33760 research outputs found
Sort by
LEARNING STRUCTURED ALIGNMENT: FROM VISUAL UNDERSTANDING TO ROBOT CONTROL
No full textThe rapid growth of visual data and the increasing demand for intelligent robotic systemshave created a pressing need for methods that can establish meaningful correspondences and
relationships across diverse visual modalities and robotic tasks. This dissertation addresses the
fundamental challenge of learning structured alignment, which involves establishing correspondences
between different representations, temporal sequences, and task domains to enable more
effective visual understanding and robot control.
In the first part of this thesis, we advance visual understanding through three key contributionsthat demonstrate the power of structured alignment in perception tasks. We begin by
tackling semantic correspondence, where we propose a teacher-student learning paradigm that
enriches supervision from sparse keypoint annotations, enabling dense correspondence learning
through spatial priors and loss-driven dynamic label selection. We then address video instance
segmentation through two complementary approaches: UVIS, which leverages foundation
models (DINO and CLIP) for unsupervised segmentation without dense annotations, and
PointVIS, which achieves competitive performance using only point-level supervision through
class-agnostic proposal generation and spatio-temporal matching. Finally, we develop Trokens
for few-shot action recognition, introducing semantic-aware point correspondence sampling and
relational motion alignment that captures both intra-trajectory dynamics through Histogram of
Oriented Displacements and inter-trajectory spatial relationships, effectively aligning appearance
features with motion patterns through trajectory-based token alignment.
While the first part focuses on establishing correspondences within visual data, real-worldapplications require bridging the gap between visual understanding and robot control. In the second
part of this thesis, we present two frameworks that demonstrate how structured alignment
can be extended to robotic applications. We introduce ARDuP, a novel method for video-based
policy learning that aligns generated visual plans with language instructions for effective control.
This innovative framework integrates active region (i.e. potential interaction areas) conditioning
with latent diffusion models for video planning and employs latent representations for direct
action decoding during inverse dynamic modeling. By utilizing motion cues in videos for automatic
active region discovery, our method eliminates the need for manual annotations of active
regions. Finally, we present TREND, which addresses robust preference-based reinforcement
learning through a tri-teaching framework that filters noisy preference labels while incorporating
few-shot expert demonstrations, demonstrating effective alignment between human preferences
and robot behaviors even under high noise conditions
AN INVESTIGATION OF THE PRINCIPAL SUPERVISOR’S ROLE IN SUPPORTING PRINCIPAL’S INSTRUCTIONAL LEADERSHIP SKILLS
No full textThis study investigated the role of principal supervisors in educational leadership within the K–12 Grand Royal School System (GRSS). It examined what principal supervisors do to build principals’ instructional leadership capacity—and to what extent their actions support principals in leading teaching and learning. Using a mixed-methods approach, including surveys, interviews, observations, and document analysis, the study explored the daily practices of principal supervisors and their interactions with principals.Findings show that GRSS principal supervisors value instructional leadership and often use goal-setting and formal observations to build principals’ capacity. However, their daily work is frequently consumed by operational demands, community concerns, and compliance tasks, which limit their ability to focus on instruction. Although most participants agreed that principal supervisors were accomplished instructional leaders, neither principals nor supervisors cited specific examples of support that led to instructional improvements. This gap was attributed to systemic barriers such as competing district priorities, ambiguous role expectations, and limited time.
In addition to these core findings, interviews revealed several unexpected insights, including supervisors’ lack of control over their schedules, inconsistent implementation of site visits, and a shared sense of nostalgia for earlier support structures. Supervisors also described the emotional toll of the role, marked by stress and strain from balancing competing demands without sufficient authority or systemic support. These perspectives provided valuable context for interpreting survey results and underscored the disconnect between the role’s intended purpose and its day-to-day realities.
While participants believed in the potential of the supervisor role to enhance leadership and improve learning outcomes, they acknowledged that current conditions often prevent it from functioning as intended. The study concludes with recommendations for restructuring the principal supervisor role and strengthening support for central office leaders who serve in this capacity
Engineering Biologically Inspired Tissue Hybrids Towards Treating Patient And Defect-Site-Specific Bone Defects
No full textMorphological Characterization of Quasi-Circular Features Across the Solar System
No full textThe interior dynamics and geologic activity of a planetary body are linked. Due to instrumentation limitations, probing or studying the interior of planetary bodies directly is difficult. I argue that by studying surface features whose formation mechanism is linked to the material exchange between the interior and crust, I can gain information on the evolution and dynamics of the planetary body. For this dissertation, I have chosen to study microfeatures on Europa, coronae on Venus, and calderas on Earth.New methods were developed to determine the orientation and circularity of surface features across three planetary bodies. Three circular statistical tests were conducted to test the distribution of the orientation data for the quasi-circular features. A new definition of misorientation based on spherical geometry was defined for the Venus and Earth studies.
