1983 research outputs found
Sort by
Spectromer: a Visual Transformer-Based Model for Spectral Data
No full textWe present Spectromer, a novel framework leveraging Vision Transformers (ViTs), a class of deep learning models originally developed for image recognition, for the analysis of astronomical spectral data. By converting traditional one-dimensional
spectral data, where each spectrum is represented as a sequence of intensity values over wavelength, into two-dimensional image-like representations, Spectromer enables Vision Transformers to leverage their spatial self-attention mechanism for capturing both local and global spectral features.
We fine-tune a base model, pretrained on ImageNet, using images constructed from SDSS and LAMOST spectral data, which together encompass several million spectra from diverse astronomical objects. These images are generated by transforming one-dimensional spectral data into two-dimensional image representations suitable for vision transformer architectures. We then validate this model on key downstream tasks including stellar object classification and redshift estimation, demonstrating strong performance and scalability. Spectromer has provided either comparable or better results depending on downstream tasks with other models, showing similar R^2 to AstroCLIP’s spectrum encoder even when including data from different astronomical objects as well as showing higher classification accuracy versus solutions based on Support Vector Machine and Random Forests. Our results highlight Spectromer’s potential to advance spectral analysis by leveraging pretrained vision models to enable precise interpretation of large-scale astronomical datasets beyond their original design. To our knowledge, this is the first application of ViTs to spectroscopic data and among the first to demonstrate results on a large-scale, real observational dataset without relying on synthetic data.Extension Studie
Preparing Students for College-Level Mathematics Through Secondary Advanced Placement Courses – IB, A-level and AP
No full textStudents in international schools are asked to learn subjects from foreign curricula, enabling them to pursue international education in college. Three of the more common curricula chosen by parents and students are International General Certificate of Education (IGCSE), the International Baccalaureate Diploma (IB DP) and the Advanced Placement (AP). At the end of their intensive one- or two-year studies with these foreign curricula, students must take exams that not only serve as benchmarks for schools’ reputations, but also signal the students’ readiness for college mathematics. The advanced curricula also play crucial role in shaping the students’ learning styles and significantly influence their perceptions of mathematics as they enter undergraduate studies.
This thesis examines the extent to which advanced high school mathematics exams (A-Level, IB and AP) prepare students for college-level mathematics by analyzing the cognitive demands of the associated exam questions. Past exam questions from May and June of 2021 2022, and 2023, published by the Cambridge International Examination (CIE), International Baccalaureate (IB), and College Board, are analyzed. The cognitive demands of these assessment papers are evaluated using two distinct frameworks: the Complexity, Abstractness, and Strategy (CAS) framework and the Mathematical Assessment Task Hierarchy (MATH) Taxonomy.
This thesis reveals that all three exams contain many low-cognitive demand questions that can be solved through rote memorization and mimicry of mathematical reasoning while lacking probes for students’ conceptual understanding. As a result, these exams often fail to bridge the gap between high school and undergraduate mathematics, perpetuating misconceptions about the nature of the subject. Despite these commonalities between the programs, the findings also indicate differences between the exams, which could serve as stepping stones toward undergraduate mathematics
Dynamic, Welfare-Maximizing Pooled Testing
No full textFrom isolated cases to global pandemics, testing for disease is fundamental to modern medicine. However, shortages of time, testing supplies, and staff means that it is often not possible for each biological sample to be individually tested. Thus, pooled testing is often used by the American Red Cross, medical clinics across the United States, and health centers throughout the Global South. Under pooled testing, biological samples are grouped together and tested as one sample. In this thesis, I first build upon the work of Finster et al., who developed a general model for this problem. I first provide practical implementations for the algorithms theoretically presented there, and develop methods to determine the posterior probability of health given a history of test results. I then introduce novel algorithms to improve pooled testing assignments by using a dynamic approach. In this setting, tests are sequentially assigned, increasing the expected welfare of pooled testing by conditioning on the results of past tests before determining the next. I develop and discuss four algorithms under the dynamic setting: one that computes the optimal dynamic testing allocation, a