22689 research outputs found
Sort by
Toward the Generalization of Vision Models
Get PDFIn the field of computer vision, the ability to generalize is essential for ensuring robust and reliable performance across diverse datasets and real-world scenarios. Generalization refers to a computer vision model's capability to accurately interpret and process new, unseen data beyond the specific examples it was trained on. Real-world visual data can vary widely in terms of lighting conditions, backgrounds, viewpoints, occlusions, and other factors, making generalization crucial for effective performance. A computer vision system with strong generalization ability can identify objects, recognize patterns, and make accurate predictions even when presented with variations it has not encountered during training. Achieving robust generalization necessitates careful algorithm design, sufficient training data that captures real-world variability, regularization techniques to prevent overfitting, and robust evaluation methodologies. Without robust generalization, computer vision models may perform well on training data but fail to generalize to new, unseen data, limiting their real-world applicability.
In this dissertation, I will present my research efforts in enhancing the generalization ability of vision models from several perspectives. Firstly, i focus on designing vision algorithms that can generalize to novel viewpoints and occlusions, tackling challenges commonly encountered in real-world scenarios. Secondly, I investigate the architectural disparities between vision transformers and convolutional neural networks (CNNs) to better understand their impact on generalization, particularly concerning out-of-distribution data and adversarial attacks. Thirdly, I propose innovative self-supervised learning algorithms, Point-Level Region Contrast for Object Detection Pre-Training, aimed at providing strong backbone features for generalization across various downstream tasks. Additionally, I introduce efficient masked autoencoding techniques for enhanced representation learning. Finally, I propose Sequential Modeling Enables Scalable Learning for Large Vision Models, a methodology designed to effectively leverage large volumes of real-world visual data, thereby further enhancing generalization capabilities. Through these contributions, I aim to advance the field towards more generalized and robust vision models with broad applicability across diverse real-world scenarios
CHARACTERIZING CHROMOBACTERIUM SUBTSUGAE AS A MODEL ORGANISM TO STUDY IMMUNITY AND HOST-MICROBE INTERACTIONS IN DROSOPHILA MELANOGASTER
Get PDFThe innate immune response is an organism’s first line of defense against infection across animals, from invertebrates to vertebrates, including humans. It encompasses both defensive response and repair mechanisms, playing a vital role in host survival. Studying host-microbe interactions helps us understand the environmental and host factors that lead to disease and that maintain homeostasis. In this dissertation, I characterize the response of the model host Drosophila melanogaster to a novel pathogen, Chromobacterium subtsugae, a Gram-negative environmental bacterium known for its insecticidal activity and quorum-sensing production of the purple-pigmented metabolite violacein. Chapter 1 discusses the benefits of using D. melanogaster as a model organism for host-microbe interactions and the significance of understanding C. subtsugae virulence mechanisms. Chapter 2 establishes a C. subtsugae oral infection model to study the dynamics of gut disease that lead to the loss of whole-organism homeostasis, such as in mammalian sepsis. In Chapter 3, I expand this host-microbe model by examining D. melanogaster's response to C. subtsugae systemic infection and found flies exhibit a phenotypic switch in violacein production during infection, resulting in increased survival. Chapter 4 describes the transcriptomic response of D. melanogaster to C. subtsugae systemic infection and explores host inter-population variation. Chapter 5 describes the sequencing and assembly of C. subtsugae WT and ∆vioS using short and long reads to understand their genetic properties contributing to virulence. In conclusion, my dissertation characterizes a new infection model to examine host-microbe interactions and identify innovative mechanisms to limit the damage microbes cause during infection
CHAPERONE MEDIATED SOLID-TO-LIQUID PHASE TRANSITION IN THE DISSOLUTION OF MISFOLDED-PROTEIN AGGREGATES
