71446 research outputs found
Sort by
Intuition in Silico: Representational Alignment of Deep Neural Networks with Human Brain Dynamics in Intuitive Reasoning
No full textPredictive reasoning—the capacity to anticipate future states—lies at the core of both biological cognition and artificial intelligence, guiding decisions in diverse domains ranging from object manipulation to strategic planning. In humans, a growing body of research points to a specialized “intuitive physics network” in parietal and frontal regions that supports rapid judgments about stability, collisions, and other physical events. However, standard deep neural networks (DNNs) trained on image recognition tasks often fail to replicate such nuanced physical inference; they can succeed at object categorization without internalizing the causal and dynamic structure underlying real-world interactions. This thesis investigates whether tailoring DNN training objectives—via explicit stability classification—can produce internal representations more aligned with the human intuitive physics network. We employ Representational Similarity Analysis (RSA) to compare layer-wise activations from multiple network architectures (including untrained, ImageNet-pretrained, and video-pretrained models) to pre-existing functional Magnetic Resonance Imaging (fMRI) data collected as participants judge block-tower stability. By correlating model-based and brain-based representational dissimilarity matrices, we measure how closely each network’s encoding mirrors neural activity in parietal regions. Preliminary findings suggest that specialized “physics modules,” or other task heads, can enhance alignment with dorsal-stream fMRI signals compared to conventional architectures. This highlights the importance of task-specific learning objectives for capturing human-like physical inference. More broadly, the findings inform debates about whether the human brain’s intuitive physics emerges from a dedicated forward-simulation mechanism or from a hybrid model that blends simulation with learned heuristics.Computer Scienc
Post-operative Pain Control with Resolvin E1 and Lipoxin A4 Releasing Scaffold Following Periradicular Surgery
No full textMechanical allodynia is a critical indicator of pain hypersensitivity following surgical procedures. The mechanisms underlying mechanical allodynia typically involve nerve fiber sensitization and alterations in pain processing, making it an essential target for therapeutic intervention to enhance patient outcomes post-surgery. Traditional pain management strategies, such as the use of analgesics and anti-inflammatory agents, frequently fail to provide sufficient relief and are associated with various side effects, underscoring the need for novel and more effective treatment approaches. This study aimed to assess the impact of hydrogel-based treatments on mechanical allodynia in rats that underwent mandibular osseous surgery. The Von Frey filament test was used to measure mechanical allodynia at baseline, Day 4, and Day 7 post-surgery. Fifteen rats were divided into three groups: Hydrogel (HG), Hydrogel + TCP (HG+TCP), and Hydrogel + Specialized Pro-Resolving Mediators (HG+SPM). Statistical analyses, including one-way ANOVA followed by multiple comparisons and area under the curve (AUC) analysis, revealed a significant reduction (p 0.05) in mechanical allodynia on Day 4 in the HG+SPM group compared to both the HG (p = 0.032) and HG+TCP (p = 0.041) groups. While the HG and HG+TCP groups showed moderate increases in mechanical allodynia over time, the HG+SPM group exhibited a marked reduction, as confirmed by AUC analysis. The inclusion of SPM in the hydrogel formulation likely contributed to this decrease, with SPM’s known role in resolving inflammation and promoting tissue homeostasis as a potential mechanism. These findings suggest that specialized pro-resolving mediators may enhance post-operative pain management by modulating inflammation and facilitating faster tissue recovery. Although the anti-inflammatory benefits of SPM have been well-documented in other medical fields, further research is needed to explore their clinical applications in dentistry, particularly in endodontic microsurgery, where managing post-operative pain after periradicular surgery continues to be a significant challenge.Endodontic
Essays in Macroeconomics and Finance
No full textThis thesis contains three chapters in macroeconomics and finance. In the first chapter, co-authored with Pierfrancesco Mei, we propose a theory of shallow thinking to capture people’s limited understanding of the long causal chains involved in shock propagation. We cast general equilibrium as a system of causal relations in a directed cyclic graph. Estimation from our qualitative survey suggests that, on average, people understand only about 2.6 steps of propagation, overlooking much of the graph and significantly deviating from rational expectations. Our theory implies that longer causal chains have diminishing influence on beliefs. Applying shallow thinking to a New Keynesian model with active monetary policy reconciles several puzzles about long-term interest rates and inflation: (i) long-term interest rates underreact to cost-push shocks but overreact to monetary policy shocks; (ii) inflation expectations negatively predict bond excess returns; (iii) news about future cost-push shocks triggers inflation; and (iv) more persistent cost-push shocks lead to higher inflation. Notably, (iii) and (iv) contradict the predictions of rational expectations. In a real business cycle model, relative to rational expectations, shallow thinking amplifies and prolongs output fluctuations from productivity shocks and predicts negative future stock excess returns.
