62588 research outputs found
Sort by
Standardizing Blockchain Layer 2 Benchmarking
No full textAs blockchain adoption continues to grow, developers and businesses face an ever-expanding ecosystem of platforms, each offering unique trade-offs. For developers, efficient blockchain performance enables faster transaction times, lower fees, and a more responsive user experience. For businesses, the choice of blockchain platform directly impacts scalability, cost-effectiveness, and the ability to meet user demands. However, the lack of standardized benchmarking tools has made it difficult to objectively assess and compare the performance of blockchain platforms. Performance benchmarking is thus crucial to determining the suitability of a blockchain for specific use cases and applications.To address this pressing need, the Blockchain Benchmarking Standardized Framework (BBSF) was developed to provide a consistent methodology for evaluating blockchain performance. The framework defines standardized workloads, metrics, benchmarking drivers, and reporting formats, offering a foundation for transparent and reproducible comparisons. Previous work on the BBSF focused on benchmarking layer 1 (L1) blockchains through the Blockbench-v3 implementation, using Web3-style workloads. These efforts provided valuable insights into the comparative performance of different L1 platforms, uncovering strengths and limitations that have practical implications for developers and enterprises alike.As the blockchain ecosystem has evolved, the emergence of layer 2 (L2) solutions has introduced new opportunities—and challenges—for scaling and cost efficiency. L2 platforms aim to address the limitations of L1 systems, such as transaction throughput bottlenecks and high fees, but their varied architectures and design choices present new complexities. These include differences in finality assumptions, decentralization models, and interactions with L1 systems. While the foundational principles of the BBSF remain applicable, directly applying an L1-oriented framework like Blockbench-v3 to L2 systems fails to capture critical performance characteristics unique to L2 architectures.This thesis presents a comprehensive study of blockchain benchmarking, extending the BBSF to encompass the complexities of L2 solutions. It evaluates a new standardized benchmarking framework tailored to L2 systems, addressing their unique properties and challenges. Empirical testing on prominent L2 platforms, including zkSync and additional candidates, highlights the framework’s effectiveness and provides actionable insights into their performance. Furthermore, this thesis explores benchmarking results for L1 platforms, such as Sui, to provide a comparative foundation for analyzing L1 and L2 performance. By integrating theoretical advancements with practical experimentation, this work seeks to establish a robust and adaptable approach to benchmarking that can guide developers, researchers, and enterprises in making informed decisions.</p
Learning from Implicit Feedback for Unbiased Learning to Rank
No full textSearch engines serve as one of the most important tools for accessing information online. In modern search engines, learning to rank~(LTR) algorithms play a critical role by creating models to accurately order a list of candidate documents based on their relevance to the query. Though the logs of the use of a search engine provide sufficient data to train a better ranker, it is well known that such implicit feedback reflects biases. Therefore, LTR with implicit feedback such as clicks is an important research question in the IR community.In this dissertation, we consider how to achieve genuine unbiasedness in learning to rank with implicit feedback.
We start by providing an end-to-end description of a ranking service following the lifecycle of workflow, including the feature representation learning, ranking model optimization, and dynamic update. To achieve the unbiasedness in each step, we propose three methods to address the bias within implicit feedback, respectively. In particular, we first study an under-appreciated bias -- click attraction bias -- and unbiased semantic matching, where the learned feature between the query and document via implicit feedback is immune to the click attraction bias. We then study the high variance problem of existing inverse propensity weighting methods, and propose a novel model-based unbiased learning to rank framework, which combines a user simulator to generate pseudo clicks and doubly-robust learning to obtain a ranking model with low bias and low variance. At last, we identify the propensity overestimation phenomenon of applying automatic unbiased learning to rank methods. We propose a novel logging-policy-aware propensity model and its distinct two-step optimization strategy which solves the difficulty of backdoor adjustment in unbiased learning to rank.Extensive experiments across multiple tasks demonstrate the superiority of our proposed methods.
</p
Orchestrating Coding and Learning for Reliable and Secure Neural Network Processing
No full textError correcting output codes (ECOCs) have been proposed to improve the robustness of deep neural networks (DNNs) against hardware defects of DNN hardware accelerators.Unfortunately, existing efforts suffer from drawbacks that would greatly impact their practicality: 1) limited effectiveness due to error propagation and accumulation when DNNs are deep; 2) robust accuracy (with defects) improvement at the cost of degraded clean accuracy (without defects); and 3) absence of theoretical foundations that can elucidate the relationship between codeword design, weight-error magnitude, and network characteristics, so as to provide robustness guarantees. In this dissertation, we bridge the gap and tackle the aformentioned problems in three works.
