1983 research outputs found
Sort by
A Bouquet of Petals
No full textThis thesis includes opening chapters from an adult literary novel set three years in the future and tentatively titled A Bouquet of Petals. Told in retrospect, the fictional work follows fifteen-year-old Josie Benner as she navigates coming-of-age at a time when the U.S. birthrate has plummeted by thirty-three percent. In the shadow of her mother—a renowned sociologist—Josie begrudgingly relocates from Maine to Cambridge where she becomes privy to university students analyzing the links between fertility and forever chemicals in the soil, microplastics in the water, global warming and
our nation’s food supply. Everyone talks of consequence. No one speaks of choice.
But as violence against women increases, so does Josie’s awareness. A trusted new friend helps Josie realize what’s at stake for women—and children—when healthcare is metered out by politics before medicine. And Josie can’t help notice
something more. Resistance? And a secret language? One without words, but manifested in quiet brushes on the subway, petals of paper tucked under a tip at a café. So much like the small note she’d found in her home by the shore.
As Josie comes to understand the discreet, hushed correspondence of women and the true revolution it’s inspired, she can’t help think her father’s disappearance might have something to do with the rebellion of women banding together in the silent solidarity of protest and empty wombs—a new kind of choice—in a quest to protect their human (and reproductive) rights.Extension Studie
Deep Learning as a Scientific Method and a Model Organism of Intelligence
No full textThe nature and origin of intelligence is a fundamental question in science that has been studied throughout history in psychology, neuroscience, and artificial intelligence. Recent advances in machine learning point to a promising direction: deep learning. Training neural networks by optimizing their parameters via gradient descent has shown success in both practical AI applications and in the pursuit of artificial general intelligence. This thesis investigates both the practical applications of deep learning and its scientific foundations.
The first part focuses on using deep learning to accelerate experimental neuroscience. I present two applications developed during my PhD: one that uses deep learning and synthetic data generation to track neurons in multi-channel 3D videos with improved efficiency by generating training data for rare postures not covered in the original dataset; and another that employs an uncertainty-aware system to actively guide electron microscope image acquisition in real time, achieving higher throughput by focusing the time budget on critical pixels.
The second part addresses the robustification of scientific data analysis using deep learning. I discuss how neural networks can either correct systematic errors in data or generate synthetic samples for better calibrated error estimates. The first approach is applied to hyperspectral data to remove cloud shadow’s effects on acquired spectra, while the second is used to generate probabilistic dark matter maps that quantify uncertainties in density fields without known ground truth.
The third part examines how intelligent abilities emerge in modern AI models during training. I first explore how AI models learn underlying concepts and compose them, discovering that compositional abilities may emerge without obvious behavioral signs. I then investigate how models develop in-context learning abilities based on their training data distribution, revealing a phase diagram composed of different algorithms the model implements.
The final part analyzes how large language models perform complex intelligent tasks. One study reveals that models generate task-specific representations in their internal activations when presented with new data generation processes at inference time. Another evaluates how language models integrate new information into their internal world models. I conclude by discussing the fundamental cognitive abilities that current models need to improve on to arrive at a general form of intelligence.
In summary, this thesis presents investigations of deep learning both as a tool to enhance scientific discoveries and as a model organism for studying intelligence.Physic
Deployable Online Reinforcement Learning Algorithms
No full textOnline reinforcement learning (RL) algorithms are being increasingly used in real-world set-
tings where dynamic environments may render offline algorithms ineffective. Such algorithms are
desirable in this setting because they learn and improve future decision-making using continually
collected data. Applications include robotics, recommender systems, fine-tuning large language
models, and digital health. However there are many constraints to deploying online RL algorithms
in the real world. Common challenges include limited data (sparse, partially-observable, etc.), ac-
counting for the complexity of the environment, ensuring stability and autonomy of the algorithm,
and facilitating intepretability and explainability of the algorithm. In this thesis, we have the use-
inspired goal of making online RL deployable and stable in real-world settings. To do so, we provide
a full end-to-end pipeline for online RL deployment. We start with guidelines for making various
design decisions for the algorithm before deployment. We highlight the reward design as one of
the most important design decisions. Next, we provide a framework for creating a monitoring sys-
tem to ensure the algorithm runs stably and autonomously during deployment. Then, we cover
post-deployment analyses that can be conducted to (1) explain what the algorithm learning and (2)
re-evaluate algorithm design for the next deployment. To make ideas concrete in the previous three
stages, we use real examples from the online RL algorithm deployed in the Oralytics clinical trial.
