19304 research outputs found
Sort by
Towards Fast and Robust Emerging Neural Networks
No full textNeural networks(NNs) are foundational to next-generation artificial intelligence, driving revolutionary advancements across diverse fields such as computer vision, healthcare, finance, and autonomous driving. However, the widespread adoption of these models introduces several significant challenges. Real-time applications, in particular, demand stringent adherence to execution delays and power consumption constraints, without compromising reliability. Furthermore, the increased deployment of neural networks has attracted malicious users, resulting in threats such as model disruption and theft of sensitive training data. This dissertation addresses these critical challenges within emerging neural network paradigms in computer vision, specifically focusing on quantum neural networks (QNNs), object detection models, and vision transformers. NNs have become integral to computer vision, powering key applications like facial recognition, image classification, and autonomous driving. The drive for higher accuracy and efficiency has spurred a steady flow of new, advanced models, each pushing the boundaries of what’s possible in these fields. Among the latest innovations are quantum neural networks (QNNs), which leverage rapid advancements in quantum computing—a new computational paradigm that uses principles of quantum mechanics to solve complex problems at remarkable speeds. These systems can handle tasks that traditional computers would find nearly impossible to complete within a practical timeframe. Early results with QNNs on current quantum hardware are promising, showing that they could potentially outperform conventional neural networks in both speed and computational power. Simultaneously, conventional neural network architectures have made substantial strides in performance. Convolutional neural networks (CNNs), especially the YOLO models, remain the preferred choice for object detection tasks due to their high performance and accuracy. Recently, transformer-based neural networks have gained traction for their ability to excel in both natural language and image processing through the self-attention mechanism, which captures global features of the images effectively, thereby producing superior results compared to the CNNs. This dissertation investigates the performance limitations and security vulnerabilities inherent in these emerging neural network models. We propose methodologies to enhance their robustness, efficiency, and resilience to attacks, with particular emphasis on computer vision applications. Additionally, we outline potential directions for future research to further improve the reliability and effectiveness of neural networks in real-world applications
Sergei Rachmaninoff’s Piano Concerto No.4, Op.40 – A Historical Survey and Motivic Analysis of the Third Movement
No full textRachmaninoff’s Fourth Concerto is an anomaly in his output. It is a mystery whether the work was largely produced in Russia or America; therefore it is hard to make a claim as to which musical and extramusical influences pre- and post-emigration contributed to the aesthetic. In addition, the composer twice revised it, to make it more accessible to audiences. Despite these revisions, it remains the least accessible of Rachmaninoff’s five works for piano and orchestra. This study will be in two chapters. Chapter one is a historical review of the concerto. It traces the origins of the work, specifically the genesis of the concerto pre-1917 and the completion of the work after Rachmaninoff’s immigration to the United States. It also outlines Rachmaninoff’s process of revision. Of particular interest here is that despite two revisions of the work, Rachmaninoff seems never to have questioned that the distinct contrast in style between the Fourth Concerto and its predecessors may have contributed to public dissatisfaction with the work. The question of what is so different about the Fourth is taken up in chapter two, which focuses on its motivic density and rhythmic complexity. Looking specifically at the finale, which is perhaps the least accessible of the three movements, through a motivic analysis, I demonstrate that the movement is highly cohesive, with all motivic content derived from the introduction, primary theme area and transition. From the development onwards I provide a report of Rachmaninoff’s ingenious treatment of melodic and rhythmic motives. Finally, I ask the question, what does a motivic analysis reveal beyond a masterful treatment of motives
Racism Persists, But So Do We: Investigating the Intersection of Exclusionary Discipline Rates and the Sense of Belonging of African American Students in a Multicultural Learning Environment
No full textBackground: African American students have been disproportionately subject to exclusionary discipline, which can negatively impact students’ sense of belonging. However, most of this research has occurred in majority-White contexts, not accounting for shifts in American demographics. Purpose/Research Questions: This study’s purpose is to investigate African American students’ sense of belonging in a site that has undergone a substantial shift in demographics and to explore if that sense of belonging is related to African American students’ rate of removal from classrooms. Research questions include: 1. How do exclusionary discipline rates differ among racial groups in the East ISD? 2. To what extent do African American students experience a sense of belonging within a school district numerically dominated by another historically marginalized group? 3. What is the relationship between African American students’ sense of belonging and their discipline rates within East ISD? Methods: This study analyzed an East ISD data set, which included student responses to a sense of belonging survey. Each response included the student’s race and discipline incidents for the school year. ANOVA and Kruskal-Wallis testing were used to compare discipline rates by student race (Research Question 1) and to compare survey responses by student race (Research Question 2). Pearson correlations were used to investigate the relationship between students’ discipline and their survey responses (Research Question 3). Results: East ISD’s African American students were disproportionately disciplined more than their peers and were more likely to report being involved in school activities. They were less likely to feel a part of their school, believe their teachers cared, and feel as though teachers listened. These students surprisingly demonstrated a positive relationship between discipline rate and two areas of sense of belonging: being excited to come to school and believing their teachers cared about them. Conclusion: Despite no longer being in a majority-White setting, African American students are still experiencing disparate outcomes, such as higher discipline rates and lower sense of belonging. However, the surprising finding of a positive relationship between discipline rates and sense of belonging for East ISD’s African American students raise interesting questions on how students of color perceive discipline
