12466 research outputs found
Sort by
Explicit Programming Strategies
Get PDFSoftware engineering research has highlighted the importance of both tools and knowledge-sharing in developer problem-solving. Despite the critical role of strategic knowledge, it is rarely shared explicitly among developers. This thesis addresses this gap by introducing new techniques and tools for supporting developers' problem-solving through the explicit sharing of programming strategies. I developed a novel notation called Roboto to help developers articulate their problem-solving strategies step-by-step and created HowToo, a platform to support the sharing and use of these strategies. Through five research studies, I examined the impact of following explicit programming strategies on developers' productivity and success, demonstrating that systematic use of expert strategies leads to more efficient and successful problem-solving. Additionally, this thesis explores the contextual factors that influence developers' choice of debugging strategies in web development, highlighting the complexity and variability in decision-making processes. These contributions demonstrate the value of explicit strategy sharing and provide practical tools for improving developer productivity and knowledge sharing
GROUP IDENTITY AND SOCIAL PREFERENCE
Get PDFThis dissertation focuses on the effect of group identity on social preferences. I examine different ways of building group identity and how it affects economic decisions. The first chapter examines whether anticipated future interaction can serve as a method to build group identity. The second chapter explores how anticipated future interaction can alter social distance. The third chapter investigates the economic consequences of losing group identity
Ground-based light curve follow-up validation observations of TESS object of interest TOI 5356.01
Get PDFContext: The current definition of an exoplanet is changing the more we learn about them. Several exoplanet detection missions have been launched over the years since the first one was discovered, each using different detection methods with the most successful one being the transit method. Detecting exoplanets does not always mean they are confirmed, and they must be checked and reviewed to determine their existence.
Aims: The goal of this paper is to present the results of a follow up observation of the exoplanet candidate: 5356.01. The observation and conclusion help to build upon our definition of an exoplanet and how they form
Exploring the role of polymorphonuclear myeloid-derived suppressor cells and soluble Lox-1 during thrombosis using a tissue-engineered blood coagulation model
No full textNeutrophils, the largest subset of an individual’s white blood cell population, are known as the “first responders” in wound and infection, beginning an immune response. Polymorphonuclear myeloid-derived suppressor cells (PMN-MDSCs) are a subset of immunosuppressive neutrophils that instead curtail the immune response. PMN-MDSCs and the soluble form of its protein marker, Lox-1, are implicated in pro-thrombotic disease states including acute respiratory distress syndrome (ARDS). Lox-1 is a receptor for negatively charged microparticles and apoptotic cells. It retains its binding capacity in its soluble form, leading us to hypothesize that soluble Lox-1 (sLox-1) may be part of the PMN-MDSC’s immunosuppressive response by binding to and neutralizing procoagulant microparticles. In this work, we used RNA sequencing and proteomics to determine that healthy human neutrophils in an ex vivo thrombosis model can shift to a PMN-MDSC signature within 4 hours of clotting, including expression and copious shedding of Lox-1. Western blot and proteomic analyses of plasma from ARDS patients showed that survivors and those that did not experience coagulopathy had increased sLox-1 bound to circulating microparticles, and survivors had increased sLox-1 plasma concentrations, supporting the hypothesis that sLox-1 may be protective in mitigating coagulopathy. Further proteomic and lipidomic analyses revealed that sLox-1 concentrations are closely related to platelet byproducts. However, platelet byproducts cannot induce shedding, rather a constituent in plasma is necessary. This finding opens the door to future studies for how to induce sLox-1 shedding as a therapeutic during thrombosis
Computational Approaches to Lexical Complexity Prediction and Simplification
Get PDFLexical Simplification (LS) automatically replaces difficult to understand words with easier alternatives whilst maintaining the original meaning of a sentence. LS serves as a preliminary step to Text Simplification with the aim of enhancing a text's accessibility for different target demographics, including children, adults with low literacy, second language learners, or for individuals with a reading disability, such as dyslexia or aphasia. The arrival of recent large language models (LLMs), including the latest GPT series models, Mistral, and others, has drastically changed how we approach LS. Previous research questions within the field have been answered or have become less popular, whereas others have increased in popularity. To this end, this thesis provides a synopsis of previous approaches to LS as well as my own contributions to the field. It discusses several past and new research questions regarding feature performance, demographic variables, domain adaptation, and the development of end-to-end LS systems