The first work aims to denoise in the early layers of DNNs to diminish the significance of error propagation and accumulation on memristive DNN accelerators. Specifically, we propose a minimum mean square error (MMSE) based method to compensate the weight variations at each layer without extra hardware costs. What\u27s more, we propose a weights-to-crossbar mapping scheme by inverting bits to mitigate the impact of stuck-at-faults (SAFs). Additionally, we propose to use L1 regularization to increase the network sparsity, as a greater sparsity not only further enhances the effectiveness of the proposed bit inversion scheme, but also facilitates other early denoising mechanisms. Experimental results show that our schemes can achieve 40%--78% accuracy improvement under different tasks and DNNs.
In the second work of this dissertation, we first identify the root cause of ECOCs\u27 degraded clean accuracy is error correlation, and then propose a novel comprehensive error decorrelation framework, namely COLA. Specifically, we propose to reduce inner layer feature error correlation by adopting a separated architecture, where the last portions of the paths to all output nodes are separated, and orthogonalizing weights in common DNN layers so that the intermediate features are orthogonal with each other. We also propose a regularization technique based on total correlation to mitigate overall error correlation at the outputs. The effectiveness of COLA is analyzed theoretically, and evaluated experimentally, e.g., up to 6.7% clean accuracy improvement compared with the original DNNs and up to 40% robust accuracy improvement compared to the state-of-the-art ECOC-enhanced DNNs.
The third work of this dissertation is a fundamental analysis of ECOC through the lens of neural tangent kernels (NTKs). We found that utilizing one-hot code and non-one-hot ECOC is akin to altering decoding metrics from l_2 distance to Mahalanobis distance in clean models, which are defined as those free of weight errors. A distance threshold exists between clean models and non-clean models such that if the distance between a clean output and its nearest codewords is smaller than this threshold, then the DNN can make predictions as if it is free of weight-errors. The threshold is determined by the normalized distance among codewords, the DNN architecture, and the scale of weight-errors. Based on these findings, we further demonstrate how to practically use them to identify optimal ECOCs for simple tasks, which have small number of classes, and complex tasks, which have large number of classes, by balancing the code orthogonality and code distance. Extensive experimental results across four datasets and four DNN models validate the superior performance of constructed codes, guided by our findings, compared to existing ECOCs.
The drawbacks of ECOCs mentioned in the first paragraph are addressed through the techniques proposed in this dissertation. Specifically, the early denoising techniques introduced in the first study effectively correct errors in the initial layers, significantly mitigating error propagation and accumulation. By reducing error accumulation, the robust accuracy of ECOCs is substantially improved. To address the second drawback, we identify the root cause of clean accuracy degradation—error correlation—and propose the COLA framework to decorrelate errors. This approach enhances both clean and robust accuracy. Furthermore, the third contribution provides a theoretical foundation that offers guidelines to the research community for further improving ECOC. Together, the three works proposed in this dissertation synergistically enhance each other to develop a practical and robust DNN with ECOC on memristive devices.</p
Efficacy of Organizational Skills Training for Students with Emotional and Behavioral Disorders
No full textOrganizational skills interventions teach strategies for materials organization, time management, and planning and are effective in reducing organizational skills deficits among students with attention-deficit/hyperactivity disorder (ADHD; Bikic et al., 2017; Langberg et al., 2008). Thus far, research has focused primarily on students within general education settings and without significant co-occurring diagnoses. Although many students with emotional and behavioral disorders (EBD) have significant organizational skills deficits (Coleman, 2012; Eusebio, 2010), sample characteristics of randomized control trials limit generalizability of organizational skills interventions to these students. The purpose of this study was to evaluate the implementation and efficacy of Organizational Skills Training (OST; Gallagher et al., 2014) for students with EBD within a self-contained special education setting. First, four special education teachers within this setting completed a