Search

Despite over 30 years of research, face recognition is still one of the most difficult problems in the field of Computer Vision. The challenge comes from many factors affecting the performance of a face recognition system: noisy input, training data collection, speed-accuracy trade-off, variations in expression, illumination, pose, or ageing. Although relatively successful attempts have been made for special cases, such as frontal faces, no satisfactory methods exist that work under completely unconstrained conditions. This thesis proposes solutions to three important problems: lack of training data, speed-accuracy requirement, and unconstrained environments. The problem of lacking training data has been solved in the worst case: single sample per person. Whitened Principal Component Analysis is proposed as a simple but effective solution. Whitened PCA performs consistently well on multiple face datasets. Speed-accuracy trade-off problem is the second focus of this thesis. Two solutions are proposed to tackle this problem. The first solution is a new feature extraction method called Compact Binary Patterns which is about three times faster than Local Binary Patterns. The second solution is a multi-patch classifier which performs much better than a single classifier without compromising speed. Two metric learning methods are introduced to solve the problem of unconstrained face recognition. The first method called Indirect Neighourhood Component Analysis combines the best ideas from Neighourhood Component Analysis and One-shot learning. The second method, Cosine Similarity Metric Learning, uses Cosine Similarity instead of the more popular Euclidean distance to form the objective function in the learning process. This Cosine Similarity Metric Learning method produces the best result in the literature on the state-of-the-art face dataset: the Labelled Faces in the Wild dataset. Finally, a full face verification system based on our real experience taking part in ICPR 2010 Face Verification contest is described. Many practical points are discussed.

This dissertation is a collection of three essays in applied microeconomics. In the first essay, I study the enrollment effects of Tennessee Promise, a state-funded financial aid program in the two-year college sector. I document a substantial response in full-time first-time enrollment and some degree of substitutability between the unsubsidized and subsidized college sectors following the inception of the Tennessee Promise program. The second essay takes initial steps to examine the broader implications of free college. I first use institution-level data to construct three metrics of diversity, namely the share of underrepresented minorities in the student body, Simpson’s index, and Shannon’s index. I then leverage the staggered timing of statewide “free college” programs across the United States to document these programs’ impacts on college diversity. Overall, free-college programs have not catalyzed an economically significant effect on the demographic composition of public two-year colleges, although initial results do offer grounds for optimism. This essay places the first essay in perspective. In relation to the literature, the findings complement recent work, which has documented the early success of promise programs in encouraging community college enrollment. To this end, I provide several policy implications to improve practice. The third essay is a joint work in which my coauthor and I seek to understand US immigrants’ health-related behaviors and outcomes. We simultaneously examine risky consumption choices (smoking and drinking) and physical health conditions (asthma, diabetes, vision problems, and coronary heart diseases) using data from the National Health Interview Surveys (1989-2018). We incorporate cohort fixed-effects and the interactions between cohort effects and years since immigration into our empirical framework to capture the dynamics of immigrant health over time. For all health indicators, we find that there are important differences between arriving immigrants and natives. Despite some heterogeneity in the dynamics of unhealthy behaviors, this heterogeneity seems to dissipate as we explore longer-term health outcomes. Overall, our findings provide an interesting outlook on how the integration into the host society affects American immigrants’ health. We contribute new results to the immigrant assimilation literature, which has primarily focused on obesity and wages.

Cancer is the second leading cause of death in the United States, with an estimated 1.8 million new cases and 600,000 deaths annually in 2020. The biggest hurdle in cancer treatment is in completely eradicating all invasive tumor cells, which prevents cancer recurrence and enhances survival rate. Laser based interventions offer unique abilities to sense cancer cells and stimulate biological responses, both of which can be used to enhance cancer treatment. This thesis addresses two interventions aimed at improving cancer survival rate: 1) guidance of surgical resections and 2) systemic immune therapeutics. In the first part of the thesis, we applied non-destructive, fast, and label-free spectroscopy approaches for surgical guidance in skin cancer excision. Our study utilized multimodal spectroscopy with diffuse reflectance, Fluorescent, and Raman to assist tumor margin assessment. Our findings suggested that Raman spectroscopy provided the best diagnosis power among the three techniques and combined modality. We further studied the correlation between Raman measurements and its underlying tissue anatomy. We identified key diagnostic markers to accurately differentiate nonmelanoma cancer from normal skin, highlighting this method's potential in intraoperative tumor margin assessment. In the second part of the thesis, we used laser nanobubbles (bubbles generated around nanoparticles from irradiation with nanosecond pulsed laser radiation) as novel cancer therapy to trigger immunogenic cell death in breast cancer cells. We also applied a new approach to characterize laser nanobubbles' dynamics with optical coherence tomography. We demonstrated for the first time that laser nanobubbles provided a fast, highly specific therapy to eradicate tumor cells and elicit immunogenic cell death, highlighting this approach's potential as a candidate combination strategy for immunotherapy.