Dynamics of Runs on Banks: Assessing Vulnerable Banks Through a Quadrant of Instability, Identifying the First-to-Fail, and Tracking Potential Knock-On Effects
Get PDFThe three bank failures in the spring of 2023 required prompt and extraordinary measures, via a systemic risk exception and a Federal Reserve emergency lending facility, by financial regulators to attenuate the possibility of another financial crisis. The run on Silicon Valley Bank (“SVB”) revealed how a large bank could cause contagion among similarly situated financial institutions. Although stress testing imposed on banks since the 2008 financial crisis required more stringent liquidity thresholds, the testing is conducted by individual banks while making it difficult to analyze the contagion effects on the network of banks. The Bank Quadrant of Instability Model (“BQI model”) is a dynamic model that tracks the phases of the banking network from a pre-run stage of banks with exposure to liquidity risk, deemed as “vulnerable banks,” to a post-run stage where potential contagion, or “knock-on effects” are sequentially tracked by time. The three bank factors typically used to gauge liquidity are uninsured deposits, long-term investments, and interest rates. Banks with uninsured deposits and long-term investments greater than or equal to 50 percent form the quadrant of instability comprised of banks that may be vulnerable to a run. This BQI agent-based model can use either actual U.S. bank data or create proxy bank data. The U.S. bank data is available from the Federal Financial Institutions Examination Council (“FFIEC”). Bank regulators may assign risk amplifiers based on their supervisory assessments to use the model on a confidential basis. Whereas, proxy data may be preferable when modeling a different banking topology than exists today, such as researching the effects of changing the number of uninsured depositors per bank or the number of banks. The heterogenous bank agents have quantitative and qualitative variables. The quantitative variables include uninsured deposits, long-term investments, and interest rates, which are indicators of liquidity. This dissertation introduces a qualitative variable called an amplifier, which represents exposure to an exogenous shock, such as commercial real estate collapse, housing bubbles, counterparty risk, and negative social network communication, as occurred with SVB. The depositors are so called “zero-intelligence” agents that are memoryless and not aware of what other agents are doing. The agents represent a generic population of depositors who exhibit panicked herd mentality and find the vulnerable bank with the highest amplifier rating to withdraw funds. The depositors spread contagion from bank to bank as they drive banks into failure. A cascade of bank failures is a social phenomena that is deemed “run contagion” in this dissertation and may also be referred to as “knock-on effects” in bank regulator vernacular. The BQI model tracks the first bank to fail, the subsequent order of failures, and the days to failure. By understanding which banks are the most vulnerable to a run and how run contagion spreads throughout the banking network, bank risk managers and examiners may proactively monitor the risk exposure and shore-up liquidity before the exogenous shock occurs. The model, with an average of about 180 vulnerable banks, indicates that after the initial shock the first bank will fail on day four with all banks with the same high amplifier rating will fail by day eight. All vulnerable banks will fail by day 21. Lastly, the two variables that affect liquidity - interest rates and available-for-sale securities - were tested at high, medium, and low rates. The changes in interest rates resulted in run equilibrium between 11 to 37 days with the first bank to fail on day four or five. The tests with the available-for-sale securities showed less variability with run equilibrium occurring 19 to 21 days and all first-to-fail on day four
CONFIRMATION AND VALIDATION OF TESS TRANSITING EXOPLANETS WITH THE ISHELL RADIAL VELOCITIES SPECTROMETER
Get PDFThe Precise Radial Velocity (PRV) technique is one of the most effective techniquesused to search for exoplanets and characterize them. We use the iSHELL spectrograph on NASA Infrared Telescope Facility (IRTF) to obtain follow-up PRV observations to continue monitoring and confirming Transiting Exoplanet Survey Satellite (TESS ) mission candidate planets orbiting cool and low mass stars. Observing cooler and lower-mass stars is unique because the reflex motion from orbiting exoplanets is larger, the transit depths are deeper, and the Habitable Zone (HZ) orbital periods are shorter. The importance of PRV follow-up observations is to confirm the orbital properties of the planets, determine their masses and bulk densities, look for evidence of additional planets, and constrain their eccentricities. My motivation for this study is to help NASA achieve its Level 1 goal. TESS was launched in Spring 2018 with a Level 1 mission requirement to search for small planets around nearby and bright stars with the goals of measuring the masses of 50 planets that are smaller than 4 Earth Radii (RL). While this may be sufficient to meet the primary TESS goal of determining the masses of 50 planets, it is not utilizing TESS to