Treasures @ UT Dallas
Not a member yet
7697 research outputs found
Sort by
Gaussian Process Emulation: Theory and Application to Coupled Physics
No full textThis dissertation focuses on uncertainty quantification (UQ) for complex multiphysics mod-
els. One of the defining features of these multiphysics models is large output dimensions
because the quantities of interest typically depend on both space and time. We develop
Gaussian process (GP) emulators as fast surrogates of computationally expensive coupled
computer models for uncertainty estimation. A GP can be thought of as an interpolator for
a limited number of computer runs of a physics simulator. The predictive mean of the GP
is conditioned to agree with the output of the multiphysics model. A GP is computation-
ally inexpensive, enabling numerous rapid evaluations of the emulator across various input
parameter regimes. Furthermore, Gaussian process emulators offer a mechanism to quantify
uncertainty at untested inputs (where the computer model is not evaluated) because GP’s
provide credible intervals estimates. However, standard GPs are not designed to handle
coupling or high-dimensional outputs.
We develop a Gaussian process methodology suitable for coupled, vector-valued systems. The
approach combines a vector-valued emulator (parallel partial emulator) with an emulator for
coupled multiphysics functions (a linked emulator). The main assumption of the parallel
partial emulator is that the predictive mean and variance of each component of the vector-
valued function are independent, while the correlation parameters are shared. Meanwhile,
the linked emulator mimics the coupling structure of the emulated function. We test this
new method on a simple composition of trigonometric functions as well as on the Terzaghi
consolidation problem. The classical Terzaghi problem models fluid flow and compaction of
a one-dimensional column of mud when a large load is dropped on the top. We choose the
following metrics to measure emulation performance: training and prediction time, root mean
squared error and the average length of the credible intervals. For both test problems our
parallel partial linked emulator outperforms both traditional composite emulation techniques
and techniques that involve a separate emulator at each value of the independent variable
(here depth).
Simulation studies of hydraulic fracturing can identify and isolate parameter regimes that
create the most volume in the fracture, allowing oil and gas to flow more freely to the wells.
In the field, volume creation can be estimated by monitoring microseismic activity (namely
locations and magnitudes of these events). We use a hydraulic fracturing simulator, the
Complex ReseArch Fracture Code (CFRAC) for numerically realistic experiments. However,
the CFRAC simulator is computationally prohibitive for a large range of input parameter
values. To avoid running the physics-based computer model for all of these input values, we
employ a GP in place of CFRAC to search the input space for combinations of parameters
that lead to simultaneous changes in void aperture and sliding displacement which indicate
successful volume creation. Through emulator evaluations, we show that seismic data is not
always reliable for selecting parameter regimes that ensure volume creation. In some cases,
input parameter values identified by observing larger magnitudes of cumulative moment do
not lead to the opening of the fracture, and thus do not result in volume creation
Three Papers on International Political Interactions and Event Data
No full textThis dissertation is of the three-paper format. What Factors Influence Foreign Public Opin-
ion? finds that several factors influence foreign public opinion. An analysis is performed
using regressions that rely on data for 34 states from 2013 through 2019. The results sug-
gest the following: 1.) An increase in a state’s public diplomacy spending, which aims to
positively influence a foreign public, leads to a decrease in the favorability of the foreign
public toward the spending state. 2.) Armed conflict during the present or past century
between a foreign public’s state and another state leads to an increase in the favorability
of the foreign public toward the other state. 3.) An increase in the amount of foreign aid
money dispersed by a donor state to a recipient state leads to an increase in the favorability
of the recipient state’s public toward the donor. 4.) An increase in the confidence that a
foreign public has in the leader of another state leads to an increase in the favorability of
the foreign public toward the other state. An Analysis of U.S. Embassy Tweet Sentiment
using GARCH Models investigates the following research question: What drives the volatil-
ity in the sentiment of diplomatic tweets? For each of four cases, a DCC-GARCH (dynamic
conditional correlation - generalized autoregressive conditional heteroskedasticity) model is
applied to a bivariate time-series constructed with a tweet sentiment volatility time-series
comprised of data sourced from official U.S. embassy Twitter accounts, and a theoretically-
motivated, covariate political event data time-series. It is found that, in only two of the
cases, political events are potentially driving the volatility in tweet sentiment, as indicated
by statistically significant results for the Joint Dynamic Conditional Correlation - B1 pa-
rameter. Measurement in Event Data Research examines measurement concerns about the
research and application of political event data, with a particular focus on the state-of-the-
art after 2015. Here, the “measurement of event data” is determining the mathematical
and/or conceptual distance between what a machine codes/classifies from information de-
scribing an event and what actually is happening in the event. Three primary arguments
are made: 1.) It is important for users of event data to understand the measurement side
of these data to avoid faulty inferences and make better decisions. 2.) Avant-garde event
data systems are still not free from some of the fundamental problems that plague legacy
systems (investigated are theoretical and real-world examples of measurement issues, why
they are problematic, how they are dealt with, and what is left to be desired even with newer
systems). 3.) One of the most crucial goals of event data science is to attain congruence
between what is machine-coded/classified versus the actual circumstances of the event. To
support the arguments, well-documented sources of measurement error are detailed and new
ones are identified. This information is used to gauge the advancement of the literature.