survey regarding the necessity of possible modifications to OST for implementation with their students. The intervention was then adapted accordingly by focusing solely on materials management skills, shortening intervention sessions, incorporating instruction on coping strategies and conflict management, and replacing OST behavior management and reinforcement systems with existing school procedures. Second, a multiple baseline single subject research design evaluated the efficacy of the adapted OST intervention in improving materials management skills for two middle school students with EBD. The single subject analysis (comprised of visual analysis, percentage of non-overlapping data, and Tau-U) demonstrated significant improvements in materials management skills for both students. Teachers and students reported high acceptability for OST. As such, it is both worthwhile and effective to implement organizational skills interventions for students with EBD within self-contained special education settings. </p
Effects of Drug Control Interventions on Human Capital, Social Behavior, and Labor Market Outcomes
No full text{"value":"This dissertation focuses on drug control interventions and explores their impacts on human capital development, social behaviors, and labor market outcomes. The first chapter pertains to the Drug-Free Communities (DFC) Support Program, which aims to mitigate substance use among youth. This study examines the impact of the DFC program on youth development outcomes in the United States from 2008 to 2019. Using a difference-in-differences (DID) model that leverages the staggered allocation of grants to communities, we document that the DFC program significantly reduces drug-related criminal activities and improves academic performance among juveniles. Our cost-benefit analysis unveils the substantial social welfare gain of this program. Potential mechanisms encompass youth behavioral responses, such as reductions in marijuana use and fewer opioid-related inpatient stays, and community-level changes in the drug misuse environment, including reductions in retail opioid shipments and drug-related mortality. To the best of our knowledge, this study is the first to causally identify the impacts of the DFC program on community well-being. Our comprehensive evaluation underscores the importance of collective action within communities in combating substance use, particularly in light of the recent drug crisis.
The second chapter examines drug courts, which provide eligible individuals with drug treatment as an alternative to incarceration and aim to reduce recidivism. However, their impact on overall drug crime has not been extensively studied. In this study, we leverage the staggered implementation of county-level Adult Drug Courts (ADCs) from 2001 to 2012 and use the difference-in-differences approach that accounts for treatment heterogeneity to examine their effectiveness in curbing drug crime. Innovatively, our study considers the order of commitment and focuses on its impacts on first-time drug offenses and recidivism. We explore these effects both through a theoretical framework and by empirically testing them. Our primary results show that the implementation of ADCs significantly reduced the one-year, two-year, and three-year recidivism of drug offenses by 36.00%, 25.10%, and 28.44%, respectively, indicating positive specific deterrence effects. Moreover, we find that ADCs significantly increased the first drug arrest by 20.79%, indicating a weakened general deterrence effect. Aggregating these two effects across different data sets yields a net increase in total drug offenses of 13.10% to 15.82%, demonstrating that the weakened general deterrence effect dominates the other. In sum, our study unveils the unintended consequences of leniency in drug crime punishment, suggesting a need for intensifying ADC programs, possibly by adding more phases and enhancing graduation rates for participants.
The third chapter is motivated by the limited research on the relationship between medical CBD laws and the labor market. In this paper, we use the Current Population Survey Outgoing Rotation Group data and examine the effects of CBD legalization on the labor market outcomes of the working-age labor force. To address the concern of low exposure to this policy for the general population, we use the linear probability model to predict the probability of marijuana use and create high-probability and high-recall groups. Using the stacked difference-in-differences approach, we consistently find a reduction in employment and no effects on hours worked and hourly wage for these two groups. Potential mechanisms include an increase in arrests for adult cannabis offenses and an increase in the number of Supplemental Security Income (SSI) recipients due to disability.