Finally, we study a theoretical non-stationary bandit problem inspired by the non-stationarity in
many real world problems. We conclude by discussing various open research problems for online
reinforcement learning deployment.Engineering and Applied Sciences - Computer Scienc
Large Language Models for Automated Evaluation of Radiology Reports with Fine-Grained Scoring
No full textThe current gold standard for evaluating generated chest x-ray (CXR) reports is through radiologist annotations. However, this process can be extremely time-consuming, especially if there are large numbers of reports to evaluate. In this work, we present a Large Language Model (LLM)-based automated evaluation metric for generated CXR reports called FineRadScore. Given a candidate and a ground truth report, FineRadScore gives the minimum number of line by line corrections required to go from the candidate to the ground truth report. Additionally, FineRadScore assigns a severity rating for each correction and generates comments regarding why the correction was needed. We demonstrate that FineRadScore is able to generate the corrections in a way that aligns with radiologists and has an understanding of how clinically meaningful each error is. We also demonstrate that, when used to get a sense of the quality of the report as a whole, it aligns with radiologists at a similar level to current state of the art automated CXR evaluation metrics. Finally, we analyze FineRadScore's shortcomings to pave the way for future works.Computer Scienc
R / Python Pipelines for Biomedical LLM Semantic Search Apps
No full textLeveraging Pytorch's GPU indexing and R's data management, evaluation and visualization capabilities
At the CELEHS laboratory we are particularly interested by LLM-based embeddings as BGE and BERT. As the number of models increases, we need methods to compare their clinical usefulness. While some R packages exist to leverage GPU capabilities, Pytorch is by far more used for GPU computation. In contrast, R is efficient for data management and visualization. How should one build robust and reproducible pipelines incorporating them both ? My answer is well-designed pipelines with Docker, Makefile, and Elasticsearch. In this talk I will showcase my design approaches to such challenges.Author's Origina
Beyond the Triangle: Leading District MTSS Improvement at the Speed of Trust
No full textThis Capstone explores the persistent challenge of implementing coherent, equitable Multi-Tiered Systems of Support (MTSS) in public school districts, focusing on a strategic improvement initiative in Somerville Public Schools (SPS). Despite federal mandates and widespread adoption of MTSS as a best practice, many districts struggle to move beyond symbolic compliance. As the Superintendent Fellow and utilizing a trust-centered change management approach inspired by Frances Frei and Anne Morriss, my project aimed to improve coherence in MTSS by recoupling school-based intervention blocks (X-Block), aligning data meeting protocols, and codifying district-wide expectations in a revised District Curriculum Accommodation Plan.
Through relationship-building, targeted data collection, and compelling communication, the work engaged educators throughout the district, catalyzed early shifts in practice, and laid the groundwork for long-term change. While constraints limited full-scale implementation, the project demonstrates how district leaders can build trust, generate urgency, and activate stakeholders to drive adaptive change while leveraging technical solutions. This work offers practical insights for education leaders seeking to move MTSS from theory to impactful practice.Educatio
Designing Networks for Educational Transformation: The Art & Science of Possibility
No full textThe American high school system is facing a crisis of disengagement, with students expressing disinterest in an outdated model that prioritizes efficiency over deep learning and personal growth. This capstone examines how learning networks can be a catalyst for large-scale educational transformation, offering a future-oriented approach to high school redesign. Anchored in the research by Tony Bryk, Paul LeMahieu, Etienne Wenger, Marshall Ganz, and Ronald Heifetz, this work explores how networks can accelerate learning among systems, build collective capacity, and inspire scalable change.