Get PDFAccumulation of protein aggregates is a hallmark of cellular aging and degenerative disorders. This could result from either increased protein misfolding and aggregation or impaired dissolution of aggregates formed under stress, the latter of which is poorly understood. In this study, we employed quantitative live-cell imaging to investigate the dynamic process of protein disaggregation in yeast. We present evidence that heat-shock stress drives misfolded proteins into solid-like state of aggregation, which gradually transitions into a liquid-like state during dissolution as the stress attenuates. This solid-to-liquid phase transition (SLPT) accompanies the reduction in aggregate number due to fusion of the liquid condensates. The chaperone activity of Hsp104, a Clp/HSP100 family chaperone, is required for both SLPT and subsequent dispersal of the liquid-like condensates. Additionally, Sse1, a yeast HSP110 chaperone, also facilitates SLPT. These results illuminate an unexpected mechanistic framework of cellular control over protein disaggregation post stress attenuation
LIFE GOES ON A COLLECTION OF BIOLOGY-FOCUSED WRITING
Get PDFFrom plants and animals to the unexpected liveliness of urban environments, life springs up to fill every nook and cranny on this planet. In our complex world, the smallest changes can have large and unexpected impacts. This thesis is about biology and the many deeply entwined concepts that fall beneath its gaze
Application of trans-Tango for trans-synaptic tracing in the zebrafish visual system
No full textResolving neural connections in the brain is an important goal of neuroscience research. An extensive toolbox of methods is available for this purpose but can be hindered by challenges such as the physical accessibility of neurons, the non-specificity of reagents, and the toxicity of materials delivered to the animal. In my project, I helped adapt trans-Tango, a genetically encoded transsynaptic tracing system to a vertebrate nervous system. Previously, trans-Tango was used to visualize and manipulate a wide variety of neural circuits in the fly nervous system, including those involved in taste, visual system orientation, and addiction. The goals of my work were to validate this transsynaptic tracing system to the zebrafish Danio rerio, and specifically, to use it to visualize connections in the retina.
trans-Tango is built upon the human glucagon signalling pathway. Specificity is conferred by a promoter driving the Gal4/UAS transcriptional regulatory system, which activates expression of a human glucagon ligand that is presented extracellularly and anchored at the presynaptic membrane by domains of the Neurexin protein. The ligand spans the synaptic cleft by a portion of an intracellular adhesion molecule (ICAM) and binds to a modified G protein-coupled human glucagon receptor on postsynaptic membranes. At its cytoplasmic domain, the receptor is fused to the QF transcription factor through a proteolytic cleavage site (TEVcs). Upon activation by ligand binding, the receptor recruits an arrestin protein fused with a tobacco etch virus (TEV) protease, which cleaves its recognition sequence TEVcs. This frees QF to translocate to the nucleus, where it promotes transcription of genes under control of the QF upstream activating sequence (QUAS).
I contributed to the adaptation of trans-Tango to the zebrafish central nervous system and validated synaptically coupled neurons in the larval zebrafish retina, a tissue whose anatomy and connectivity is well understood. Functional circuits across the retina have been widely studied, and trans-Tango allows us to isolate neuronal partners that communicate by chemical synapses. My results support the known synaptic partners of inhibitory amacrine subtypes and support the formation of direct synaptic connections between dopaminergic amacrine cells and retinal ganglion cells in the developing zebrafish retina
HIV and Tuberculosis Stigma in KwaZulu-Natal, South Africa: A Mixed-Methods Analysis of Experience, Theory, and Method
No full textInfectious disease stigma such as tuberculosis (TB) stigma, and stigma associated with human immunodeficiency virus (HIV) negatively impact quality of life and impede care engagement at all levels of the care continuum. As a result, infectious disease stigma is a major impediment to HIV and TB elimination efforts. Stigma is a complex and dynamic social construct that varies across culture, geography and by local disease burden. Quantifying stigma is difficult with many scientists electing to use unvalidated or proxy variables. We have little knowledge about how hallmarks of clinical disease such as viral suppression and TB culture conversion impact stigma or illness identity. Research examining infectious disease stigma is generally done cross sectionally without consideration location in the TB/HIV continuum. While there are a variety of validated HIV stigma scales, there is only one scale validated to measure TB stigma, the Van Rie Patient and Community Perspectives Towards Tuberculosis Scales. While the scale was psychometrically validated in Thailand, the Patient Perspectives Towards Tuberculosis have not undergone qualitative exploration in South Africa. To address all these questions, we conducted a parallel mixed-methods study that examined and explored stigma, identity, fluctuations in stigma over time, and interrogated the underlying validity of the Patient Perspectives sub-scale.