Guidance is then provided for users on how to make performance comparisons within and
across language models, identify opportunities to improve event data systems, and more
articulately discuss and present their findings
Towards Improving End-to-end Network Slices Control Through Reinforcement Learning
No full textIn the era of 5G networks, customer expectations in the realm of networking have surged
significantly. They now demand highly personalized and tailor-made services that precisely
align with their specific criteria for Quality of Service (QoS), Service Level Agreements
(SLAs), and Key Performance Indicators (KPIs). Traditional 4G networks, which adhere to
a “one-size-fits-all” model, are ill-suited to cope with these evolving demands. In response
to this challenge, the concept of network slicing has emerged as a highly promising solution.
Network slicing allows the partitioning of a single physical network into distinct isolated
logical slices, each proficiently catering to the diverse needs of individual users.
However, the dynamic nature of network traffic and computer networks introduces an additional layer of complexity to the implementation of network slicing. In this dissertation,
we aim to enhance the strategy for network slice admission and allocation. Our primary
approach is to evolve from the traditional static network slice model to an elastic network
slice. This transition is designed to amplify system resource utilization efficiency and reduce
costs for customers. As evidence of the utility of the elastic network slice, we present a
use-case within a federated learning context. This illustrative case underscores the benefits
and practical applicability of our proposed model. Furthermore, we delve into the intricacies of network slice provisioning within a multi-domain environment. This exploration
encompasses both the strategies for partitioning and allocation
Stability of Solutions of the Complex Ginzburg-landau Equation
No full textShort-pulse lasers are modeled using equations related to the cubic-quintic complex Ginzburg-
Landau equation (CQ-CGLE), which is a generalization of the nonlinear Schr¨odinger equation (NLSE). These lasers balance loss, gain, nonlinearity, spectral filtering, and dispersion
in order to generate regular trains of stationary or periodically stationary pulses. While stationary pulses maintain a constant shape as they propagate, periodically stationary pulses
change shape as they propagate around the laser loop, returning to the same shape once each
round trip. With the advancement of laser technology, there has been a dramatic increase
in the amount by which the pulses breathe each round trip of the laser. In this dissertation,
we describe a method for determining the regions of parameter-space in which stable stationary pulses can be generated. Unlike other existing methods for finding stable stationary
solutions, we anticipate that it will be possible to extend our proposed method to the case
of periodically stationary pulses.
We describe the linearization of the the CQ-CGLE about a stationary solution, and we
employ a method that uses the spectrum of the linearized operator to determine the stability
of the pulse. The spectrum of the pulse is composed of the essential and point spectra, and
while formulae exist for the essential spectrum, in general no such formulae exist for the
point spectrum. We determine the point spectrum with the aid of a compact operator with
a matrix-valued Green’s kernel that is associated to the linearized operator.
We consider two types of compact operators, those which are trace class, and those which
satisfy the weaker condition of being Hilbert-Schmidt. We review the theory of the Fredholm
determinant of a trace class operator and the 2−modified Fredholm determinant of a Hilbert-
Schmidt operator, and we extend this theory to the case of matrix-valued kernels. We derive
a formula for the numerical approximation of such Fredholm determinants and quantify the
error between the true and approximated determinants. We then establish a result which
quantifies when a Hilbert-Schmidt operator is trace class. We prove that if the matrix-valued
Green’s kernel associated to the linearization of the CQ-CGLE defines a trace class operator,
then the Fredholm determinant of this operator is equal to the well-known Evans function.