In the second chapter, co-authored with Xu Lu, we argue that institutional portfolio rebalancing across asset classes plays a key role in transmitting monetary shocks to equity markets. All else equal, around FOMC announcements, a stock with 10% higher ownership by institutions that routinely rebalance loses an additional 3.7 basis points following a 10-basis-point surprise rate hike. Our cross-sectional approach links this return difference to the aggregate stock market reaction through the price elasticities of both the aggregate market and individual stocks. We find that portfolio rebalancing accounts for roughly one-third to two-thirds of the aggregate stock market reaction attributed to changes in expected excess returns.
In the third chapter, I investigate the determinants of demand elasticities and distortions across industries, acknowledging that firms’ demand and pricings are interdependent in input-output networks. I propose a modeling approach of competition in markups, which micro-founds demand elasticities and markups via “network Lerner indexes” based on the network structure of the economy in general equilibrium. At the equilibrium, each firm’s markup depends on four industry-level network statistics—cycle, substitution in production, substitution in consumption, and impact on consumer price, in addition to its market share. Empirical analysis of Compustat firms from 1997 to 2019 confirms the theory predictions. Taking into account the sectoral heterogeneity explains about three times more variations in markups. Under the theory-predicted markups, the loss in total factor productivity due to misallocation is about 13% over the sample period, which is four times larger than the implied loss that ignores sectoral heterogeneity.Economic
Composable visualization systems for biological data
No full textThe field of biology demands innovative visualization approaches to handle its expanding, integrated data landscape. This need has driven advancements in visualization design—such as genome browsers, molecular renderers, and image viewers—that enhance understanding of complex datasets. However, the push for specialization and novelty has fostered development practices that produce isolated, single-purpose applications. Monolithic tools require substantial engineering and often lack interoperability, making them difficult to repurpose or integrate into broader workflows. This thesis investigates whether composable, reusable components can break down these silos and increase adaptability and impact.
To address these challenges, I introduce infrastructure that bridges layers within the visualization ecosystem for distributing interactive tools in data-centric workflows. First, I present anywidget, a high-level specification and toolkit for authoring reusable, web-based visualizations across environments like computational notebooks and standalone applications. As a standard, anywidget facilitates interchange between visualization authors and users, enabling both general-purpose and domain-specific visualizations to interoperate. Second, I introduce Viv, a web-based bioimaging toolkit that visualizes multi-terabyte datasets directly from cloud storage without software installation. Viv operates at the layer between data producers and consumers, demonstrating how alignment with open data standards enhances accessibility and scalability. Together, these contributions promote modularity, accessibility, and sustainable visualization practices within and beyond biological data systems.
Building on this infrastructure, I show both the adaptation of existing toolkits and the creation of new, composable applications in computational notebooks. I introduce Gos, a Python library that embeds the Gosling visualization grammar into data-centric environments, reducing context switching and enhancing usability for data analysis. Finally, I present a framework for comparing embedding visualizations—originally motivated by single-cell biology—that leverages the modular components developed in this thesis. Unlike standalone applications tied to specific data types, this framework is adaptable across domains and data types, highlighting the flexibility of a modular approach.