The study sample (N=59) had a mean age of 36.6 years, 59.3% were male and most were unemployed (59.3%). At entry into TB treatment 51.0% were virally suppressed with a mean time of 6.4 years since HIV diagnosis. The qualitative aims (n= 30 participants selected from among the quantitative sample) used data from in-depth interviews exploring the construct of stigma. We also employed cognitive interviewing techniques to highlight discrepancies between the lived experience of stigma and traditional quantitative measures. Participants described the negligible impact that viral suppression or undetectable equals untransmittable (U=U) had on their identity or lived-experience of stigma. Over time, participants described changes in stigma from three perspectives which was an unexpected finding. Finally, we found that while it performed well psychometrically, the Patient Perspectives Towards Tuberculosis did not accurately quantify the lived experience of TB stigma and likely reflects community rather than individual stigma
Varieties of state-building: Ecology, Clientelism, and Bureaucratic Rule in Latin America
No full textThis dissertation explores the conditions under which ruling coalitions expand their authority over time and space, and such efforts’ effects on state capacity. Current research suggests that all states have a perpetual appetite to extract resources from society. However, this research overlooks the fact that subnational regions present different appeals and challenges to ruling coalitions. While states seek to extend bureaucratic rule over peripheries with valuable assets and favorable geography, they might instead seek to preserve local patrimonial bastions when those areas offer substantial electoral support. In turn, these strategies lead to broad subnational heterogeneity in the reach of the state and different patterns of state capacity.
I test this theory by studying Chile, a case that stands out as a successful instance of capacity-building in a region characterized by weak states. Notably, this occurred without geopolitical pressures, a typical catalyst identified in the literature. Prompted by a fiscal crisis in the mid-1850s, Chile's ruling coalition launched state-building projects to offset its budgetary deficit. The state negotiated a transition to bureaucratic rule in Atacama to the north and Concepción to the south, peripheral areas with favorable ecological conditions that coalesced to make a military threat. Then, the state imposed bureaucratization in Araucanía, an area of dense forests in the far south dominated by the native Mapuche. Finally, to shore up political support, the government allied with landowners in the capital's hinterland, the Central Valley.
The empirical chapters use original datasets with information from censuses, budgets, and statistical yearbooks that covers ninety years (1845-1935) to test these claims. Using geographic Information Systems (GIS), I map the cross-sectional and temporal variation in the territorial reach of the state. Using other archival and secondary sources, I trace the historical process by which elites engaged in these strategies and suggest how the effort to rule Araucanía in particular created tangible gains in state capacity. Although Chile is often considered the poster child of evenly projected of state power in Latin America, my argument suggests that state-building was instead highly selective and uneven, with the very center being the most important deviation to uniform, bureaucratic rule. These results challenge prevailing narratives about the projection of political authority and Chile’s territorial uniformity
Sulfur cycling in estuarine sediments: Implications for paleoenvironmental signatures
Get PDFRedox transformations of sulfur are integral to the redox evolution of Earth. The reduction of sulfate to sulfide and ensuing burial of pyrite (FeS2) in sediments increases the Earth’s ocean-atmosphere oxidation state, and the distribution of stable sulfur isotopes in pyrite provides a deep-time proxy record for redox changes and biogeochemical processes. However, interpretations of the geological sulfur record must be informed by study of sulfur cycling in modern environments. To this end, my dissertation focuses on sedimentary sulfur cycling in Chesapeake Bay, a large estuary with strong spatial and temporal biogeochemical gradients, which may offer a modern analog for ancient sediments in a low-oxygen world.