As a Strategist for the Carnegie Foundation for the Advancement of Teaching, I supported the conceptualization and launch of the Future of High School (FHS) Network, a national network of high school systems moving beyond the limitations of the Carnegie Unit to advance a new Education Architecture consisting of a broader set of goals, rigorous and engaging learning experiences, and meaningful and actionable signaling systems. This capstone documents the design and execution of the FHS Network through three phases: entering a liminal state of organizational transition, conceptualizing a network that fosters deep and sustained learning, and launching a network to influence national change. By carefully recruiting and supporting twenty-five innovative school systems, the FHS Network seeks to understand the catalyzing forces needed to root, scale, and sustain a new Education Architecture.
This work argues that designing networks for transformation requires both science and art. While both are grounded in research, the science of network design involves strategic recruitment, structured knowledge-sharing, and goal-setting processes. The art, however, lies in navigating human dynamics, fostering trust, and ensuring the network remains responsive to emerging insights. Through this dual approach, the FHS Network aims to generate knowledge that not only supports its members but also informs the broader field of education.
Ultimately, this capstone contributes to the field of educational leadership by providing insights into how networks can drive large-scale change, sustain momentum, and reimagine high school as a place where all students can thrive. While it offers a roadmap for designing networks of learning, it also serves as a compass for navigating the uncertain yet hopeful terrain of national transformation.Educatio
re/collection: artist as chaplain, chaplain as artist
No full textI invited acquaintances to draw images on my arm; images that represented a story from their life. As they sat with me, taking a marker to my skin, I listened to their stories, collecting their recollections. These marks were permanently tattooed on my skin, rendering visible the everyday collection of memories and impressions. Through a consideration of storytelling and memory, I probe at the touchpoint where artist and chaplain meet. With a brief history of performance art and a poetic reinterpretation of the spiritual assessment model, this work presents collected fragments that touch on human interdependence, the body as site of memory, and the experience of caring for another.Author's Origina
Exploring Interfacial Phenomena During the Electrochemical Oxygen Evolution Reaction
No full textHydrogen plays a critical role in industrial processes such as chemical manufacturing, petroleum refining, and fertilizer synthesis—industries that together contribute approximately 4% of global carbon emissions. These industries currently utilize hydrogen produced via the emissions-intensive steam methane reforming (SMR) process, but production of green hydrogen via electrochemical water-splitting is a promising avenue for wide-scale decarbonization. The production of abundant green hydrogen also has the potential to unlock new use-cases including H2 as a direct iron reductant, H2 as an energy carrier, or H2 as a fuel precursor to decarbonize steel production, shipping, and aviation. However, the emergence of green hydrogen relies upon the development of improved water-splitting catalysts and electrolyzers that exhibit better energy efficiencies (related to the electrochemical overpotential) and lower costs (related to the catalyst materials). Water-splitting involves both the cathodic hydrogen evolution reaction (HER) and the anodic oxygen evolution reaction (OER). Of these, improvement of the OER is particularly challenging as it is a kinetically hindered process requiring four electron transfers and often involves operating conditions which corrode and destabilize earth-abundant catalyst materials.