This thesis introduces practices that extend the reach and adaptability of visualization tools by rethinking software development approaches. The approach amplifies the impact of specialized tools, fostering a more connected, sustainable ecosystem for visualization across diverse research domains.Medical Science
Interferometry of Integer and Fractional Quantum Hall Edge States in Graphene
No full textIn this thesis, we develop Fabry-Pérot quantum Hall interferometers in graphene and measure anyon braiding in the fractional quantum Hall effect. Our results demonstrate the potential of van der Waals materials for constructing quantum coherent electronic devices with advanced functionalities in order to unveil a wealth of physics that is otherwise inaccessible via transport measurements. We begin by first demonstrating clear Aharonov-Bohm resistance oscillations in integer quantum Hall states, overcoming a major technical challenge of “Coulomb dominated” oscillations, which plagued decades of experiments in traditional semiconductor-based platforms. Next, we develop an improved, density-tunable interferometer and measure tunable Coulomb coupling between copropagating integer edge states, revealing the physics behind anomalous interference phase jumps and Aharonov-Bohm oscillation frequency doubling in the integer quantum Hall effect. Similar observations in other semiconductor platforms had been unexplained for a decade. The combined theoretical developments and precise tuning knobs added by our work enable further experiments probing correlations in strongly coupled one-dimensional chiral edge channels. Finally, we observe robust Aharonov-Bohm oscillations in two distinct fractional quantum Hall states, filling fractions ν=1/3 and ν=4/3, and discover 3-state telegraph noise consistent with localized anyon number fluctuations, which we put to use to directly measure the 2π/3 abelian anyon braiding phase in both states. This final work enables further experiments to demonstrate control of the localized anyon number and eventually measure the braiding properties of non-abelian anyons in even-denominator fractional quantum Hall states. Many open questions, such as whether non-abelian order describes these states, how robust topological order really is, which excitations belong to which fractional states in real devices, and whether we can build a technology leveraging the exotic physics of the fractional quantum Hall effect will soon be directly addressable.Engineering and Applied Sciences - Applied Physic
Deep-Learning to Assess Biological Aging From Spinal Dual-energy X-ray Absorptiometry
No full textSpinal dual-energy absorptiometry (DXA) is an x-ray imaging method typically used to assess bone mineral density (BMD) and diagnose osteoporosis in post-menopausal women. However, these images may contain additional information beyond BMD to predict mortality and disease risk to help estimate biological age, the concept that chronological age can differ from the actual effects of aging.
In this study, we aim to use deep learning to estimate biological age from spinal DXA images. A convolutional neural network was trained and tested on a 44,082 patient dataset from the UK Biobank database using a 5-fold cross-validation procedure and a hold-out test set of 8,951 individuals. An additional external data set from the Mass General Brigham (MGB), which consisted of 2,059 images, was used to evaluate the model on DXA obtained during routine care. We saw that the deep learning model output had a high association with several age-related incident and prevalent diseases, captured signals of aging and long-term risk for those diseases, and could categorize individuals at higher risk. In the MGB dataset, these associations were not able to be replicated, indicating that fine-tuning the model on DXAs obtained during routine clinical care may be necessary. As a result, our deep learning approach analyzing spinal DXA shows promise to be an effective way to measure biological age and as a method for doctors to help identify and prevent age-related diseases.Computer Scienc
A knowledge graph foundation model for neurological disease
No full textNeurological disorders are the leading global cause of disability, yet most lack disease-modifying treatments or cures. Efforts to understand, diagnose, and treat these conditions have been frustrated by disease complexity and heterogeneity. To address these challenges, we introduce CIPHER, a foundation model for neurological disease. CIPHER is a 578-million-parameter heterogeneous graph Transformer for neuroscientific discovery and precision medicine. CIPHER is trained on NeuroKG, a multiscale knowledge graph assembled from 36 large-scale biomedical databases and containing 147,020 nodes and 7,366,745 edges. To provide brain-specific context, NeuroKG integrates 2,480,956 neurons and 888,263 non-neuronal cells from a human brain single-cell RNA-sequencing atlas, as well as 387,483 nuclei from patients with