Finally, we implement our numerical method in the special case of a known solution of
the NLSE, and we show that in this case the kernel is trace class. We derive an explicit
formula for the Evans function in this case and obtain excellent agreement between it and the
numerically calculated Fredholm determinant. We quantify the behavior of the 2−modified
Fredholm determinant and present results showing the accuracy of our numerical method,
thereby validating the error bounds we derived
Confidential Computing with Trusted Execution Environments
No full textIn an era where digital interconnectedness is a fundamental aspect of daily life, the safeguarding
of user data against unauthorized access, misuse, and exploitation emerges as a critical
concern. The reliance on digital platforms and services calls for robust protection measures
for personal information, where financial losses, psychological distress, and erosion of trust
between consumers and digital service providers are just some of the potential consequences
of inadequate data protection. Companies that fail to prioritize strong security frameworks
are at risk of incurring financial penalties, facing legal challenges, and suffering reputational
damage. Eventually, user trust in the digital age hinges on transparent and responsible data
management.
The rapid expansion of the Internet of Things (IoT) introduces new vulnerabilities in data
security and privacy, as billions of connected devices collect and transmit sensitive data. This
makes IoT systems prime targets for cyber-attacks, posing risks ranging from identity theft
to physical threats. The solution extends beyond encryption and secure protocols, requiring
a holistic approach that includes both hardware and software security measures. Confidential
computing, which leverages hardware-based security technologies, offers a promising avenue
for creating secure environments for processing sensitive data within the ever-expanding IoT
landscape.
We propose an advanced approach to IoT data security and privacy through Confidential
Computing, emphasizing the potency and limitations of current Trusted Execution Environments (TEEs) and encryption methods. Our design involves using state-of-the-art TEE Intel
SGX in the untrusted cloud to safeguard IoT data and popular trigger-action automation
platform, offering an end-to-end privacy preserving solution. Additionally, by demonstrating
an attack that can infer sensitive information of IoT devices from the cloud-based TEE with
high accuracy, we analyze the shortcomings of TEEs with access pattern-based side-channel
attacks and advocate for a comprehensive security strategy incorporating data oblivious
execution and padding to mitigate such threats.
Further, our research broadens to include the critical area of confidential AI at the edge,
where we introduce an innovative system designed for deep learning inference on edge devices
equipped with limited trusted memory. This system is crafted to safeguard the confidentiality
and integrity of deep learning models and data on untrusted edge devices, challenging the
constraints of existing edge-based TEEs without compromising the accuracy or performance
of the models. Through extensive experimentation across various deep learning architectures,
our findings confirm the system’s effectiveness in deploying inference services for complex
models on edge devices, ensuring data confidentiality and maintaining model accuracy with
minimal performance overhead
Utilization of TMV-TEMPO as an in Vivo MRI Sensor of ROS Production in Liver Inflammation and as a DNP Agent
No full textMagnetic resonance imaging (MRI) is a powerful tool that can noninvasively generate exquisite
images of soft tissues in living subjects for medical diagnostics. This PhD dissertation details the
i) utility of a nitroxyl-modified tobacco mosaic virus (TMV) as an in vivo MRI contrast agent for
detection of reactive oxygen species (ROS) in murine model of liver inflammation and ii) its
feasibility as an agent in MRI signal-enhancing dynamic nuclear polarization (DNP) technology.
Chapter 1 entails discussion of the basics of magnetic resonance and hyperpolarization. Chapter 2
involves discussion of the identification and physical characterization of spontaneous mutation of
TMV in a laboratory environment. Chapter 3 shows the successful utility of reduced TMV-
TEMPO as an in vivo T2 -weighted MRI sensor of ROS in lipopolysaccharide (LPS)-induced liver
inflammation in Balb/c mice. Chapter 4 details the 13 C DNP testing of TMV-TEMPO as a
polarizing agent and the use of gadolinium coordination polymer GduDEP as a 13 C DNP signal
enhancer. Finally, the overall conclusion and outlook of these projects are detailed in Chapter 5.
