Treasures @ UT Dallas
Not a member yet
7697 research outputs found
Sort by
Machine Learning Based Prediction in FPGA CAD
No full textTechnological advances have allowed the continuous improvement of modern electronic systems. Enabled by the scaling of technology nodes, current integrated circuits are becoming
increasingly complex. These intricate designs require EDA tools to ensure the rapid creation of complex new-generation architectures. Traditionally, a hardware IC design engineer
designs the chips with the help of RTL language like Verilog, VHDL, or System Verilog.
However, creating chips using RTL descriptions becomes challenging for the new generation
of complex architectures addressing applications like deep learning and computer vision. As
a result, designers nowadays use high-level languages like C/C++ or System C to design the
chips. The design flows consist of multiple stages, from C-synthesis (converting C/C++ code
to RTL code) to place and route. Each step is highly time-consuming, and the performance
of each stage is very much dependent on the characteristics of the previous stage.
The use of machine learning (ML) to help speed up electronic system design is becoming
prevalent across the industry. This dissertation is a fusion of ML and EDA tools. We have
solved multiple electronic design automation (EDA) problems for field-programmable gate
arrays (FPGAs) technology. This dissertation applied ML to solve five significant FPGA
physical design automation problems. Design closure in general VLSI physical design flows,
and FPGA physical design flows are important and time-consuming problems. Routing can
consume as much as 70% of the total design time. The first contribution in the dissertation
is a machine learning-based post route routing congestion estimation tool, where we applied
regression models on post placed FPGA netlist. Limited availability of training data is
one of the significant challenges researchers face in applied machine learning in EDA flow.
Therefore, training data quality and quantity play an essential role in the generated model in
any machine learning application. The second contribution we addressed in this dissertation
is to propose a methodology to create vast training design sets from a single HLS code. Highlevel synthesis (HLS) tools allow designers to prototype ideas on various FPGAs and ASIC
platforms quickly. However, the Quality of Results (QoR) reported after HLS synthesis is
highly inaccurate. This suboptimal QoR may result in false-positive closure of timing and
area, which may not close after routing. The third contribution addressed in this dissertation
is a robust ML-based design flow that can accurately predict post-route QoR for a given
behavioral description without the need to synthesize the design. The fourth contribution of
this dissertation is a generalized resource and performance estimation model for convolution
neural network (CNN) architecture. Design space exploration (DSE) of HLS designs is the
process of finding a set of optimized designs for area and performance. Traditional DSE is
time-consuming, and the resulting design points are not always optimal as the designers rely
on post HLS synthesis results. This dissertation’s fifth and final contribution is to design
a fast design space explorer that combines traditional metaheuristics-based methods and
machine learning to generate an optimal set of designs that are minimum both in terms of
area and latency
Algorithms for Complex Explanation Queries
No full textRecently, there has been growing interest in developing machine learning systems that are
robust to minor perturbations in input and are explainable in that they are able explain
why they made a particular decision to a user. In this dissertation, we focus on a widely
used graph-based probabilistic representation called probabilistic graphical models (PGMs)
and propose a new unified constrained optimization task called constrained most probable
explanation (CMPE) problem over them. We show that in addition to the two aforementioned tasks, making models robust and explainable, many real-world tasks over PGMs can
be reduced to CMPE. We thoroughly analyze the computational complexity of CMPE and
develop specialized, practical algorithms for solving it. Specifically, we show that CMPE
is strongly NP-hard in general and thus unlikely to admit an efficient, polynomial time algorithm. However, it is only weakly NP-hard when the PGM exhibits certain structural
properties such as small k-separators.