its full extent where many interesting targets will be left out. Therefore, ground-based PRVs help advance NASA’s Level 1 requirement. Our ground-based observations supplied us with the transits of the planets around their host stars which, in turn, provided an estimate of the orbital period and phase of a planet through which we can constrain the planetary mass. Thus, a combination of our PRVs and photometric transits will provide the radii and masses that drive the densities, which is the main scientific return needed for stars with planetary transits. In this thesis we confirm the validation of the TESS transiting exoplanet candidates with the iSHELL spectrograph at the NASA Infrared Telescope Facility (IRTF). We use a modified GP kernel to simulate and constrain the stellar activity in the stars and validate the discovery of a two-planet transiting system orbiting the TESS object of interest, (TOI) 560. We obtain follow-up spectroscopy and corresponding PRVs with iSHELL and the HIRES Spectrograph at Keck Observatory to validate the planetary nature of these signals, which we combine with the published Planet Finder Spectrograph (PFS) RVs from the Magellan Observatory. We estimate the age of the star and detect the masses of both planets. Next, we apply an optimization to our forward model technique that generates our PRVs. Optimization of the forward model is a unique approach and has not done before, where we deep dive to analyze each of its parameters to obtain the best Line Spread Function (LSF) model that fits our data and then we apply different techniques to produce PRVs. We selected TOI 461 as a model star to apply these optimizations because TOI 461 is a young, bright star and so is an excellent target for detailed characterization studies with JWST to constrain its composition and test theories of planet formation and evolution. We found that the deep telluric lines in our data tend to be poorly fit and have some residuals which causes a bias in our RVs. These RVs biases are called systematics. We applied new template processes to smooth the residuals of our forward model and change their deshifting. These new template processes helped improve the systematics and stabilized them. We implemented a new technique of clipping the saturated flux lines in our data because they motivate the systematics to appear. This clipping technique helped in mitigating the systematics and improved our PRVs. We trimmed the data at the edges of the spectral orders where data is noisy due to the lower signal-to-noise (SNR) ratio. We utilized a chunking approach to the spectrum by chopping it up into pieces and modeled each piece independently to test the behavior of the systematics in each part of iSHELL spectrograph. The trimming and chunking approaches both confirmed a variability in the distribution of our Line Spread Function (LSF), and so our LSF model is not constant across the star spectrum. We need a chromatic LSF that model the LSF width as a Gaussian changing slowly over the spectrum
Hardware Support for Machine Learning Security and Privacy
No full textMachine Learning (ML) security and privacy is a pressing issue. While ML is being increasingly deployed in diverse sectors due to its promising performance in regular scenario, ML models are found vulnerable to adversarial attacks—security attack and privacy attack. Security attack exploits ML vulnerability in order to confuse models and eventually leads to misclassification. In privacy attack, ML models are found to leak sensitive information of their private training data. However, insuring data privacy in these systems, especially protecting training data from any leakage is crucial towards building trustworthy ML systems. With the rise of security and privacy attacks, researchers have proposed various defense techniques to make ML models robust. However, existing defenses against security attack show various drawbacks in terms of training multiple models, requiring extra model data, retraining or fine tuning model, long-latency inference due to additional computation, etc. These limitations pose performance and power consumption overhead, challenging practical deployment of the solutions in resource- and power-constrained devices such as embedded systems, mobile devices, IoT, smart watch, and edge devices, etc. Similarly, prior defenses against privacy attacks are either extremely compute-intensive because of using additional models or drastically drop the model utility (i.e., classification accuracy) due to adding noise to gradient and outputs. Now, our goal is to come up with low-overhead defense solution through hardware support. As such, this dissertation pursues three research directions that demonstrate how hardware support can boost ML security and privacy. In the first direction, we explore ML security vulnerability for malware detection and overcome them using hardware support. Novel malware and other security exploits are continuous threat for computing systems. With the rise in variety and complexity of new attacks, software-based malware detection is also growing complex, posing increasing challenge to protect systems in real time. The drawback is overcome with the emergence of hardware-based malware detector (HMDs), which are ML classifiers built with low-level features collected from hardware. HMDs can be always on and detect malware and otherattacks in real time. However, our