Chapter 1: High-skilled migration and higher-education attainment rose globally in the period between 2000 and 2010. These trends raise the question of whether increases in higher-education attainment led to increases in high-skilled emigration in source countries during this period? To the contrary, I found a strong negative correlation between percentage changes of high-skilled emigration rates and percentage changes of high-skilled proportions in non-OECD countries. Using the IAB Brain Drain, Barro-Lee and World Development Indicators Datasets, I estimated the elasticity of high-skilled emigration rates with respect to high-skilled percentages in non-OECD countries with ordinary least square, least absolute shrinkage and selection operator (LASSO) and post-double-selection LASSO. The estimates ranged between -0.78 and -0.98. Chapter 2: In 2015, the state of Georgia eliminated exit exams from high school graduation requirements in an education reform. I study this reform's impact on 4-year high school graduation rate, post-secondary institutions enrollment and employment rate among new high school graduates. Using the state of Florida as control, I estimate weighted difference-in-difference analyses at the school level for graduation rates and at the district level for post-secondary institutions enrollment and employment rates. The reform increased graduation rate by 4% for all students, and by 1.4% for White, 5.98% for Black and 5.59% for Hispanic subgroups. The reform decreased post-secondary institutions enrollment rate by 1.8% but had no statistically significant impact on employment rate among new high school graduates who don't pursue post-secondary education. In the pre-periods, the trends are parallel for graduation rates, post-secondary institutions enrollment rates and employment rates between Georgia and Florida. I conclude that although exit exams elimination increased 4-year graduation rates (especially among Black and Hispanic students), the additional diploma holders are less likely to enroll in post-secondary institutions and no more likely to be employed when choosing not to pursue postsecondary education. Chapter 3: I propose a new semi-parametric estimator for the Tobit Type-3 sample selection model. I first generalize the mode regression estimator to its bivariate counterpart. I then apply this bivariate mode regression estimator to the Tobit Type-3 model under two separate sets of assumptions. First set will be computationally simpler but requires more restrictive assumptions, while second set will be more computationally taxing yet requires less restrictive assumptions. I then show that the estimators are identified and strongly consistent under both sets of assumptions.

Wave propagation is ubiquitous in science and engineering applications, but solving the second-order wave equation in a parallel way is still computationally challenging. Specifically, as efficiency gained from spatial domain decomposition is saturated, time-domain becomes the next candidate for parallelization. However, most parallel-in-time methods are not effective in solving hyperbolic problems, including the wave equation. Motivated by the simple parareal algorithm developed by Lion, Maday, and Turinici, we propose a new parallel scheme called [theta]-parareal that generalizes the original parareal. The convergence and stability analysis of the [theta]-parareal scheme reveal the deficiency of the parareal method when applying to highly oscillatory problems. We then develop a new parallel-in-time iterative method for solving the homogeneous second-order wave equation. The new approach is a data-driven strategy in which we use pre-computed data to stabilize the iteration by minimizing the wave energy residual. We propose two techniques, a linear algebra-based method, and a deep neural network method. Numerical examples, including a wave speed with discontinuities, are provided to demonstrate the effectiveness of the proposed methods on the wave equation.

Ras is a key regulator of multiple signaling networks controlling a variety of important biological processes such as cell proliferation, survival, differentiation, and migration. Interestingly, Ras also plays a critical role in the regulation of proper learning and memory. Neurofibromin is one of the most common and well-studied negative regulator of RasGAPs, which inhibits Ras activity. Neurofibromin is the protein product of neurofibromatosis type 1 (NF1) gene, which inactivates the proto-oncoprotein Ras through its GTPase activating protein domain. Mutations on the NF1 gene will cause neurofibromatosis type 1 disease, which is a common human genetic disorder characterized by phenotypes ranging from increased tumor susceptibility to learning disabilities. NF1 alternative exon 23a lies in the center of the GAP domain and is skipped in brain but included in other tissues. To study the molecular and biological functions of this regulated splicing event in vivo, we generated mutant mice (Nf123aIN/23aIN mice) with constitutive Nf1 exon 23a inclusion in all tissues, and found increased Ras activation in their brains. In addition, the brain of Nf123aIN/23aIN mice also has elevated levels of phosphorylated ERK1/2 downstream of Ras. Notably, Nf123aIN/23aIN mice show deficiencies in both short- and long-term spatial learning, as well as in fear associative learning. Taken together, these results strongly suggest that the regulated inclusion of Nf1 exon 23a modulates Ras signaling, and is important for proper learning and memory in mice. Therefore, our knock-in mouse is a very helpful tool for future studies on the role of alternative splicing in Ras/ERK signaling and in learning and memory.