Overall, this PhD dissertation ties up the experimental details and discussions of a new in vivo
MRI biosensor for ROS at least in the pre-clinical level, along with some details on improving the
13
C DNP signals
Effects of Selective Inner Hair Cell Loss and Cochlear Synaptopathy on Psychophysical Intensity Increment Detection Tasks in the Chinchilla Animal Model
No full textIntensity coding of sound plays a critical role in the perception of complex auditory stimuli such
as speech. These stimuli are considered complex, in part, due to rapid fluctuations along the
intensity and temporal domains. These essential temporal and intensity processing abilities are
thought to be negatively impacted by age-related hearing loss (Vermiglio et al., 2012) or noise-
induced hearing loss (Altschuler et al., 2019) and deficits in these two domains are thought to
contribute to speech processing deficits (Kumar et al., 2012). In the periphery, prior to cortical
levels of auditory processing, hearing loss is highly correlated with damage to the outer hair cells
(OHC) and/or inner hair cells (IHC), the two sensory cell types that reside in the inner ear. At
this level, OHC play a critical role in active basilar membrane non-linear gain in the cochlea.
However, OHC only transmit about 5% of afferent information to auditory nerve fibers
(Spoendlin, 1971). Loss of OHC has been directly correlated with increased thresholds and
poorer frequency tuning due to loss of non-linear amplification for low intensity acoustic stimuli.
For example, it has been shown that damage to OHC from noise exposure results in reduced
hearing sensitivity evidenced by commensurate elevations in auditory thresholds (Boettcher et
al., 1992a). IHC on the other hand have extensive auditory nerve fiber innervation and transmit
over 95% of afferent acoustic input to the central auditory system. Despite the extensive
innervation, selective loss of IHC in animal models has been shown to have little effect on
physiological or psychophysical auditory thresholds but has been shown to reduce the neural
signal coming from the cochlea.
Despite these findings, few studies have evaluated the effects of selective IHC loss or damage on
suprathreshold auditory perception. Over the last decade, animal studies have suggested that
noise and age-induced IHC deafferentation in mice could play a significant role in auditory
temporal processing (Liberman and Kujawa, 2017) and may also impact intensity coding.
Other previous studies have shown that selective IHC loss in chinchillas following carboplatin
treatment, a commonly used anticancer drug, has little effect on pure tone thresholds in quiet
(Lobarinas et al., 2013b; 2016; Salvi et al., 2016), but significantly affects pure-tone thresholds
in noise (Lobarinas et al., 2016) and detection of gaps in continuous noise (Lobarinas et al.,
2020). In contrast, selective IHC loss has little effect on temporal summation assessed by
psychophysical detection of short-duration pure-tones (Trevino et al., 2022). The results of these
two studies suggest that the gap detection deficits may be the result of poorer intensity coding
given that temporal summation does not change even after IHC losses of over 70%.
To assess whether the deficits observed in the gap detection studies stem from poorer intensity
coding, I developed software based on previous studies to measure intensity increment detection
(IID). This psychophysical task was used to evaluate chinchilla sensitivity to intensity changes
before and after carboplatin, a treatment that has been shown to reliably and selectively destroy
IHCs in this species.
Alternatively, selective damage but not loss of IHCs can occur as a result of exposure to noise at
levels that are below those that would cause mechanical and metabolic trauma to OHC (Kujawa
and Liberman, 2009). Studies suggest that this noise-induced loss of afferent synaptic
connections between IHC and auditory nerve fibers may be a potential cause of suprathreshold
hearing deficits (Hickox and Liberman, 2014; Hickox et al., 2017). This pathology has been
termed cochlear synaptopathy, and it produces a pattern of change in which auditory thresholds
are not affected but suprathreshold deficits such as poorer hearing in noise may be present. To
evaluate whether intensity coding, a critical component of hearing in noise, is also affected by
noise-induced cochlear synaptopathy, my studies will assess chinchilla IID threshold before and
after a noise exposure aimed to induce synaptopathic damage to IHCs.
I hypothesized that IID thresholds would increase following both carboplatin treatment and noise
exposure, suggesting that IHCs synaptic damage or loss of IHCs is correlated with deficits in
intensity coding. If this hypothesis is supported, IID tasks could be used as a suprathreshold
assay of IHC cochlear damage.