Parkinson’s disease and matched controls. We evaluate CIPHER across a broad range of tasks, diseases, and datasets: a CRISPR/Cas9 essentiality screen in hPSC-derived dopamine neurons; GWAS and RVAS hits in Parkinson’s and Alzheimer’s; genome and proteome-wide alpha-synuclein-related experimental screens, including overexpression and proximity labeling assays and a targeted exome screen in n = 496 synucleinopathy patients; pesticides toxic to patient-derived dopaminergic neurons; FDA-approved treatments for 25 neurological diseases; and cerebrospinal fluid proteomics from n = 956 subjects representing multiple genetic subtypes of Parkinson’s disease and healthy controls. CIPHER demonstrates strong performance at disease-associated gene and pesticide discovery, in silico experimentation, hypothesis generation, therapeutic prioritization, and disease subtyping, paving the way toward AI-driven neuroscientific discovery.Computer Scienc
Utilizing Abandoned Oil and Natural Gas Wells in the LA Basin for Carbon Sequestration and Hydrogen Storage
No full textThis thesis evaluates the potential of abandoned oil and natural gas fields in the LA Basin for carbon sequestration and hydrogen storage by gaining a deeper understanding of the storage capacity of these depleted reservoirs. The research began with conducting a thorough literature review of all the data collected on the 69 well fields since the 1920s. This involved sorting through well logs and contour maps in search of tops and indicators of the top producing reservoir in each field. Identifying the top producers made it possible to define one regional horizon that represented all top producers and therefore could be used to accurately simulate the storage capacity of reservoirs across the basin. This basin-wide surface computation mesh was created by digitizing data from 20 fields, converting all the surfaces across the basin into point sets, pulling data from other surfaces, and adding regional faults from the California Statewide Earthquake Center’s (SCEC’s) Community Fault Model (CFM). After iterating different surface interpolation methods, a mesh was created that could be used to quantify storage capacity and analyze how faults could be impacted by fluid injection. Utilizing this horizon, a volumetric assessment was done to evaluate the potential of the top 5 fields across the basin. Simultaneously, the research process was started for a hydrogen flow simulation to determine how its injection could differ from carbon dioxide. Through these final results, a contribution was made to further the research into the plausibility of carbon sequestration and hydrogen storage in the LA Basin.Environmental Science and Engineering A
ACCESSIBILITY AND REAL ESTATE PRICES IN MEXICO CITY: A HEDONIC APPROACH
No full textAccessibility—the ease of reaching destinations—is widely regarded as a key driver of real estate prices. Yet much of the literature limits its focus to a single destination type, coarse spatial units, and a single accessibility metric. This thesis addresses these limitations by estimating the impact of accessibility on housing prices in Mexico City through a hedonic price model. It draws on granular real estate data and incorporates destination-specific accessibility—covering jobs, health, education, recreation, gastronomy, public spaces, and retail—using both cumulative and gravity-based metrics across public and private transport modes. The findings reveal substantial variation in how accessibility influences housing prices, with positive, negative, and nonsignificant associations depending on destination type and transport mode. By jointly estimating multiple accessibility measures within a hedonic framework, the analysis offers a nuanced comparison of how residents value different forms of access, as capitalized into housing prices.Department of Urban Planning and Desig
Essays in Finance, Technology, and Behavior
No full textThis dissertation consists of three essays that study how patterns in human behavior inform models of finance and technology. The first essay develops methods to measure economic representations from language, and uses these representations to study how the market values firms. It finds that firms can be misvalued when they are misperceived, including during waves of investor attention to new technologies. The second essay, co-authored with Keyon Vafa, builds a framework to evaluate lookahead bias in pretrained language models, and finds evidence of the bias in applications to finance and political economy. It describes how to prevent lookahead bias by using time-indexed language models in forecasting analyses. The third essay, co-authored with Johnny Tang, analyzes how partisanship relates to economic beliefs, and finds partisan differences in optimism, attention to economic topics, and interpretation of economic shocks. It discusses how these differences in stated beliefs can help to interpret partisan differences in economic decisions.Economic