quicker but less optimal algorithm that uses a greedy heuristic, and two classes of machine learning-based algorithms, one using supervised learning and one using reinforcement learning. I evaluate these algorithms by their computational complexity, optimality guarantees, and performance across various testing instances. My results show that the benefits of dynamic testing are significant, especially under real-world constraints. While the optimal dynamic testing allocation is welfare-maximizing, the computational complexity of the algorithm makes it infeasible for practical use. The greedy dynamic-assignment algorithm yields an increase in welfare of over 2.75% compared to the previous state-of-the-art algorithm for large populations and outperforms this benchmark in over 53% of instances. While the performances of the machine learning-based algorithms are not currently competitive, the results suggest promise and motivate further work on such an approach. I also use the empirical results to discuss how various changes in constraints, such as testing budget, affect the variance of welfare yielded under dynamic plans. The results show that current pooled testing practices can be meaningfully improved through dynamic assignment, particularly through the greedy dynamic-assignment algorithm. Adopting the proposed methods would result in a variety of welfare-increasing outcomes, from family members being able to visit immunocompromised loved ones more often to an increased number of life-saving blood donations delivered to patients in need.Applied Mathematic
Associations Between Socioeconomic Status Indicators, Baseline Activity and Continued Elevated Physical Activity following a 1-Week Step Challenge
No full textPhysical activity is an important dynamic of well-being. However, in 2020, the Center for Disease Control determined that only about 46.9% of adults met the recommended standards for leisure-time physical activity and nearly half as many for those below the federal poverty line. The current study investigated how socioeconomic status may be related to engagement in a short-term walking challenge and improved physical activity following the intervention. The study aimed to understand correlations between socioeconomic status and baseline step volume(Aim 1), achievement of the intervention goal of 15% increase in step volume (Aim 2). The investigation also considered who improves in terms of step volume during the intervention period/Week 2 whether or not the met their target step goal (Aim 3), and who improves in terms of step volume during the follow-up period/Week 3 above their baseline, whether or not the met their target step goal (Aim 4), and any other possible relationship between baseline steps, goal achievement and overall improvement in physical activity (Aim 5). The sample included 17 participants, recruited broadly throughout eastern Massachusetts with local recruitment through the East Bridgewater YMCA, Center for Wellness and Health Promotion at Harvard University. Results showed minimal significant relationships between any of the socioeconomic variables and intervention variables. Occupation, when coded for active vs sedentary showed a moderate correlation to intervention achievement. Additionally, the study findings confirm prior research suggesting that
those already primed for physical activity are those who are most likely to engage and benefit from physical activity and wellness interventions, with results showing a correlation between those with higher baseline step counts more likely to have higher step counts in the follow-up week and those who achieved their target goal being most likely to have higher step counts in the follow-up week. The study failed to confirm the presumed relationship between these variables and socioeconomic status. This however may be associated to the small sample size as well as possible human error in the self- report step count measures
Orchestrating Change: The Case for Cross-Sector Collaboration in Child Welfare Transformation
No full textResearch shows that one in 17 children in the U.S. will spend time in foster care, with that rate increasing to one in nine for Black children. And even if a child isn’t removed from their family and placed into foster care, 37% of all children in the U.S.—and over 50% of Black children—will be the subject of a child welfare investigation before reaching adulthood. In addition to these racial inequities, youth in foster care experience disproportionately worse academic experiences and outcomes. Students in foster care are about twice as likely to be chronically absent or receive an out-of-school suspension and about three times as likely to be expelled compared to their peers not in foster care.
This capstone describes my residency at Foster America, a national nonprofit working to transform the child welfare system to reduce the number of kids in foster care and increase support for families. My strategic project occurred during a period of immense growth and change for the organization and consisted of three interrelated components: developing media strategies, convening cross-sector leaders through a series of events, and connecting with organizations in sectors adjacent to child welfare.