Effects of Crystallization Temperature and Film Thickness on Low-Density Polyethylene (LDPE) Crystallization
No full textThe effect of film thickness and crystallization temperature on low-density polyethylene (LDPE) crystallization was investigated in an isothermal crystallization process. LDPE films of different thicknesses were prepared by flowcoating and crystallized at TC, for 30 minutes. The morphologies of the spherulites after crystallization were captured and used to determine the average size of the spherulites. It was observed that the spherulite size of LDPE increased with an increase in Tc. This is because at higher TC, the driving force for nucleation is low, leading to fewer nuclei but allowing more space and time for each spherulite to grow larger. Also, the results indicated that thinner films typically exhibit higher nucleation density and smaller spherulite sizes due to an enhanced interfacial interaction at same TC. In comparison, thicker films tend to exhibit lower nucleation density and larger spherulites, as the interior regions are less affected by surface effects and provide more room for crystal growth. Finally, the variation of spherulite density with crystallization time at Tc of 100 °C was examined, revealing a significant increase in spherulite density during the early stages of crystallization, followed by a plateau in the later stages as the system reached near-complete crystallization. These findings highlight the dynamic transition from rapid nucleation to slower crystal growth over time
Two Truncated Second Main Theorems in Nevanlinna Theory and Their Applications in Defect Relations
No full textWe generalize the concept of a non-integrated defect, first introduced by H. Fujimoto, and derive two generalized non-integrated defect relations. Based on the work of G. Heier and A. Levin, M. Ru, and Y. He, we improve Heier and Levin’s second main theorem for holomorphic maps into a projective variety intersecting arbitrary effective Cartier divisors in general position. This is achieved by incorporating Seshadri constants into a truncated second main theorem, from which we deduce a non-integrated defect relation. In the second part, we improve K. Yamanoi’s second main theorem for holomorphic maps from C to a smooth projective variety intersecting a simple normal crossing divisor. We achieve this by developing a truncated second main theorem with a good truncation level and a precise estimation of the error term. Consequently, we obtain a non-integrated defect relation in this case
Reaction-Diffusion Systems with Dynamic Boundary for Modeling the Spread of Disease Causing Pathogens
No full textReaction-diffusion systems are extensively employed in epidemiology and population dynamics due to their mass preservation. However, they pose some basic mathematical problems. We developed a reaction-diffusion model to describe the spread of a waterborne disease within a spatially distributed host population. The population inhabits an invariant rectangle domain, Ω, in the upper half-plane, with part of its boundary coincident with the real axis. Infection is established when a pathogen enters through a waterway across this boundary. We investigate existence, boundedness and long-term behavior for dynamic boundary problems. We also analyze the impact of spatial heterogeneity and boundary specificity on disease distribution. Numerical simulations are provided to illustrate key theoretical results. We derive new analytical insights regarding hydrodynamic transport and spatial infection
On the State-Dependence of Forecaster Inattention and Progressive Taxation
No full textThis dissertation consists of two essays on macroeconomics that study how allowing certain variables and parameters to be state-dependent creates new empirical results and has ramifications for theoretical modeling. I show that ignoring these state-dependencies can make research conclusions overstate or understate certain policy implications. In the first essay, I provide new estimates of forecaster inattention and information rigidity related to the sticky information model of expectation formation. While most papers use aggregate-level regressions or model calibrations to estimate the amount of information rigidity, I employ a more granular estimation strategy using micro-level data from the US Survey of Professional Forecasters (SPF). I provide evidence that the true amount of forecaster inattention is much smaller than previously thought. I also document novel state-dependence and time series facts. My results imply that some other, stronger source of information rigidity must exist to account for the discrepancy between the aggregate- and micro-level results. A model accounting for information rigidity should not use sticky information as neither its only nor main mechanism, as doing so could result in incorrect model predictions. In the second essay, I study optimal progressive income taxation in an incomplete markets model with a fat-tailed idiosyncratic income growth process with cyclical skewness. The process follows a two-state regime switching process representing recession and expansion. A Ramsey planner is constrained to two log-linear tax and transfer functions, one for each aggregate state, to allow state-dependent tax progressivity. I take recently created computational techniques used for aggregate MIT shocks and adapt them to work for regime switching processes. I find that the overall negative skewness of income growth leads to higher optimal progressivity, but how it is spread between aggregate states barely matters; progressivity should be higher in expansions and recessions, but allowing policy to be state-dependent has negligible effects on social welfare. I also find that the welfare change from current to optimal policy in a misspecified model without skewed and fat-tailed income growth is inflated to 2.9% compared to 0.9% using the correctly specified model