To solve CMPE and upper bound its optimal value, we present novel efficient approaches
that combine graph-based partitioning techniques with approximation algorithms developed
in literature on the multiple choice knapsack problem (MCKP), number partitioning problem
and subset sum problem (SSP). We derive approximation and complexity guarantees for
our new algorithms and demonstrate via a large scale experimental evaluation on several
benchmark graphical models that compared to baselines such as random sampling and local search, our new algorithms are more accurate. We also encode CMPE as a mixed integer
linear program (MILP) and show that our approach is superior to open-source MILP solvers
such as SCIP
Cornerstones and Intersections of the Dairy Industry: Regulation, Labor, and Demand
No full textThe dairy industry, directly and indirectly, contributes $753 billion and 3.3 million American jobs to the economy. Dairy by itself directly contributes 1.1 percent to the U.S. GDP. In this dissertation, we look at three cornerstones of the dairy industry: regulation, labor, and demand. In Chapter 2, we first look at a general regulatory theory based on current dairy regulations and we test the theory using an experiment. We find both theoretically and experimentally that regulation can beget a win-win situation for both buyers and sellers. In Chapter 3, we switch gears and look at the supply side of the dairy industry, focusing on the wages for dairy farm laborers. We find that both agglomeration and immigration enforcement impact dairy farm laborers’ wage rates. In particular, we find agglomeration increases the wage rate, as we expect. But that in industries, like dairy, where there are a large immigrant workforce, immigration enforcement reduces the wage rate. Next in Chapter 4, we pivot and analyze the regional demand for finished dairy products using IRI scanner data. Here we find using scanner data provides us a detailed look at regional demand for dairy products. Finally in Chapter 5, we study the intersection of these three cornerstones in a linear programming multi-commodity model. This chapter uses the knowledge gained in Chapters 3 and 4 to update the data and add supply and demand curves to a model that has a long standing history in the dairy industry. We then use this updated model to test the theory outlined in Chapter 2, and find that in this multi-commodity model the regulatory scenario is more efficient than the unregulated scenarios
Affectionate Touch and Well-being: the Role of Health Behavior and Coping Behavior
No full textFriends embrace upon greeting, and couples cuddle on the couch while watching a movie.
Affectionate touches like these are commonly experienced throughout the lifespan, and growing
empirical evidence suggests receiving affectionate touch in the context of adult, close
relationships enhances individuals’ well-being. The two studies in this dissertation tested a
conceptual model for how experiences of affectionate touch in close relationships may relate to
health behavior and coping behavior as well potential mediating factors (i.e., self-efficacy and
self-control as an indication of self-efficacy) and moderating factors of this association (i.e.,
motivations for using affectionate touch, perceived responsiveness, and the type of relationship).
Lastly, these studies tested whether affectionate touch experiences in close relationships act as a
moderating factor between perceived stress and health behavior and coping behavior. Study 1
assessed this conceptual model cross-sectionally (N=861 individuals) and examined affectionate
touch experiences within romantic relationships, friendships, and in sibling relationships. Study 2
tested this model in a two-week daily diary study of couples (N=116 couples). Across the two
studies, it appears that affectionate touch experiences in romantic relationships may be more
relevant for coping behavior than for health behavior. In Study 1, there was preliminary evidence
that greater affectionate touch in romantic relationships predicts greater support seeking behavior
and that perceiving one’s partner to be more responsive to them may potentiate the influence of
affectionate touch on emotional and practical support seeking behavior. Study 2 confirmed that
daily affectionate touch experiences are also associated with greater support seeking behavior.
However, perceived partner responsiveness on the daily level appears to be an independent
predictor of support seeking. In Study 1, self-efficacy was not a mediating factor, and
motivations for providing touch and the type of relationship were not moderating factors.
Finally, it does not appear that affectionate touch buffers the impact of stress on these tested
health behaviors and coping behaviors. This dissertation is the first project to assess how
affectionate touch experiences in close relationships relate to health behavior and coping in an
effort to improve our understanding of the broader link between affectionate touch and wellbeing and, in doing so, provides important insight into the ways affectionate touch may promote
adaptive coping behavior to enhance well-being which has implications for future research and
intervention work
Neural Oscillations Underlying Automatic Word Reading
No full textAutomatic word reading is the process of instantaneously and effortlessly reading words from
memory. Although a quotidian and seamless task for many, automatic word reading requires the
systematic orchestration of multiple fast and simultaneous neural and cognitive processes. Here I
used EEG, a temporally robust tool that also provides a window into the simultaneously engaged
neurocognitive processes, to study how neural oscillations in the beta and the alpha frequencies
change as skilled adult readers learn to automatically read novel pseudowords. Automaticity was
elicited via a rhyming task that included four presentations of each target word and required
participants to decide if the target word (e.g., daik) rhymed with its word pair. Participants also
completed a set of behavioral assessments that provided measures of few cognitive skills
believed to be important for automatic word reading. Findings revealed that beta power
decreased across exposures during 150-250 ms post target onset over the left posterior areas of
the brain. Alpha power also decreased across exposure during 150-350 ms post target onset over
bilateral posterior regions. A Pearson correlation analysis between exposure related changes in
neural oscillations and the collected behavioral measures revealed a significant correlation
between exposure related beta changes and performance in the CTOPP elision – a behavioral
assessment that consistently predicts later reading abilities. Taken together, findings revealed
that neural oscillations in the beta and the alpha frequencies are implicated during the process of
acquiring automatic word reading. This suggests that examining these EEG frequencies has the
potential to provide key insights about the neurocognitive architecture underlying the complex
and important process of automatic word reading
Beyond Data: Efficient Knowledge-guided Learning for Sparse and Structured Domains
No full textThe field of AI has made great advances in recent years. Most of these advances have focused
on leveraging more data and finding new architectures to improve system performance. However, collecting data can lead to exorbitant costs. This is especially the case for structured
domains where the data conforms to some standardized format (like tabular data, relational
databases, etc.). In structured domains, an expert might be required to collect and organize
data; necessitating time and effort. Further, learning explicitly from data is neither sufficient
nor favorable. Enormous data can cause concerns for safety, lack of fairness, and a substantial carbon footprint. So looking beyond learning from data, this dissertation focuses on
finding principled ways to leverage rich human knowledge for sparse and structured domains
to guide the learning procedure.