Utilizing The Panarchy Cycle systems change framework and research related to co-design, Indigenous wisdom, systems psychodynamics, psychological safety, felt accountability, and roles, my project supports the case that transformation in child welfare requires meaningful contributions from those outside of child welfare. My work holds important implications for Foster America’s organizational strategy and its approach to partnerships and calls on the child welfare sector to break out of its silo and the education sector to play a larger leadership role in child welfare transformation.Educatio
Essays in Financial Economics
No full textThis dissertation consists of three essays in financial economics with a common focus on the consequences of firm financing frictions for labor markets, entrepreneurship, and innovation. Chapter 1 provides novel evidence on the career returns to startup employment using quasi-exogenous variation in venture capital (VC) investment. The chapter shows how VC market conditions impact the labor market for startup talent, job outcomes, and worker career progression. It then documents how these effects vary with the technology-skill specificity of workers' human capital investments. Using novel administrative data on healthcare providers, Chapter 2 demonstrates the negative externalities of financial distress for employees and consumers: filing for corporate bankruptcy increases voluntary worker turnover, leading to worsened quality of care and patient health. By collecting a novel dataset on public-sector entrepreneurial finance programs, Chapter 3 provides a new look at government efforts to finance early-stage ventures, shedding light on their global scope, scale, and consequences.Economic
How antibiotic resistance evolves in patients with acute bloodstream infections
No full textAntibiotic resistance can cause treatment failure when a patient with a previously antibiotic-suscpetible infection develops a resistant infection. This can occur through three distinct processes: (1) de novo evolution, where mutations in the bacterial genome confer resistance; (2) horizontal gene transfer, where bacteria acquire resistance genes through gain of mobile genetic elements; or (3) change of strain, where the patient is reinfected with a distinct antibiotic resistant strain. Understanding these frequencies can inform treatment and improve patient outcomes; however, the relative contribution of each mechanism to resistance evolution is unclear for most types of infections. Here we quantify the frequencies of resistance gain mechanisms using a dataset of paired blood samples from patients in the Mass General Brigham hospital system. We developed a bioinformatics pipeline to classify the resistance gain mechanism between two samples from whole genome sequencing data, detecting 10 instances of de novo evolution, 4 instances of horizontal gene transfer, and 5 instances of change of strain. Our findings show that, in contrast to other infections, all three resistance mechanisms contribute to within-patient antibiotic resistance evolution in bloodstream infections. We further identify reinfection from a bacterial reservoir to be a potentially prominent but overlooked mechanism of resistance gain. We anticipate our analyses to lay the groundwork for future studies investigating resistance evolution at the within-patient scale and ultimately inform treatment through the framework of evolution.Applied Mathematic
A Comparison of Natural Language Models to Subtype Ischemic Stroke from Electronic Health Records
No full textIschemic stroke is a leading cause of disability and death worldwide. While the prevalence of ischemic stroke varies across race and ethnicity, it is particularly pronounced in low-resource medical institutions, where patients may experience higher rates of post-stroke complications and mortality. Despite concerted attempts to implement established interventions and explore novel treatment modalities, the coarse classification of ischemic stroke obscures the underlying heterogeneity in both pathophysiological mechanisms and clinical manifestations, rendering these efforts insufficient. As such, there is huge value in investigating the underlying etiology of ischemic stroke in large, diverse cohorts of patients that could power refined subtype discovery. While electronic health record (EHR) data represent a valuable resource for stroke subtyping, there are few studies that have utilized EHR data for building these cohorts due to the significant human effort required to label features and cases.
To this goal, the rapid progress of Natural Language Processing methods has opened the door to fully automated stroke subtyping. This study investigates the performance of two Natural Language Processing approaches, namely Logistic Regression, a statistical model, and Clinical Longformer, a pre-trained transformer model, in subtyping ischemic stroke directly from the EHR. The models were trained and tested on EHR data of about 3000 stroke patients adjudicated by board-certified neurologists. Both models achieve commendable performance, with the transformer model slightly outperforming the statistical model with recall and precision of 0.83 and 0.74 respectively. This finding highlights that Natural Language Processing offers a more consistent and scalable approach to subtyping ischemic stroke from EHR, which could significantly enhance the statistical power and facilitate large-scale stroke research. In particular, this outcome has enabled us to infer toast subtypes across a sizable cohort of 30,000 coded strokes at Massachusetts General Hospital, thereby opening up novel avenues for investigating stroke risk prediction and genetic underpinnings.Computer Scienc
Reading the Rasavāhinī: A Religious and Literary Reading of a Buddhist Narrative
No full textThis dissertation argues for the value of reading Buddhist narratives as literature. Specifically, it argues for the value of reading them as Literature—with a capital “L”—which is to say, for their artistry and also for their enduring relevance for readers across vast time and place. I suggest that reading Buddhist narratives, not just as religious texts but also as Literature, reveals their deepest valences and full range of intellectual and aesthetic complexity. Most importantly, I contend that it is by virtue of the very affordances of Literature that Buddhist narratives have the capacity to be ethically edifying, not just as didactic texts that convey clear ethical instructions to passive recipients, but as ethically formative exercises that engage readers and provoke introspection, reflection, and meta-reflection. This manner of engaging with Buddhist narratives is at odds with the way narrative has largely been read in the history of Buddhist studies. However, I contend that its value lies in its capacity to help readers see more in the texts, and it can help scholars better understand how religious narratives shape the moral lives of their readers.