Improvement of OER catalysts (OEC’s) and overall electrolyzer systems requires a detailed understanding of processes occurring at the electrode-electrolyte interface and further development of catalyst materials, electrolyzer systems, and techniques for studying the electrochemical interface. This thesis explores multiple phenomena occurring at the electrode-electrolyte interface including (1) the formation of local pH gradients, (2) catalyst surface deconstruction, and (3) fundamental catalyst-water interactions, as well as highlights the development of novel techniques for examining catalysts in operando. First, an acid-stable catalyst was used to investigate the formation of locally acidic environments during OER under varied operating conditions. These experiments culminated in the development of a model for quantifying local pH gradients during OER. Secondly, mixed-metal OER catalyst materials with improved OER activity were developed and the origins of activity enhancement were determined. These materials, comprised of rare-earth cations incorporated into transition-metal oxide host materials, revealed increased surface oxide deconstruction, leading to increased active site density. Surface oxide deconstruction, determined here to be the cause of increased activity, is likely a widespread phenomenon, relevant in many OEC systems. Thirdly, a surface-sensitive X-ray technique, ambient-pressure X-ray photoelectron spectroscopy (APXPS), was utilized to explicitly study the solid-water interface of mixed-metal oxide catalyst films. Interfacial catalyst-water studies reveal that lanthanide incorporation results in an increase in partially negative surface oxygen species. The mixed-metal catalysts have multiple protonation states, resulting in a buffering effect which temporarily prolongs mixed-metal catalyst operation in acidic conditions. Finally, this thesis culminates in the development of a proton-exchange-membrane (PEM) electrolyzer assembly incorporating the mixed-metal catalyst materials developed herein. The advanced technique of operando time-resolved APXPS was utilized to examine mixed-metal OEC operation in an industrial-type electrolyzer system for the first time.
This dissertation includes studies ranging from the elucidation of fundamental surface interactions to the engineering of electrolyzer assemblies for advanced characterization. These projects have all furthered our understanding of the OER, enabling future advancements in electrolyzer technologies.Chemistry and Chemical Biolog
Essays on the Social Origins of Economic and Political Behavior
No full textThe first chapter asks why people rarely move to places where they can earn higher incomes. We use individual-level data from Facebook to find that social ties play a crucial role in explaining this puzzle: social ties are concentrated locally and shape migration decisions. On average, individuals live within 100 miles of nearly 80% of their friends, with less-educated individuals having even more concentrated social networks. To establish a causal link between the location of one's friends and migration, we exploit plausibly exogenous variation in the timing of friends' moves around individuals' college graduation. Having one more friend in a given commuting zone at the time of graduation increases one's likelihood of living there by 0.3 percentage points, which is comparable in magnitude to the effect of a $470 increase in annual wages. We incorporate these findings into a spatial equilibrium model and show that the magnitude of social network effects can explain why people stay in poorer places and why less-educated people are much less responsive to economic shocks. Overall, this study shows that social networks play a first-order role --- as important or more important than canonical economic factors such as wages and rents --- in determining residential choice at the individual and aggregate level.
The second chapter asks how childhood environment shapes political behavior. We measure young voters’ participation and party affiliation in nationally comprehensive voter files and reconstruct their childhood location histories based on their parents’ addresses. We compare outcomes of individuals who moved between the same origin and destination counties but at different ages. Those who spend more time in the destination are more influenced by it: Growing up in a county where their peers are 10 percentage points more likely to become Republicans makes them 4.7 percentage points more likely to become Republican themselves upon entering the electorate. The effects are of similar magnitude for Democratic partisanship and turnout. These exposure effects are primarily driven by teenage years, and they persist but decay after the first election. They reflect both state-level factors and factors varying at a smaller scale such as peer effects.
The third chapter uses friendship data from Facebook to study the social integration of Syrian migrants in Germany. Our analysis establishes five key findings: (1) Places differ substantially in their propensities to socially integrate migrants. This regional variation in integration outcomes largely reflects causal place-based effects. (2) Spatial variation in migrants' social integration can be decomposed into the rate at which Germans befriend their neighbors in general and the particular rate at which they befriend migrants versus other Germans. We follow the friending behavior of Germans that move across locations to show that both forces are more affected by local institutions and policies than by persistent individual characteristics or preferences of local natives. (3) Integration courses causally affect place-specific equilibrium integration levels by increasing the rate at which Germans befriend Syrian migrants. (4) Social integration helps migrants obtain help from natives across a range of settings such as finding jobs and housing. (5) Natives quasi-randomly exposed to a migrant in high school are more likely to befriend other migrants later in life.Political Economy and Governmen