research shows that current generation of HMDs can be reverse engineered and subsequently be bypassed by evasive malware, which severely undermines HMDs security. Next, we identify the ML vulnerability exploited by evasion attacks and developed evasion-resilient HMDs in three unique approaches as follows: Monotonic-HMDs, Stochastic-HMDs, and Non-differentiable HMDs (ND-HMDs). i) Monotonic-HMDs: In our experimental findings, regular malware can be detected using HMDs, whereas evasive malware can bypass the model by perturbing most-negative weighted features. In fact, evasive malware appears as benign application, while continuing their malicious activity, by increasing most-negative weighted features. We address this ML vulnerability by building HMDs using only the positive weighted features, which can detect both regular malware and evasive malware significantly. We call HMDs with positive-weighted features as Monotonic-HMDs since they impose increasing monotonicity constraints to model, which means if the attackers try to increase any of the malware features, they are more likely to be detected as malware. ii) Stochastic-HMDs: We experiment with state-of-the-art evasion-resilient HMDs (called RHMDs), which can confidently detect evasive malware but pose substantial overhead (area, power, latency) because of comprising multiple diverse detectors and randomly switching between them. We overcome the drawback by proposing Stochastic-HMDs, which is a sin-gle detector that basically injects stochastic faults to HMDs computation; it makes the Stochastic-HMDs a moving target defence, achieving the same goal of RHMDs with much less overhead (time, power, and area). iii) Non-differentiable HMDs (ND-HMDs): Our experiment demonstrate that HMDs can detect micro-architectural attacks such as spectre, meltdown, foreshadow, and zombieload. However, evasive version of these attacks can hide from HMDs detection. We observe that evasive attacks succeed more on HMDs classifiers that have continuous, differentiabledecision boundary. We overcome this vulnerability by building HMDs using classification algorithms that use discrete, non-differentiable decision boundary, which can detect evasive as well as regular micro-architectural attacks. In the second direction, we explore hardware support to overcome deep learning (DL) vulnerability against security attack such as adversarial image samples. While DL models can accurately classify clean images, they get confused and predict incorrect labels for adversarial images, posing security threat for image- and vision-based applications. Adversarial samples are generated by inserting visually imperceptible noise to clean images so that they cross model’s decision boundary and get misclassified. Here, we propose to randomizes the inference computation by lowering the CPU supply voltage below the nominal voltage, which randomly violates critical path and induce faults. In effect, this unpredictably moves target model decision boundary, which hardens successful adversarial sample generation and reduces the misclassification rate. The proposed defense does not require any software/hardware modifications and offers a by-product reduction in power consumption due to supply voltage reduction. In the third direction, ML systems are susceptible to privacy attacks, such as membership inference attacks (MIAs), which leak private information from the training data. Specifically, MIAs are able to infer whether a target sample has been used in the training data of a given model. Such privacy breaching concern motivated several defenses against MIAs. However, most of the state-of-the-art defenses such as Differential Privacy (DP) come at the cost of lower utility (i.e., classification accuracy). In this work, we propose Privacy Preserving Volt (VP P ), a new lightweight inference-time approach that leverages undervolting for privacy-preserving ML. Unlike related work, VP P maintains protected models’ utility without requiring re-training. The key insight of our method is to blur the MIA differential analysis outcome by comprehensively garbling the model features using random noise. Unlike DP, which injects noise within the gradient at training time, VP P injects computationalrandomness in a set of layers’ during inference through carefully designed undervolting. Specifically, we propose a bi-objective optimization approach to identify the noise characteristics that yield privacy-preserving properties while maintaining the protected model’s utility. Extensive experimental results demonstrate that VP P yields a significantly more interesting utility/privacy tradeoff compared to prior defenses. For example, with comparable privacy protection on CIFAR-10 benchmark, VP P improves the utility by 32.93% over DP-SGD. Besides, while related noise-based defenses are defeated by label-only attacks, VP P shows high resilience to such adaptive MIA. Moreover, VP P comes with a by-product inference power gain of up to 61%. Finally, for a comprehensive analysis, we propose a new adaptive attacks that operate on the expectation over the stochastic model behavior. We believe that VP P represents a significant step towards practical privacy preserving techniques and considerably improves the state-of-the-art
THE CHARACTERIZATION OF EARLY AND LATE EXTRACELLULAR VESICLES RELEASED FROM HIV-1-INFECTED CELLS AND THEIR EFFECT ON THE RECIPIENT CELLS.