Electrical stimulation of tissues induces cell alignment, directed migration, extended processes, differentiation, and proliferation, but the mechanisms involved remain largely unknown. To reveal effects of electric fields (EF) through the media on cell behavior, voltage (7.45 – 22 V), current density (36 – 106 mA/cm2), duration (2 – 24 hrs), and alternating currents (AC, 2 – 1000 Hz) were varied independently when exposed to cell cultures. It was determined that current density and duration are the primary attribute Schwann cells respond to when an EF is applied through the media. This implies that the number of charges moving across the cell surface may play a key role in EF-induced changes in cell behavior. Identical conditions were used to stimulate cells grown on the surface of a conductive substrate to examine if a scaffold can provide structural and EF cues. The effects of an EF through the substrate were examined by placing a protein gel on the surface during stimulation and observing the morphology of subsequent cell cultures and the physical topology of the gel. EFs were shown to create Ca2+ redistribution across gels and subtle changes in collagen I fibril banding. Stimulated gels were able to induce perpendicular Schwann cell alignment on newly seeded cultures days after initial EF exposure, and the cell response decreased when seeded at longer times, indicating the effects of EF on the matrix environment has a relaxation time. These findings were then integrated into a biodegradable, electrically conductive polypyrrole-poly-ε-caprolactone polymer developed by collaborators. Dorsal root ganglia placed in matrix gels on top of conducting polymer exhibited significantly longer axons when stimulated with DC and AC signals. The overall results demonstrate that EFs have a significant effect on the extracellular environment. The broad implication of this data grants researchers with the ability to physically and metabolically control cell behavior with EFs, including improved wound healing or reduced cancer metastasis.

The International Electrotechnical Commission (IEC) standard 61400-1 for the design of wind turbines does not explicitly address site-specific conditions associated with anomalous atmospheric events or conditions. Examples of such off-standard atmospheric conditions include thunderstorm downbursts, hurricanes, tornadoes, low-level jets, etc. This study is focused on the simulation of thunderstorm downbursts using a deterministic-stochastic hybrid model and the prediction of wind turbine loads resulting from these simulated downburst wind fields. The wind velocity field model for thunderstorm downburst simulation is first discussed; in this model, downburst winds are generated separately from non-turbulent and turbulent parts. The non-turbulent part is based on an available analytical model (with some modifications), while the turbulent part is simulated as a stochastic process using standard turbulence power spectral density functions and coherence functions. Tower and rotor loads are generated using simulation of the aeroelastic response for models of utility-scale wind turbines. The main objective is to improve our understanding from the point of view of design so that we may begin to address transient events such as thunderstorm downbursts based on the simulations carried out in this research study. The study discusses as well the role of control systems (for blade pitch and turbine yaw), of models for representing transient turbulence characteristics, and of correlated demand and loads on multiple units in turbine arrays during thunderstorm downbursts.

In this dissertation, we formulate and implement p- adaptive and hp-adaptive finite element methods to solve elliptic partial differential equations. The main idea of the work is to use elements of high degrees solely (p- adaptive) or in combination with elements of small size (hp-adaptive) to better capture the behavior of the solution. In implementing the idea, we deal with different aspects of building an adaptive finite element method, such as defining basis functions, developing algorithms for adaptive meshing procedure and formulating a posteriori error estimates and error indicators. The basis functions used in this work are regular nodal basis functions and special basis functions defined for elements with one or more edges of higher degree transition elements). It is proved that with our construction of these basis functions, the finite element space is well- defined and C⁰. Several algorithms are developed for different scenarios of the adaptive meshing procedure, namely, p-refinement, p-unrefinement and hp-refinement. They all follow the 1-irregular rule and 2-neighbor rule motivated by [Bank and Sherman, 1983 - MR751598]. These rules help to limit the number of special cases and maintain the sparsity of the stiffness matrix, and thus to simplify the implementation and reduce the cost of calculation. The work of formulating a posteriori error estimates and error indicators is the core of this dissertation. Our error estimates and error indicators are based on the derivative recovery technique proposed by [Bank and Xu, 2003 - MR2034616, MR2034617] and Bank et al., 2007 - MR2346369]. Using the information in formulating the error indicators, we define a hp-refinement indicator which can be utilized to decide whether a given element should be refined in h or in p. Numerical results show that the combination of the two indicators helps automatic hp-refinement to create optimal meshes that demonstrate exponential rate of convergence. In this dissertation, we also consider hp-adaptive and domain decomposition when they are combined using the parallel adaptive meshing paradigm developed by [Bank and Holst, 2000 - MR1797889]. Numerical experiments demonstrate that the paradigm scales up to at least 256 processors (maximum size of our experiments) and with nearly 200 millions degrees of freedom