On Europa, the orientation of microfeatures is not uniformly distributed and does not depend on latitude. Microfeatures are found to have formed episodically and under various stresses. Chaos and hybrids have a more extended formation period than pits and domes. These findings support formation mechanisms dependent on a diapir or liquid water body interacting with crustal horizontal stresses.
A new digital map of 525 coronae was created for Venus. Orientation results show that coronae have a non-uniform orientation distribution. Their orientation is not dependent on size, formation location, or circularity. Comparing the orientation of coronae to implied stress orientations shows that coronae are likely to be elongated in the direction of maximum horizontal compressional stress. These results align with coronae formation models that show a preferential injection of melt in the direction of maximum horizontal compressional stress or a rising plume that pools along a step in lithospheric thickness.
I compared the orientation of calderas on Earth to direct stress measurements. These findings show that calderas are more likely to be elongated perpendicular to the direction of maximum horizontal compressional stress. These results suggest that local stresses and preexisting structures influence calderas’ orientation. I advocate for caldera formation models based on the deformation of magma chambers or the preferential injection of magma along preexisting fault structures
Characterizing the biophysical properties of mouse lymph nodes using multiple particle tracking
No full textLymph nodes (LNs) are scattered throughout the body and essential for mounting an adaptive immune response. They are highly spatially organized, which helps to ensure that these immune responses are efficiently mounted. LN spatial organization is supported by an extracellular matrix (ECM), which is thought to be more densely packed in the inner paracortex and less densely packed in the outer cortex. ECM provides tissue with viscoelasticity, which is a property of a material that exhibits both viscous and elastic responses. Elasticity is the ability of a material to be pulled or compressed without permanent deformation, whereas viscosity is the liquid-like response of the material. Recent studies have shown that LN viscoelasticity affects cell functions, including cell proliferation, migration speed, and morphology. However, the viscoelasticity of LN has not yet been characterized.
Traditional methods for studying material rheology are inadequate for studying the intricate architecture of LNs. In this study, live LN tissue slices were used to maintain LN architecture, and viscoelasticity was measured using an established method called multiple particle tracking (MPT). In MPT, fluorescent probes are placed on tissue slices, and the Brownian motion of the particles is recorded. The displacements of probe trajectories can be applied to the Stokes-Einstein relationship to calculate the values of moduli, pore size, and viscosity.
In this thesis, we developed an inert probe for use with MPT that does not interact with other probes or the environment. We then characterized skin draining and mesenteric LN viscoelasticity at homeostasis. We used an acute mouse model in which lipopolysaccharide (LPS) was intradermally injected to inflame inguinal LNs, euthanized at specific time points, and viscoelastic properties were characterized using MPT. To observe any changes in ECM components during inflammation, we used immunohistochemistry and image analysis to quantify changes in ECM components such as collagen, elastin, and fibronectin. Additionally, we characterized viscoelasticity in a chronic inflammation mouse model (weekly LPS injections for 5 weeks) and euthanized the mice at peak inflammation and resolution. We found that LN viscoelasticity is site-specific, where skin-draining LNs exhibit stiffer properties than mesenteric lymph nodes. We also found that during acute inflammation, LNs exhibited tissue stiffening at peak inflammation and were permanently restructured after resolution. Lastly, we found that chronic inflammation did not significantly change LN viscoelasticity. Our findings show the characteristics required to use MPT in LN tissues to characterize the viscoelasticity of LNs while maintaining their structure. We also successfully measured the viscoelastic properties of LNs during acute and chronic inflammation and began probing how the ECM components change in relation to viscoelasticity
Combinatorial Optimization for Barter Exchange and Tumor Phylogeny Inference
No full textCombinatorial Optimization is a field with rich history encompassing problems that are fundamental to Computer Science with important applications. Despite the many decades-long maturity of Combinatorial Optimization, new important problems continue to arise in response to emerging technologies and datasets. Additionally, even the slightest variations in problem statements can render previous approaches unworkable. In this doctoral thesis we will focus on two such problems. First, we explore a centralized barter exchange problem, motivated by the rise of online marketplaces and social media. Second, we explore the problem of tumor phylogeny inference, motivated by recent advances in single-cell sequencing technologies and the continued demands of ever-growing datasets.