","attr0":"abstract"
On the Word-Representability of Chordal Near-Triangulations
No full textIn this dissertation we study the word-representability of a class of graphs called (plane) chordal near-triangulations. In general, we say a graph G = (V, E) is word-representable if there exists a word, w, over the alphabet V(G) such that letters x and y alternate in w if and only if xy ∈ E(G). Word-representable graphs were first introduced in the literature in 2008 by Kitaev and Seif to study a problem in Semi-group theory. In that same year, Kitaev and Pyatkin published a first systematic study of word-representable graphs. Since its introduction, the theory of word-representable graphs has been an active area of research, and has even resulted in a book, “Words and Graphs,” published in 2015, dedicated to its study. Therehave also been variations of word-representability investigated in the literature, one such variation will be discussed in this dissertation.In this dissertation, we investigate a class of graphs called chordal near- triangulations. In 2016, Glen characterized the word-representable K4-free near-triangulations in terms of 3-colorability. The introduction of an induced K4, however, ensures that every such near-triangulation is necessarily 4-colorable. It was shown by Halldo´rsson, Kitaev, and Pyatkin, in 2011 that a graph is word-representable if and only if it admits a semi-transitive orientation. Using an approach via semi-transitive orientations, in Chapters 2 and 3, we characterize the word-representable chordal near-triangulations in terms of forbidden induced subgraphs. This characterization will be a stepping stone to the bigger problem of characterizing all word-representable near-triangulations, in hopes to eventually characterize to all word-representable planar graphs.In Chapter 4, we introduce a variation of word-representability called 12- representability. While this variation is not studied as extensively in the literature as classical word-representability, we outline several important results before investigating the 12-representabiity of Apollonian Networks, which are the chordal triangulations. We determine that for this class of graphs 12-representability is equivalent to being a comparability graph, which is not true in general.</p
The transport of free-living amoeba and their pathogenic endosymbionts through drinking water treatment plants and distribution systems
No full textFree-living amoebae (FLA) are ubiquitous microorganisms that have been found throughout both natural and man-made environments. Some FLA are human pathogens, known to cause fatal infections, but they can also harbor certain pathogenic bacteria, allowing the bacteria to survive and grow. Research has shown many species of FLA can survive and bypass typical water treatment processes, transporting internalized pathogenic bacteria with them. Therefore, presence of FLA in drinking water systems, along with the pathogens they harbor, is an area of increasing universal concern.Research has shown that Acanthamoeba sp. and Vermamoeba vermiformis are the most frequently recovered FLA in treated drinking water networks, followed by Naegleria sp. These FLA show a high resistance to chlorination, especially when in the cyst stage. This thesis investigated the presence of these three most common FLA, and the known endosymbiotic pathogenic bacteria Stenotrophomonas maltophilia, in two drinking water delivery networks in Eastern Pennsylvania. The results of the two-year study showed 74% (56/76) of all raw water samples and 22% (15/69) and 14% (12/83) of finished water samples from the two utilities, respectively, were positive for FLA by microscopy. Endosymbiotic S. maltophilia was also recovered at various locations throughout both water treatment plants and distribution systems, demonstrating that FLA can serve as vectors that transport bacteria through conventional water treatment processes.
Additionally, a fan-shaped FLA which was confirmed as Vannella sp., was observed in numerous samples throughout the water treatment process. This demonstrates that Vannella sp. is another FLA that can bypass water treatment processes, including chlorination. Results indicate that the Vannella sp. detected in this study harbored the opportunistic pathogen S. maltophilia, suggesting the potential for Vannella sp. to transport endosymbiotic pathogens through water treatment plants and into finished drinking water supplies. A detailed co-culture protocol was developed to test the ability for Vannella sp. to harbor and protect S. maltophilia from harsh environmental conditions such as disinfection. </p
Factors Influencing Disability-Inclusive Emergency Operations Plans and Perceived Competence for Implementation
No full textExisting literature lacks comprehensive research on emergency operations planning (EOP) for students with disabilities in K-12 public schools. This study aimed to address this need by investigating the presence of current EOPs and identifying factors influencing their implementation, focusing on individualized emergency plans for students with disabilities. A survey was used to gather information from in-service special educators, special education administrators, and building principals regarding school safety practices, personal experiences, available resources, and perceived competence in supporting students with disabilities during emergencies. The research questions explored various aspects of emergency planning, including building-specific organizational factors, visible security measures, training and drill practices, access to appropriate supplies, and individual factors influencing perceived competence. Data were analyzed using four ordinary least squares multiple regression models. Findings indicate a significant relationship between the percentage of special education enrollment and the percentage of individualized emergency operations plans (IEOPs) (p < .001). Reported building type was also a predictor of the percentage of IEOPs in a given school. Factors influencing perceptions of competence included awareness of the impact disabilities may have on EOP participation (p <.001), frequency of interactions with students with disabilities (p = .036), and high level of access to specialized resources to support students during emergency situations (p <.001). Common themes in reported barriers to effectively planning for students with disabilities included lack of awareness of need, staff shortages, building accessibility, and access to relevant information. Common themes in reported facilitators to IEOP included time for collaborative planning with relevant stakeholders in and out of the school, knowledge of student needs, and adequate staffing to meet student needs. </p
