5 concentrations. Results suggest a growing role for secondary PM2.5 production due in part to rising oxidant concentrations, sulfate reductions in favor of nitrate, and changing nighttime PM2.5 formation pathways.
Pendergrass et al. submitted to Geophys. Res. Lett.
Developing a chemical data assimilation platform and applying it to global methane emissions (Chapters 4 and 5). Satellite observations of pollutant concentrations do not offer direct information on pollutant sources. Bayesian optimization can fuse observational data with emissions inventories and constrain emissions based on both. Here I develop an open-source chemical data assimilation toolkit called CHEEREIO which uses the localized ensemble transform Kalman filter (LETKF) algorithm and the GEOS-Chem chemical transport model to optimize emissions and concentrations. I then apply CHEEREIO to methane, with a focus on explaining causes of the 2020-2022 methane surge. I attribute the surge to emissions from the tropics and use a satellite inundation product to suggest that wetlands play a key role.
Pendergrass et al. (2023) Geosci. Mod. Dev., Pendergrass et al. submitted to Atmos. Chem. Phys.Engineering and Applied Sciences - Engineering Science
Leveraging Passive User Context For Human-AI Collaboration
No full textThe rapid advancement of Artificial Intelligence (AI) powered tools, particularly Large Language Models (LLMs), understanding user intent has emerged as a fundamental challenge for creating effective, user-centric tools. Although users can articulate their goals explicitly, traditional approaches to eliciting intent—such as detailed prompts, additional examples, or formal specifications—often impose a high cognitive burden and fail to fully capture the subtleties of users’ evolving needs. This dissertation argues that augmenting context awareness, particularly the passive capture of environmental and interaction cues, can reduce ambiguity in user intent, leading to more intuitive and efficient human-AI collaboration.
We demonstrate the value of passive context awareness across three domains. In Chapter 1, we apply pragmatic reasoning in regular expression synthesis, reducing the need for exhaustive user examples by reasoning over examples not provided by the user as contextual cues to synthesize regular expressions. Chapter 2 introduces DynaVis, a dynamic interface for visualization editing that combines natural language input with dynamically generated UI widgets, showcasing how local workflow and task context can streamline iterative edits. Chapter 3 discusses MagicCopy, an AI-driven copy-and-paste tool that infers user intent by analyzing source and target applications alongside user instructions to automate cross-application data transformations.
Finally, we envision the future of AI design workflows, emphasizing the importance of two-way grounding where systems not only interpret user context but also reveal their reasoning and capabilities. Together, these contributions highlight the potential of passive context-awareness to transform interactive AI by delivering more seamless, contextually informed assistance.Engineering and Applied Sciences - Computer Scienc
The Price of a Neighbor's Hate: Assessing the Educational Impacts of the 2019 Xenophobic Uprising in South Africa
No full textXenophobic uprisings in South Africa have killed, injured, and displaced hundreds of Black, African migrants. Using school location as a proxy for exposure to xenophobic violence, I estimate a difference-in-differences model on immigrant performance in the South African National Senior Certificate Examinations (NSC). From this estimation, I find that the 2019 xenophobic uprising led to a 12 percentage point decline in immigrants’ NSC Overall Pass Rate. This effect is unique to immigrant students, whose 2019 pass rates declined by 8% relative to non-immigrants and immigrants far from attacks. Adverse impacts are, however, short-term, and do not persist past the year of exposure. For all other NSC outcomes, including Mathematics pass rates and Distinction attainment, noisy nulls obscure the full scope of the uprising’s educational impacts.Applied Mathematic
Differentiating on Diversity: How Disclosing Workforce Diversity Influences Consumer Choice
No full textCompanies are facing increased pressure to “walk the talk” on diversity, equity, and inclusion (DEI) in their operations. One specific call-to-action from stakeholders is the public disclosure of EEO-1s. Companies with 100+ employees are federally mandated to annually report the intersectional diversity data of their workforce in the EEO-1. We examine how consumers perceive the strategic decision companies make regarding whether to disclose workforce diversity information. We find no evidence that a company’s disclosure of its workforce diversity data negatively affects attitudes or perceived company commitment to diversity, even when it reveals racial disparities across job categories. Instead, we find that consumers perceive firms that disclose their workforce data more positively and to be more committed to DEI initiatives, relative to firms that choose not to disclose, particularly when these disclosures reveal diversity within the workforce.Author's Origina
Reshaping Remnants: Architecture Amid Inheritance and Uncertainty
No full textThe way we design architecture today is teleological in its approach. Architects face a crisis of relevance in an era of uncertainty and climate change, where rigid design processes fail to engage with ecological crises and material realities.