First, we study a centralized barter exchange model where agents enter with a set of items they have available to trade, and a set of items they want. Additionally, every item has some fixed positive value. For example, in an exchange between trading card collectors, agents looking to broaden their collections can enter aiming to exchange cards they have repeats of for cards that they do not yet have. The clearinghouse, i.e., the party facilitating the centralized barter exchange, is tasked with the goal of outputting a reallocation that maximizes the welfare of agents subject to no agent being unhappy (an unhappy agent gives away more value than they receive). We study two subproblems: Barter Exchange with Symmetric Valuations (BSV) and Barter Exchange with Asymmetric Valuations (BAV). BSV and BAV are, respectively, the settings when agents agree on item values and when agents may have their own individual item valuations. Both BSV and BAV are NP-hard. On the flip side, we design randomized algorithms for both of these problems that guarantee, in expectation, maximum welfare and that every agent receives at least as much value as they give away. Additionally, for BAV, with probability 1, every agent receives at least as much value as they give away, minus the value of two items. In the case of BSV, the guarantee is strengthened to, with probability 1, each agent receiving at least as much value as they give away minus the value of a single item. Finally, in BAV, agents can improve their utility by misreporting their valuations. We design a mechanism that is incentive compatible (i.e., an agent maximizes their utility by reporting their true valuation).
Second, we study tumor phylogeny reconstruction from single-cell sequencing (SCS) data. Recent advances in SCS allow for more a fine-grained look into tumor evolutionary history. Accurate understanding of such tumor phylogenies (evolutionary trees) can inform effective cancer treatments. However, SCS data is invariably noisy and this noise must be dealt with to produce a \textit{perfect phylogeny} (PP), which can then easily be mapped to an evolutionary tree using the well-known result of Gusfield, 1991. In particular, we want to find the most likely PP (i.e., the PP requiring the least number of corrections in the data). This is the Most Likely PP Reconstruction problem and it will be our focus. We draw connections between the Most Likely PP Reconstruction problem and the fundamental combinatorial optimization problem of Vertex Cover to build a branch-and-bound algorithm that provably returns the most likely PP. We benchmark our branch-and-bound algorithm against previous approaches on synthetically generated SCS data. In these experiments, our algorithm finds the most likely PP up to 300 times faster.
A common thread in our studied problems of Barter Exchange and Most Likely PP Reconstruction, is the Combinatorial Optimization paradigm of formulating problems as Linear Programs (LPs) to obtain a fractional solution in polynomial time.Then this fractional solution is then carefully rounded to an integral solution, in polynomial-time, while providing theoretical guarantees over the rounded solution.
Among these results, we introduce a novel refinement of the randomized dependent rounding algorithm of Gandhi et al., 2006
Sometimes The Rain, Sometimes The Sea: A Lighting Design
No full textThis thesis provides an analysis of Sometimes The Rain, Sometimes The Sea by Julia Izumi and its source material, an account of the process of designing the lighting for this production at American University in the fall of 2024, and documentation of the completed lighting design, along with the designer’s reflection on the process
OBSERVATION OF ULTRA-HIGH-ENERGY GAMMA RAYS AND SEARCH FOR DARK MATTER SIGNATURES IN THE GALACTIC CENTER WITH THE HAWC OBSERVATORY
No full textThe center of our Galaxy is an intriguing region in astrophysics, providing a unique opportunity to study various astrophysical processes. However, our line of sight is obscured by dense layers of dust and gas, making optical emission observation impossible. Fortunately, we can observe other portions of the electromagnetic spectrum, such as radio, X-rays, and gamma rays. The latter serve as probes of cosmic-ray acceleration to extremely high energies and, in theory, for indirect \ac{DM} detection. This dissertation analyzes gamma-ray data from the \ac{GC}, obtained with the \ac{HAWC} Gamma-Ray Observatory.
The \ac{HAWC} Observatory, situated on the Sierra Negra volcano in Mexico at an altitude of 4,100 m, detects gamma rays with energies from 0.1 to greater than 100 TeV. It has a wide field of view of about 2 sr, and with an operational duty cycle of over 95\%, it observes two-thirds of the sky daily. The observatory uses the water-Cherenkov detection technique to detect Cherenkov emission from secondary air-shower particles. This study presents improvements to the reconstruction algorithms that retrieve the properties of the primary gamma rays. In particular, the angular resolution and background rejection have improved by a factor of four for gamma-ray events coming from high-zenith angles and with energies above 40 TeV. Thanks to these improvements, \ac{HAWC} can detect the \ac{GC}.