The goal of this study was to investigate the effects of IHC loss or damage on the chinchilla’s
ability to detect intensity changes. Each animal completed a battery of baseline assessments,
including a psychoacoustic pure tone detection task and an intensity increment detection task.
Physiological assays of cochlear function were also obtained. These included distortion product
otoacoustic emissions (DPOAEs), a measure associated with OHC function, Auditory Brainstem
Response (ABR) thresholds, and ABR wave-1 amplitude measurement, measures associated
with the neural output of both the cochlea and the auditory brainstem. These baseline measures
were re-assessed following two distinct patterns of cochlear damage induced by (1) selective
IHC loss from 75 mg/kg of carboplatin and (2) synaptopathic noise exposure. At the conclusion
of data collection, histological analysis of chinchilla cochleae was completed including
individual animal IHC, IHC synapses, and OHC counts. The degree of IHC and OHC loss was
correlated with post-lesion perceptual test results to assess the effects of cochlear damage on
intensity coding. The results from this study suggest that IHC likely play a critical role in the
ability to code changes in intensity given that deficits in intensity coding can directly impact an
individual’s ability to process speech, especially in the presence of background noise. IID could
be used to assess the presence of IHC damage or IHC synapse damage independent of traditional
measures of hearing which are primarily sensitive to OHC loss. Identifying the cause of these
suprathreshold speech in noise deficits may lead us further in identifying specific solutions for
these patients. Unlike most current electrophysiological tests that have been used to estimate
cochlear damage, this test is quick to administer and could easily be implemented in clinical
audiology practice. In humans, this task would take approximately 5 minutes indicating that a
healthcare provider could administer this test in a standard evaluation without having to schedule
a separate testing appointment
Explainable AI Algorithms for Classification Tasks With Mixed Data
No full textWith the great power of Machine Learning techniques, numerous applications have been
created that have become an integral part of our modern life. However, the decision making
processes of many of these machine learning-based applications are being questioned and
criticized due to their opacity to users, especially for critical tasks such as disease diagnosis,
loan application, industrial robots, etc. This opacity is the result of using statistical ma-
chine learning approaches that generate models that can be viewed as solutions to optimization problems that minimize loss or maximize likelihood. Explainable Artificial Intelligence
(XAI) models or Explainable Machine Learning (XML) models are machine-learned models
in which human users can understand the decision making or prediction making process.
The main goals of XAI are to: 1) generate highly accurate models that are comprehensible
to human users. 2) explain a model’s decision-making process to a human so that they
can easily understand it, develop trust in it, and diagnose any potential problems. This
dissertation presents the FOLD family of new explainable AI algorithms for classification
tasks that are able to efficiently handle mixed data (numerical and categorical) without extra effort (i.e., without resorting to any special data encoding). These algorithms generate
a set of default rules, represented as a stratified logic program, that serves as the predictive model. Due to their symbolic nature and because they are based on logic, they can
be easily understood and modified by humans. These new algorithms are competitive in
predictive performance with state-of-the-art machine learning algorithms such as XGBoost
and Multi-Layer Perceptrons (MLP), however, they are an order of magnitude faster with
respect to execution efficiency. The FOLD-R++ algorithm has been designed for solving
binary classification problems, FOLD-RM for multi-category classification problems, and
FOLD-LTR for ranking. FOLD-SE is a further improvement over these algorithms that
leads to scalable explainability. Scalable explainability means that regardless of the size
of the data, the generated model is represented using a small number of rules—resulting
in improved human-interpretability and human-explainability—while maintaining excellent
predictive performance. The rest of this thesis presents the FOLD family of algorithms and
compares and contrasts them with state-of-the-art machine learning algorithms
Deep Probabilistic Models for Step Recognition and Step Localization in Egocentric Videos
No full textIn this thesis, we present a novel approach for solving two important procedural activity
recognition tasks. Given a video depicting users performing procedural tasks, such as executing
steps in a cooking recipe, the goal in our first task, called step localization, is to either
determine the start and end times of the step in the video or indicate that the step was not
performed. Our second task is dynamic, wherein, given a video stream and a step, our objective
is to identify whether the step was performed in the available video stream or not. To
address the first task, we propose an approach that utilizes a transformer architecture introduced
in previous work called ActionFormer, combined with various video feature extractors
such as Omnivore, SlowFast, Videomae, and 3DResnet. Our experimental results clearly
demonstrate that Omnivore features achieve the best performance. For the second task, we
introduce a deep probabilistic model that tracks the steps a user has performed and those
they have not. Our experiments indicate that our deep probabilistic model outperforms the
baseline model that relies on Omnivore features
Mathematical and Computational Frameworks for Modeling Mass Transport Phenomena Across Biological Systems
No full textMultiscale mathematical modeling of transport phenomena across different levels of biolog-
ical systems, such as cells, capillaries, tissues, and organs, has been increasingly helpful in
describing how interactions among these systems lead to their function and dysfunction. The
development of models across these scales is based on knowledge from various fields, such
as engineering, physiology, and biophysics, and it requires significant collaboration among
scientists from the relevant areas. Moreover, mathematical modeling of membrane trans-
porters and ion channels helps researchers obtain a thorough understanding of the complex
regulation process between membrane transporters, estimate the changes and uncertainties
of their mechanisms, and study how disorders in these procedures may cause disease. A
unified framework to describe the fundamental principles and unite the established models
could support this growing research community.