This thesis critiques the discipline’s short-sightedness and reliance on predetermined standardization, proposing an alternative design methodology that embraces inherited, irregular conditions and designs for geologic timescales—working with surplus materials, composing rather than dictating form, and allowing longevity to drive architectural response.
Using the post-strip-mining Appalachian landscape and discarded stone as fodder, Reshaping Remnants explores how hyper-local, material-driven sequencing and land-based timescales can redefine architectural practice. The project stitches waste rock and debris strewn in the aftermath of strip mining into a field of structures that restore the watershed while addressing immediate human and non-human needs. At the same time, it challenges architecture’s conventional temporality by designing for earthly timespans, suggesting that true sustainability emerges when architecture outlives its short-term purpose and transgresses the limited purview of teleological thinking.
Architecture, like all terrestrial life, is not static. How can it transcend the hubris of immediate authorship and instead transform through time, accommodating emergent futures and contextual abundance? By designing for longevity and leaving space for contingency, this approach resists the notion of fixed form, arguing that architecture becomes truly sustainable when it shifts from imposition to orchestration—choreographing the conditions it inherits rather than forcing material to obey a predetermined vision.Department of Architectur
Three Essays on Improving Measurement in the Study of Early Childhood Policy
No full textAs early childhood has grown as a subject of academic study and policy intervention internationally, there is increasing demand for measures of early childhood development and the quality of early childhood care and education settings that function well across diverse settings. This dissertation presents three essays on improving measurement of these important constructs and defines improvement as better understanding the psychometric properties of assessments, designing strategies that enhance reliability and validity, and addressing logistical and practical constraints alongside psychometric ones.
The first essay introduces a human-centered framework for evaluating a longer assessment of early childhood development (ECD) and creating a shorter version that maintains good measurement properties. This framework considers statistical, conceptual, and practical facets of items to maximize reliability, reduce complexity, and the conceptual breadth of the construct being studied. The framework is then applied to the International Development and Early Learning Assessment (IDELA), an assessment of ECD for three- to six-year-old children used in over 90 countries worldwide, to propose a balanced short form.
The second essay compares direct and caregiver-reported measures of ECD to evaluate the convention that direct assessment is a superior form of measurement. Using two datasets from the creation of the caregiver-reported ECDI2030 tool used to assess ECD with millions of two- to four-year-old children worldwide, this study evaluated the performance of item-pairs and compares the functioning of a directly assessed item against a caregiver-reported counterpart. The results suggest that direct assessment may deserve its status as a preferred form of ECD assessment, but that many applications of ECD measurement may be appropriate with caregiver reported tools.
The third essay evaluates the psychometric properties of the International Development and Early Learning Assessment-Classroom Environment (IDELA-CE), an assessment of process and structural quality of ECCE settings. It examined the factor structure of the IDELA-CE and how well it aligns with the hypothesized structure of the tool across countries. It also examines sources of measurement error stemming from the rater observing the class and the particular lesson observed. The analysis found moderate support for the factor structure hypothesized by the tool's creators and that while most variance in observed scores is attributable to stable differences between classrooms, that day-to-day fluctuations should not be ignored.