In particular, this dissertation looks at four challenges that arise when models are learned in
structured domains and propose to tackle them using explicit human knowledge. First, we
consider the challenge of learning from sparse and noisy data in the successful gradient boosting framework and propose to use domain-specific trend information to improve prediction.
Second, we consider the challenge of learning to generalize across multiple tasks and objects in sequential decision making. We address this challenge by proposing a framework that
takes inspiration from human’s ability to generalize by identifying compositionality and generating abstract representations. Third, we consider the challenging task of human-machine
collaborative problem solving and propose a framework that uses natural language communication for effective bi-directional interaction. Finally, the fourth challenge we consider
is the problem of a large hypothesis space when dealing with domains with heterogeneous
objects. We identify the lack of a language bias—typed object representations—in recent
neurosymbolic architectures and devise an approach to incorporate the bias.
In this dissertation, we demonstrate various ways to incorporate domain-specific knowledge
from humans in training AI systems. We conclude that using domain knowledge not only
reduces the sample complexity but also improves the performance and generalization abilities
of the model
Computational Modeling of FGF10 Mediated Buckling Morphogenesis Within the Embryonic Airway Epithelium
No full textThe embryonic lung is an excellent model system to study mechanisms of branching
morphogenesis, since the embryonic airways undergo a series of recursive branching events to
build the bronchial tree. This process involves reciprocal signaling interactions between the
branching airway epithelium and a surrounding layer of pulmonary mesenchyme. Focal regions of
fibroblast growth factor (FGF10) expression within the pulmonary mesenchyme are thought to
provide a biochemical template for the overall airway branching pattern. Recent work, however,
has shown that mechanical forces can also impact this process, and it remains unclear how patterns
of FGF10 expression interact with biophysical cues in the developing lung to sculpt incipient
branches. This is, in part, owing a lack of computational models that incorporate both growth-factor diffusion and the mechanics of growth and remodeling. Several previous studies have used
computational modeling to suggest how specific spatial patterns of FGF10 expression might arise
spontaneously within the pulmonary mesenchyme, while others have used continuum mechanics
to identify the forces involved in bud initiation, but no computational framework has been
developed which unite these different approaches. Here, we developed a finite-element model in COMSOL Multiphysics, which couples growth-factor diffusion to the mechanics of epithelial morphogenesis. We first consider the axial growth
of a constrained bar, in which the components of the growth tensor depend on the local
concentration of a diffusible molecule. This framework is then extended to determine how growth
factor diffusion from a focal source within the pulmonary mesenchyme elicits patterns of epithelial
growth and budding morphogenesis along the embryonic airway epithelium. This work highlights
the importance of both mechanical forces and growth factor diffusion during airway branching
morphogenesis
Quantum Search on Molecular Graphs and Word-representable Line Graphs
No full textThe work presented in this thesis broadly falls under two quantum search procedures: quantum search using a continuous-time quantum walk and quantum search using adiabatic quantum evolution. The searches are conducted on three different graphs: the molecular graph of
graphene, the molecular graph of alkane, and on the line graphs of (non-)word-representable
graphs. For the latter, a classical solver was also used to resolve a long-standing question.
Quantum search using a continuous-time quantum walk was implemented for a marked
node on the molecular graph of graphene, the hexagonal lattice. It was established that
including second-nearest-neighbor coupling does not have a significant impact on the search
time but does seem to improve the probability of success. This builds on the work of Foulger,
Gnutzmann, and Tanner on quantum search on graphene who considered only the nearest-neighbor interactions.
Quantum search using adiabatic quantum computation was implemented for (a) searching
for isomers of alkanes, (b) searching for the most influential node in a network, and (c)
for the decision problem of word-representable line graphs. Quadratic unconstrained binary
optimization (QUBO) formulations were proposed for these problems to be embedded and
solved on both quantum annealing and gate-based quantum computers such as D-Wave
quantum annealers and IBM quantum computers. The dynamics of the search is described
by a quantum system evolving under a slowly varying Hamiltonian. Based on the adiabatic
theorem, the search evolves from the ground state of an initial Hamiltonian H0 to the ground
state of a problem Hamiltonian HP . The primary objective in each case was to construct an
efficient HP for the search problem.