In the first study, I detail seasonal changes in the concentration and sulfur isotope composition of pyrite and other reduced sulfur compounds in the shallowest sediments of two sites in Chesapeake Bay. The findings indicate that bioturbation, the mixing and flushing of shallow sediments by animals, affects the ecology of sulfur-cycling microbes and may promote pyrite precipitation via mixed-valence sulfur compounds called polysulfides.
In the second study, I assess controls on pyrite accumulation in deeper sediment cores collected throughout Chesapeake Bay. The results suggest that mild bioturbation may increase pyrite burial rates by increasing net sulfate fluxes into sediments and by generating polysulfides, which participate directly in pyrite precipitation. I apply these findings to the early Paleozoic Era, a time of rapid biogeochemical changes on Earth, and find that the efficiency of Paleozoic pyrite burial temporarily increased in tandem with rising rates of bioturbation. This effect would have transiently accelerated ocean-atmosphere oxygenation.
In the third study, I use quantum-chemical modeling to assess the degree of equilibrium position-specific isotope fractionation (PSIF) in polysulfides. Results suggest that PSIF adds 0.1–2.5‰ to the bulk intermolecular isotopic fractionation of ~5.5‰ between elemental sulfur and the most common polysulfide compound. This may explain why shallow sedimentary pyrites are often anomalously depleted in sulfur-34 compared to pyrite-forming compounds from the same sediments, and it highlights the need to consider how rapid mass-dependent fractionation in the polysulfide pool may have affected isotopic signals in the ancient rock record
Subsurface Flow and Transport Processes with Applications to Methane Variations on Mars
No full textGround-based measurements of methane, a potential biosignature, in the atmosphere of Mars indicate that its abundance fluctuates over seasonal and sub-diurnal time scales. To date, both the source of methane on Mars and the mechanism for transmission from the subsurface to the atmosphere are not well understood, and the relative paucity of reliable measurements make it challenging to constrain the time scales of the observed methane variations. The majority of research in this area employs sophisticated global atmospheric transport and mixing models to interpret methane signatures, whereas the subsurface processes have generally been interrogated using relatively simplistic models that neglect potentially critical aspects of subsurface transport. Here I address the subsurface transport of methane on Mars using models of fractured-rock porous media gas flow and tracer transport in order to better understand the processes responsible for delivering methane from underground to the atmosphere. First, I present a gas flow and transport model to simulate methane transport from depths of 200 m to the surface of Mars driven by the barometric-pressure pumping mechanism. I determined that Mars’ atmospheric pressure fluctuations are sufficiently strong to drive significant fluxes of methane into the atmosphere from deep underground, and that the seasonality of the surface fluxes is reasonably consistent with the observed abundance variations. Next, I investigate sub-diurnal methane abundance variations to identify strategic atmospheric sampling times for the MSL Curiosity rover using a coupled subsurface-atmosphere model that predicts hourly methane abundances in northern summer. The analysis identifies two time windows not previously sampled that have high potential to constrain the apparent mid-day drop in methane abundance and modest potential to support the influence of a barometric pumping mechanism on Mars’ methane variations. Lastly, I perform a fundamental study of gas transport processes that combines laboratory experimentation and two-site adsorptive-diffusive modeling to investigate gas transport in zeolitic rock at varying degrees of saturation, which has implications for underground nuclear explosion monitoring as well as episodic gas release from hypothesized zeolitic martian methane reservoirs. I show that at lower saturations, adding water content preferentially fills zeolite pores which greatly reduces their adsorptive capacity, thereby increasing the rate of transport