Optimization via Clustering to Improve Ensemble Performance of Infectious Disease Models
No full textThis thesis discusses the implementation and optimization of two ensemble learning models. The first project presents a Cluster-Aggregate-Pool (CAP) approach to improve seasonal influenza ensemble forecasting. Several implementations of the CAP method are presented and I argue that a CAP ensemble better satisfies the assumption of independence among component forecasts, restores diversity, and effectively handles situations where a single component model fails to provide a submission. Findings in this thesis confirm that a CAP ensemble, compared to a non-CAP ensemble, shows improved calibration and performance across various metrics.The second project focuses on enhancing the performance of traditional statistical models when dealing with imbalanced data. This thesis develops a divide-and-conquer framework and demonstrates that, with an appropriate dataset partition, the performance of pooled subset models is at least as good as that of models fitted on the entire dataset. The proposed method shows improvements in performance metrics compared to traditional statistical models and performs comparably to boosting or bagging models, but with reduced complexity. A key advantage of the divide-and-conquer approach is its ability to report the partition results while preserving the interpretability of feature contributions to the target label.</p
Combining Atomic Force Microscopy with Infrared Lasers for High-Resolution Surface Analysis
No full textWhile a standard microscope provides detailed images of a sample\u27s shape, an infrared (IR) microscope visualizes the distribution of chemicals within the sample. This label-free method identifies molecules and materials through their characteristic vibrational absorptions. However, its application is restricted by Abbe\u27s diffraction limit, which confines the resolution of small features to approximately half the wavelength of incident light. This limitation brings a challenge for analyzing nanomaterials and biological specimens with spatial features smaller than 100 nm. To overcome this, combining atomic force microscopy (AFM) with IR radiation has emerged as a powerful technique, enabling nanoscale imaging. My research focuses on a specific branch of AFM-IR, known as peak force infrared (PFIR) microscopy, which was invented by our group in 2017. From one scan, PFIR simultaneously captures the IR image under a specific wavenumber, topography and mechanical channels including adhesion, modulus and deformation under a spatial resolution of ~6 nm. However, as a scanning-probe method, traditional AFM-based IR produces only one image per scan, making the comparative study with multiple images from the same sample area time-consuming. This issue is exacerbated by frame drift and distortion from consecutive scans. To tackle these challenges, we developed a Dual-Color peak force infrared (PFIR) microscopy technique that achieves simultaneous imaging of two IR frequencies with a single AFM scan. This method not only improves the data acquisition efficiency of AFM-based IR, but also enables simultaneous nondestructive chemical nanoimaging of multiple components, allowing for pixel-by-pixel comparison.
Moving from microscopy with narrowband laser source to spectroscopy with broadband laser source, we have developed AFM-based one-dimensional and two-dimensional (2D) nano-spectroscopies. Inspired by Fourier transform infrared (FTIR) spectroscopy, we developed the Fourier transform PFIR nano-spectroscopy. Through a Fourier transform, we demonstrated it is possible to retrieve IR absorption spectra from the photothermal detection of broadband IR absorption by an AFM tip on a polymer mixture and hexagonal boron nitride. Our work revealed the feasibility of time-domain detection of the AFM-IR signal in the mid-IR regime and paves the way toward multiphoton vibrational spectroscopy below the diffraction limit.
Further, we developed time domain AFM-based 2DIR nano-spectroscopy for the first time. This technology integrates the current research horizon of 2DIR spectroscopy with AFM-IR, allowing rich spectroscopic data to be gathered locally on a heterogeneous sample surface under nanoscale spatial resolution. The AFM-based 2DIR allows for in situ examinations of vibrational anharmonicity, coupling, and energy transfers in materials and nanostructures, making it particularly effective for understanding the relaxation dynamics at the IR range in 2D materials.
This Dissertation is organized as follows:
Chapter 1 introduces the fundamental principles of AFM-based IR microscopy, with a particular emphasis on PFIR microscopy. This chapter discusses the methodology and its capability to explore IR information at the nanoscale.
Chapter 2 contains the development and application of the Dual-Color PFIR microscopy to simultaneously scan two IR images under different wavenumbers, providing a more comprehensive understanding of nanoscale IR properties.
Chapter 3 details the development and application of the FT-PFIR nano-spectroscopy. Inspired by Fourier Transform Infrared (FTIR) spectroscopy, this method integrates a broadband mid-IR laser source with AFM operated under peak force tapping mode for the first time. Compared with sweeping a range of wavenumbers with narrowband laser sources for point spectrum collection, FT-PFIR retrieves point spectra from FFT of the collected interferogram.
Chapter 4 contains the development and demonstration of AFM-based 2DIR nano-spectroscopy. This chapter discusses the powerful advanced spectroscopy, 2DIR, for vibrational interactions and structural studies and highlights the innovative combination of AFM and 2DIR techniques to achieve nanoscale spectroscopic analysis.
</p