In this regard, the first part of this work provides a mathematical and computational frame-
work for modeling molecule transport, such as nutrients, inorganic ions, drugs, and toxins,
across the cell through membrane transporters and ion channels along with the associated
database. Our purpose is to substantially save time and resources needed to find the math-
ematical models available for different classes of membrane transporters and improve com-
munication between scientists with different backgrounds interested in this field. To achieve
these objectives, this work first deals with the essential terminology required to understand
and model biological transport mechanisms, as well as to compile currently available mod-
els. Afterward, an inclusive mathematical framework for predicting the mechanisms of mass
transport in various tissue compartments, such as cells, capillaries, and gland ducts, is devel-
oped with a primary focus on membrane-mediated transporters such as channels, uniporters,
symporters, pumps, and antiporters. We present a comprehensive and up-to-date set of cur-
rently available mathematical equations for modeling ion channel and membrane transporter
mechanisms. With this unified mathematical framework that includes most of the available
models for ion channels and transporters, scientists can select the right model more conve-
niently.
Finally, a comprehensive database of parameters relating to mathematical and computa-
tional transport models is developed in order to facilitate efficient research on the mem-
brane transporters. TransporterDB is a biophysical database that contains data on the
kinetic parameters of membrane transporters and ion channel parameters such as conduc-
tance and electric mobility for each ion. TransporterDB can be accessed through a web
browser (https://transporterdb.org/) and GitHub repository (https://github.com/
stamp-cell/TransporterDB). Equations for the parameters were taken from our previ-
ously published comprehensive mathematical toolkit, which contains over 200 sources. While
TransporterDB has not been fully completed, it contains information on more than 80 models
of human ion channels and membrane transporters. The source of data is from in vivo and
in vitro studies on various mammals. Users and readers are encouraged to submit additional
data and information.
We have completed the first part of this dissertation by a validation study. The developed
mathematical-computational framework is validated for simulating two different cases: 1)
human mammary epithelial cell in the breast and 2) cardiac action potentials.
The second part of this work explores, an integrated machine learning (ML) and Computa-
tional Fluid Dynamics (CFD) simulation approach to predict milk flow behavior in lactating
breasts. Using CFD to solve fluid flow problems can be both computationally intensive and
time-consuming. Artificial neural networks (ANN) are capable of learning complex depen-
dencies between high-dimensional variables. This work uses this capability to develop a
novel data-driven approach to CFD. To that end, a fully integrated CFD and ML workflow
is developed to predict milk flow behavior in lactating breasts. CFD simulations are used to
develop the training and validation data sets, and a machine learning workflow is developed
to train the ANN. In addition, different ANN designs are proposed and their prediction re-
sults are compared. Finally, we employ the design of the experiment method to determine
the minimum number of simulations required to determine an accurate prediction. Our work
shows that by integrating CFD and ML approaches (ML-CFD), one can train a neural net-
work to produce CFD results, thereby saving both time and resources required for running
CFD simulations. This ML-CFD approach provides a capability to build ML models capable
of predicting milk flow behavior at a per-lactating mother level faster than traditional CFD
solvers