The overarching theme of this dissertation is that measurement in early childhood is a study of tradeoffs. The three chapters highlight tensions in ensuring that the validity & relevance, reliability & precision, and feasibility & cost of measurement is tailored to the use of scores.Educatio
Residential Battery Storage - Reshaping the Way We Do Electricity
No full textIn this study, we investigate households’ investments in behind-the-meter battery storage alongside rooftop solar and examine the effects of these batteries on consumers, the power market, and environmental emissions. We develop a structural estimation model of residential electricity usage that separates observed demand and consumption preferences and lets us estimate a non-financial utility that households may have for using self-generated solar power over grid-procured electricity. We call this utility nonmarket valuation, provide evidence that it is driven by sustainability and autarky desires, and relate it to the early adoption of residential storage. Applying this model to a novel data set of German households, we find that the median household has a nonmarket valuation of 0.29€ per kilowatt hour (kWh). We then show that owning storage increases a household’s electricity demand (storage rebound) and marginally increases the emissions by 57 kg CO2 / year / kWh of battery capacity. However, batteries may reduce emissions if solar penetration in the grid is sufficiently high. Lastly, we estimate that, at future technology costs, 2023 European electricity prices, and without subsidies, investing in storage is optimal for 54% of households, which would reduce the residential grid load by 38%, but, counterintuitively, also make it more variable.Accepted Manuscrip
Exploring Variations and Decision-making Processes of the Utilization of Modern Contraceptive Methods During Protracted Violence and Political Instability in Rural Haiti
No full textBackground
Haiti's protracted sociopolitical instabilities greatly impact healthcare delivery, leading to low family planning coverage of 35%. Short-term contraceptive methods are predominant, making frequent and expensive visits to healthcare facilities necessary. These instabilities also cause disruptions in healthcare services, potentially resulting in missed follow-ups and unintended pregnancies ultimately.
Methodology
We investigated family planning utilization trends and decision-making in four Haitian communes during prolonged sociopolitical instabilities through a convergent mixed-method study. Secondary quantitative data on family planning service usage from healthcare facilities were collected, alongside data on major sociopolitical events from online newspapers. Linear regressions were employed to analyze the impact of these events on family planning usage. Qualitative data were gathered from in-depth interviews and focus group discussions across the communes. Then inductive content analysis was performed using Dedoose.
Results
During the study period, 163,179 women aged 15 to 49 utilized the six most common contraceptive methods across four healthcare facilities. Simultaneously, 188 sociopolitical events were recorded, with 80% being national, 10% local, and 10% regional events.
In bivariate analysis, we found a negative correlation between stockout of common contraceptives and family planning service utilization -0.64 (95%CI: -0.79, -0.49), p.001. Strikes reduced family planning utilization -0.63 (95%CI: -2.6, 1.35) but, their effect was not statistically significant, p=0.53. Having more than two trained providers was positively correlated with family planning utilization, 1.99 (95%CI:1.8, 2.0). Surprisingly we found an increase in family planning service utilization during gang clashes 1.94 (95%CI: -0.02, 3.9), police-gang clashes 1.8 (95%CI:0.64, 2.9), and gang violence 1.8 (95%CI:0.68, 2.9). After controlling for confounders, the relationship between health facilities characteristics and family planning service utilization remained statistically significant.
Inductive qualitative analysis revealed six main themes representing three barriers: 1) accessing healthcare facilities, 2) family planning side effects, and 3) workforce challenges. Additionally, three facilitators for family planning utilization emerged: 4) integrated services, 5) interpersonal influences, and 6) socio-economic challenges.
Conclusion
Prolonged sociopolitical events impact family planning access, creating both barriers and opportunities. These events may hinder the movement of people and goods, prompting proactive strategies by healthcare professionals and family planning users. Our recommendations include expanding storage capacity, promoting long-acting contraceptives, and improving community-based distribution. The observed positive correlation between family planning service utilization and sociopolitical instabilities suggests increased demand during crises. This highlights the need for timely provisioning of FP supplies to empower communities to navigate ongoing crises effectively.Graduate Educatio