Although multiple sources of cosmic-ray acceleration to PeV energies, known as PeVatrons, have been proposed within the Galaxy, they remain insufficient to fully account for the observed flux. In this context, the \ac{GC} was proposed as one of the main Galactic PeVatrons with a power-law spectrum that extends up to 50 TeV without a cutoff. Here, we present, for the first time, observations of gamma rays with energies above 100 TeV, which we suggest originated from PeV protons---accelerated in the \ac{GC}---that interact with the dense ambient gas. While the angular resolution of \ac{HAWC} at this high zenith angle is not enough to distinguish the specific PeVatron source, we provide the first confirmation of its existence.
The \ac{GC} is also one of the most promising candidates for indirectly detecting \ac{DM} via gamma rays. While the nature of \ac{DM} remains a fundamental question in modern physics, \acp{WIMP} are considered potential candidates for \ac{DM} and naturally emerge in several extensions of the \ac{SM} of particle physics. The relic density of thermally produced \acp{WIMP} in the early Universe can account for all the \ac{DM} observed in the Universe, as measured from cosmological observations. In theory, \acp{WIMP} self-annihilate in dense astrophysical environments, like the \ac{GC}, producing gamma rays in the final state from processes such as hadronization, radiation, and decay of \ac{SM} particles. We conduct a follow-up study of the above analysis and compare the spatial and spectral morphology of the residual gamma-ray emission to the one theorized for \ac{DM} annihilation. Finding no significant emission, we place for the first time to date \acp{UL} at the 95\% \ac{CL} on the velocity-weighted cross section, for \ac{DM} particles with masses well above 70 TeV using \ac{GC} gamma-ray data
Improving quantitative ventifact analysis for climate investigations using the Dyngjusandur sandsheet in Iceland as a planetary analogue
No full textWe use the Dyngjusandur sandsheet, Iceland, as a testbed to assess the type and number of measurements required to accurately represent ventifact orientations and extract palaeowind information through a statistical evaluation of the differences between photogrammetric based and in situ measurements of ventifact feature orientations. Forty representative ventifacts were selected for in situ measurement, 20 of which were imaged for photogrammetric analysis to produce oriented and scaled virtual 3D models. An additional set of measurements were made on the ‘synthetic’ models to allow for statistical assessment of erosional feature orientation. Despite the similarity between the photogrammetric (1145 measurements) and the in situ datasets (500 measurements), there are small but significant differences in mean feature orientation that become more pronounced as sample size is reduced. Results indicate that in situ and photogrammetric methods of measuring feature orientation are comparable (to within 2° ± 2° at 1 σ ( n = 20) for inferring palaeowind directions), but photogrammetric analyses require less time for data acquisition (by a factor of 0.36–0.66) and yield over five times (5.16 times) as many measurements per ventifact, with additional advantages of being able to examine digital objects under different illumination conditions and magnifications, and from angles that are otherwise not possible in the field.https://doi.org/10.1144/jgs2024-17
Investigating the Role of ppsA in ATP Production and Phage Replication in E. coli
No full textAntibiotic resistance is one of the most pressing health crises of our time, threatening our ability to treat even common bacterial infections. In response to increasing antibiotic-resistant bacterial infections, phage therapy has become a promising alternative to traditional antibiotics. Bacteriophages infect bacterial cells in order to exploit the host’s metabolic products for their own growth and replication, ultimately killing the bacterial cell. That is, phages infect only bacterial cells, therefore treating infection without interfering with human cells.
Our project aimed to investigate cell growth and bacteriophage replication in E. coli when the ppsA gene is removed to provide further insight into phage therapy. The gene ppsA is located in the gluconeogenesis I pathway, and encodes the enzyme phosphoenolpyruvate synthetase (ppsA). PpsA catalyzes the reaction that produces phosphoenolpyruvate (PEP), a precursor for glycolysis involved in the production of ATP. Comparative growth assays, lysis curves, comparative plaque assays, and an ATP quantification assay were performed to quantify the growth, viral replication, and ATP levels of parent and ΔppsA strains.
ΔppsA exhibited a slower growth rate than wild type in LB and M9 minimal media, replicating previous findings that deleting ppsA hinders E. coli growth. ΔppsA also experienced reduced T2 and T4r bacteriophage replication, a novel finding. Results of an ATP quantification assay were produced but inconclusive as to if ATP availability was directly impacted by ΔppsA. Overall, a direct link between ppsA and effective bacteriophage replication was confirmed. This link between ppsA and bacteriophage growth indicates that future phage therapy strategies may benefit from targeting bacterial metabolic pathways to improve treatments. Future work will be needed to determine the mechanisms behind this relationship, especially the role of ATP availability in ΔppsA