Seismic Data Processing in Spherical Geometries with Synthetic, Laboratory, Field, and Whole Earth Cases
No full textThe seismic reflection method, widely used for exploration subsurface imaging, generally employs densely gridded surveys and Cartesian (x,y,z) geometry to create high-resolution images of subsurface interfaces and structures. But for cases with substantial curvature, from lab-scale spheroids through planetary bodies, a spherical (r,θ,ϕ) description may be more suitable. By adapting conventional reflection processing concepts to spherical systems, we derive approximate normal moveout (NMO) travel-time equations to allow enhancement via stacking of reflection events in spherical geometries. Using a ray-tracing spherical correction, the reflection processing flow is applied to computer-generated synthetic earthquake data (using the whole Earth PREM model). On these synthetic data, we can isolate the core-mantle boundary (PcP) and inner-core boundary (PKiKP) reflections to develop a core image. To test spherical processing on real data, lab seismic data was generated using a hollow spheroid model with a diameter of 43.5 cm using ultrasonic transducers. The spheroid was filled with water, and then several different core compositions of glass and plasticine were emplaced inside the model to emulate Earth’s structure. For each core type, a circumferential survey was used to generate seismograms around the spheroid model. The derived spherical NMO equation successfully flattened the lab-recorded reflections and produced images of the actual physical model. Additionally, a suite of geophysical imaging surveys using unmanned aerial vehicles (UAVs), ground-penetrating radar (GPR), and hammer seismic methods were performed at Enchanted Rock, Texas which exhibits a remarkably smooth rounded structure. These are some of the first geophysical surveys carried out in the area. The application of spherical processing on the Enchanted Rock seismic data revealed a previously unknown subsurface reflection and possible exfoliation horizon. Finally, we assembled recorded earthquake data (~5 million source-receiver seismograms) along a great circle area circumscribing the Pacific Ocean region. We spherically processed these seismograms and interpret a resultant coherent reflection at 510 s as the core-mantle boundary. This processing flow shows considerable promise to rapidly and robustly analyze large datasets in spherical geometries
Physics-Informed Decision Tree
No full textPhysics-informed neural networks (PINNs) have become a well-known machine learning (ML) model in the physics domain where partial differential equations (PDEs) are commonly used. PINN can efficiently learn the underlying physics pattern and perform accurate predictions. However, due to the black-box nature of the artificial neural networks (ANNs), it is lack of interpretability. Furthermore, calculating the derivatives for solving PDEs makes the model computationally expensive. To address these limitations of PINN, we introduce a physics-informed decision tree (PIDT). This novel approach aims to leverage the advantages of the decision trees, such as interpretability and cost-effective computation, while still having physics-integrated architecture. The introduced model’s accuracy and efficiency are illustrated and compared with the traditional PINN and traditional decision tree on Burgers' equation and heat equation. Experiment results show that PIDT has significantly faster training time than the traditional PINN while maintaining a sufficiently similar accuracy. It has comparable training time with the traditional decision tree and ANN, but PIDT has considerably better model accuracy. The comparison of the models demonstrates a promising result that PIDT can accurately perform and accelerate training speed, which can be essential in real-world problems
Robotic Gripper and Manipulator Design and Analysis for Button Mushroom Harvesting
No full textThis study analyzes the main components of an automated mushroom harvester, specifically the robot manipulator and the end-effector, with the goal of improving harvesting performance while minimizing damage to mushrooms. The study also explores the science behind mushroom damage, identifying critical factors such as force application speed, contact time, and gripper design, which contribute to dents, bruises, and cracks. In response, a flex gripper with a deformable three-finger structure and embedded bending and twisting mechanisms was proposed. Additionally, a dynamic model for a 4-link SCARA manipulator with a PRRR configuration was developed for optimized mushroom harvesting. Compression and stress-relaxation tests were conducted on white button mushrooms of various shapes and sizes using a Texture Analyzer and P/75 probe. These tests identified the viscoelastic properties of the mushrooms, showing time-dependent deformation and stress relaxation. A viscoelastic model was created using the generalized Maxwell model and Prony’s series, which accurately represented the mushroom's mechanical behavior. The results showed that deformation increased as force application speed increased, with cracks occurring at lower forces for faster speeds. Finite element analysis (FEA) was used to model mushroom deformation during picking. It was found that at a 5N gripping force, the maximum deformation reached 2.5 mm after 5 seconds, potentially causing permanent damage. At 3N, deformation was significantly lower, reducing the risk of permanent damage. The flex gripper was tested on various mushroom types in a commercial facility, with a 100% success rate for single-grown mushrooms and 76% for low-density clusters. Finally, the dynamic model of the 4-link SCARA manipulator, which included friction and damping forces, was validated through a Simulink simulation. This comprehensive study provides valuable insights into optimizing both the end-effector design and robotic system dynamics for improved mushroom harvesting