No full textThis work is embargoed by the author and will not be publicly available until May 2029.Human immunodeficiency virus type 1 (HIV-1) is a causative agent of acquired immunodeficiency syndrome (AIDS) and, since its discovery in 1981, it has caused approximately 40.1 million deaths worldwide. By 2022, it was estimated that about 1.3 million people were infected alone and about 40 million people lived with it. The implementation of combination antiretroviral therapy (cART) has drastically reduced morbidity and mortality in HIV-1 patients. Modern cART drugs effectively target stages of HIV-1 cycle such as viral entry, reverse transcription, integration, protease cleavage of viral polyproteins, and virion maturation. However, life-long adherence to a cART is required to suppress viral replication to the safe level. We have recently shown that Extracellular Vesicles (EVs) from HIV-1-infected cells play major role in viral pathogenesis and can facilitate viral spread. In Chapter Two of the dissertation, we have attempted to address the timing of EVs and virions release from HIV-1-infected cells. Briefly, uninfected, HIV-1- and HTLV-1-infected cells were synchronized in G0 of cell cycle and then released in serum-rich media with the presence of the inducers - phytohaemagglutinin and Interleukin 2 (PHA/IL-2) to induce viral gene expression and resume normal cell cycle. The supernatants and cell samples were collected post-induction at different time points and tested for the markers of EVs (tetraspanins), autophagy and for viral proteins and RNAs. Tetraspanins and autophagy related proteins were found to be differentially secreted in HIV-1- and HTLV-1-infected T-cells. Viral proteins and RNAs were present at 6 hrs and their production increased up to 24 hrs. Finally, HIV-1 supernatant from 6-hrs sample was found not to be infectious, however, the virus from 24-hrs sample was successfully rescued and infectious. In Chapter Three, we analyzed biochemical and functional properties of HIV-1 EVs, larger than the currently accepted size range for HIV-1. First, we isolated five different fractions (2K through 167K (L)) from HIV-1-infected T-cells by differential ultracentrifugation. All fractions showed a heterogeneous presence of viral proteins and RNAs and certain fractions showed the increase of virus-related cargo overtime and large EVs (2K) showed increased presence of viral proteins and RNAs. Large EV pellet (2K) was further purified by size-exclusion chromatography and EV-rich fractions 1-5 of 2K pellet from HIV-1-infected T-cells showed the presence of viral proteins, RNAs and amphisome markers. Next, fractions 1-5 were tested on infectivity on naïve recipient cells. Fraction #2 after SEC showed the highest replication in the recipient cells and the results were replicated in HIV-1-infected primary T-cells. Immunoprecipitation with the use of antibodies against amphisome markers showed the presence of viral RNAs and proteins. Overall, our data shows that EVs with virus-related cargo are released before the virus from infected cells, thereby implicating a potentially significant effect on uninfected recipient cells prior to subsequent viral infection. Moreover, large EVs from HIV-1-infected T-cells, produced at 24 hrs post-release, contain infectious material and have amphisome markers on their surface.2029-05-1
Perceptions of Culturally Responsive Assistant Principals Serving English Language Learners
Get PDFAbstract PERCEPTIONS OF CULTURALLY RESPONSIVE ASSISTANT PRINCIPALS SERVING ENGLISH LANGUAGE LEARNERS Miguel Antonio Chacón, Ph.D. George Mason University, 2024 Dissertation Director: Dr. Samantha Viano It is essential to prepare public school leaders to be culturally aware and equipped to educate the influx of English language learners (ELLs). Some school leaders are not culturally responsive (Santamaría, 2014). The goal of this concurrent mixed methods study was to learn how assistant principals (APs) perceived their preparation regarding culturally responsive school leadership (CRSL) for the benefit of ELLs and also how the APs implemented CRSL practices to promote ELLs’ academic achievement and socioemotional learning. The sample consisted of 15 APs from kindergarten–12th-grade public schools who completed a demographic survey and participated in semistructured interviews. The findings, interpreted using Khalifa’s (2020) CRSL framework, were as follows. First, APs felt prepared to lead schools with large numbers of ELLs despite lacking preparation. Second, APs regularly practiced critical self-reflection and demonstrated commitment to continuous learning. Third, APs developed and sustained culturally responsive teachers through professional development, building teacher awareness, and capacity to use data. Fourth, APs promoted inclusive antioppressive school contexts through direct teacher feedback, professional development, and acknowledgement of cultural and social capital. Fifth, APs engaged students’ indigenous community contexts through parental engagement, cultural events, and recognition of community partners. The implications are that school leadership programs need to more intentionally prepare APs to (a) employ CRSL practices that help ELLs and others succeed academically and socially, (b) self-reflect on leadership practices that promote social justice for ELLs, (c) develop and sustain culturally responsive teachers, (d) promote inclusive antioppressive school contexts, and (e) engage students’ indigenous community contexts