This dissertation takes as a case study a Buddhist narrative anthology composed around the thirteenth century in Sri Lanka entitled Rasavāhinī. It demonstrates how the text both implicitly and explicitly communicates its literary aims and its status as Literature, and models the benefits and pleasures of reading it as it instructs us to read it; attentively, appreciatively, and receptive to the transformative emotions it promises to bring as a conveyer (vāhinī) of aestheticized emotion (rasa). My methodology entails a two-pronged approach. The first prong deploys the tools of close reading and sensitive reading in order to attend to the text’s form, literary strategies, and creative use of language. Shifting the focus of attention from the “moral at the end of the story” to its literary dimensions, I demonstrate the text’s surplus of meanings and its related capacity to engage readers in the active production of meaning. The second prong deploys the methods of ethical reading. By modeling a self-reflectively ethical reading, I show how Buddhist narratives cultivate and heuristically educate the moral lives of readers.
The introduction distinguishes between reading Buddhist narrative “as literature” (and between literature and Literature more broadly), versus reading it as purely doctrinal or propositional, or as a source of historical or social knowledge, as narratives have long been treated in Buddhist studies. Chapter One introduces the Rasavāhinī, the text that serves as a case study at the center of this dissertation, and explores some of the ways the text signals its literariness and instructs its audience to notice and enjoy its artistic expression. Chapter Two expands on the method of close reading, demonstrating the surplus of interpretive possibilities that emerge through a slow and sensory-focused reading that is attentive to the text’s formal features, as well as its creative use of language and literary devices. Chapter Three expands on the method of ethical reading, furthering the claim that the stories are not just morally instructive texts, but also moral exercises that provide opportunities for learning through direct, embodied cognition when they are engaged with as dynamic, open works by active and personally invested readers. In the conclusion I summarize the major arguments made in each chapter, reflect on some of the conclusions arrived at through my close and ethical readings, and propose new avenues for future scholarship on Buddhist narratives and their applications to ethics.Religion, Committee on the Study o
Clausal Deficiency
No full textThis thesis investigates the syntactic and semantic properties of clauses which are considered to be deficient in some manner, which are often called infinitival or nonfinite clauses. This thesis is concerned with three types of features that are relevant to the deficiency of a clause: (i) syntactic, for instance the ability of a clause to have an independent subject; (ii) morphosyntactic, or whether the verb of the clause in question is sufficiently marked for inflectional features like tense or agreement; and (iii) semantic, for example whether a clause is fully specified for tense. The fundamental goal of this thesis is to uncover crosslinguistic patterns regarding the syntactic, morphosyntactic and semantic properties of these deficient clauses.
Chapter 1 provides a summary of the current state of affairs in the literature regarding the nature of finiteness in generative grammar and lays the foundation for the upcoming chapters, while providing a summary of each. Chapter 2, based on a detailed analysis of the infinitives of 17 different languages, provides a novel syntactic universal regarding infinitives crosslinguistically: the inability to occur with a high complementizer, building on previous cartographic work on the complementizer domain of clauses. In addition, it provides several novel implicational universals on the left-peripheral properties crosslinguistically: for instance, all languages which allow topics within their infinitives also allow wh-elements within their infinitives, but not vice versa. Also based on crosslinguistic evidence, Chapter 3 proposes a theory on the relationship between subject size and clause size: for any two clauses in which one is larger than the other, the larger clause can have a subject that is equal to or larger than a subject in the small clause, but not vice versa. In doing so, I provide a new understanding of the null pronoun PRO in control infinitives. Chapter 4 provides a detailed analysis of the semantics of infinitival tense, concluding that all infinitives necessarily lack an independent tense specification. It comes to this conclusion via both an experimental study involving the lack of temporal de re in infinitives and a crosslinguistic survey of adjunct infinitives. I distinguish between three separate types of what has been referred to as "tenselessness" for clauses in the literature. Chapter 5 provides an alternate angle to the findings in Chapter 2 and its applications beyond just comparative syntax; namely, its implications on current theories of the origins of syntax and cartography. The goal of Chapter 5 is to show that comparative syntax is able to make a mark on our understanding of the origins of language