Using the QUBO results from the D-Wave 2000Q and IBM Q QASM simulators, we were
able to verify and search for all isomers of alkanes with at most 9 carbon atoms and for
the most influential node of undirected networks using eigenvector centrality. Finally, our
formulation of the decision problem of word-representable graphs showed that the line graphs
of non-word-representable graphs are not always non-word-representable. This answers a 10-
year-old open problem
Mythic Lives of Austen, Shelley, and the Brontes in Biography, Literature and Film
No full textThis dissertation considers the impact that nineteenth-century British authors Jane Austen, Mary
Shelley, Charlotte Brontë and Emily Brontë have exerted in the British and American
imagination over time through the creation of their “mythic life stories” in biography, literature,
and film. Each of these authors has attained a degree of acclaim and popularity since their time
through the repeated theatrical, literary, and cinematic adaptations of their famous novels.
Likewise, their very life stories, retold through numerous biographies and biopics, have become
larger-than-life in our cultural consciousness, and I focus particularly on these life accounts. I
find that biographical accounts function in much the same way as literary and cinematic
adaptations. That is, they are also products of a particular creator’s time and culture, they are
formed according to principles of intertextuality, and they also create “powerful emotional,
aesthetic, and ideological resonances” in a particular culture’s psyche, as biographical critic
Lucasta Miller describes it.
An examination of the mythic life evolution of these four authors merits attention because we
find particular critical approaches or orientations at play that reveal how literary and biographical
products create and sustain mythic power. In the case of these four authors, their mythic life
representations are best understood in relation to both Romanticism and feminism. These authors
are very often interpreted as “Romantic authors” who transmute their personal life experiences
into the creation of their work. Furthermore, as women living in early nineteenth-century
England where gender roles dictated they live quiet domestic lives outside of the public realm,
early biographical accounts build a picture of these women as domestic saints, as self-effacing
sufferers, as passive receptacles of creativity. What I discover to be characteristic of subsequent
biographical representations is that these earlier nineteenth-century Romantic and gender views
have continued to powerfully shape the “mythic lives” of each of these authors. In more recent
decades, twentieth-century feminism has changed the landscape for how these female authors
have been perceived and depicted in biographical accounts, and though due attention has been
given to the way gender ideology may have restricted their thought and their literary activity, not
as much attention is given to the complexity of these women’s lives. For instance, many of these
female authors found ways for meaningful self-expression within the gender ideology of their
time, despite their time period’s restrictions on such expression.
In my study I affirm the danger of what most often occurs in biographical representations: the
tendency to over-simplify and romanticize a subject’s life. A focus on molding these writers, or
any artist, into the image of the “Romantic author” is a simplification of the authorial process,
just as seeing them as models of the ideal nineteenth-century woman is a distortion of their
complex personalities and desires. Likewise, a recasting of their lives and personalities
according to modern feminist views is a misrepresentation of the nuanced and contradictory
nature of their life views. Can we become better consumers of literary and biographical works
through a greater effort to acknowledge and understand the complexity of a human life as well as
the creative process of an artist? I hope my study will shed light on this topic and its important
relevancy today
Cell Nuclei Segmentation Using Deep Learning Techniques
No full textPathological examination usually involves manual inspection of hematoxylin and eosin (H&E)-
stained images, which is labor-intensive, prone to significant variations, and lacking reproducibility. One of the fundamental tasks to automate this process is to find all the cell nuclei
in the H&E-stained images for further analysis. We attempt this problem using deep learning
techniques. First, we introduce a semantic pixel-wise segmentation technique using dilated
convolutions. We show that dilated convolutions are superior in extracting information from
textured images. H&E-stained images are highly textured, which makes dilated convolutions
an ideal technique to apply. Our dilated convolutional network (DCN) is constructed based
on SegNet, a deep convolutional encoder-decoder architecture. Dilated convolution layers
with increased dilation factors are used in the encoder to preserve image resolution. Dilated
convolution layers with decreased dilation factors are used in the decoder to reduce gridding
artifacts. Our DCN network was tested on synthetic data sets and a publicly available data
set of H&E-stained images. We achieve better segmentation results than state-of-the-art.
To further separate the instance of each cell nuclei, we adapt our DCN with a single shot
multibox detector (SSD) and achieve promising results. Our methods are computationally
efficient and can be run on a personal laptop computer. This work is the first step to
wards using mathematical models to generate diagnostic inferences and providing clinically
actionable knowledge to physicians and patients