Search

It has always been an important development strategy for software to keep pace with its evolution and meet stakeholder needs. Therefore, many software systems have very short release cycles. When many releases appear, how to test them becomes more and more important. Since the new features are often innovative, there are no conventional tools to test them, and manual testing is actually very cumbersome, not to mention that there are many non-testable features. This thesis mainly investigates the testability of the 109 features of 16 versions of the Mozilla Firefox browser from May, 2022 to July, 2023. The objects are the features of Firefox browser itself under the Windows 11 operating system, excluding additional plug-ins or Firefox browsers under other operating systems. Through manual testing and analyses, the features of the Firefox browser are divided into testable and non-testable, and the reasons why features are non-testable are given. Since the object being tested is the features of the browser, Selenium—a tool dedicated to testing browsers is considered. In this thesis, the Selenium package based on Python is used to simulate manual testing. By comparing the results of manual testing and the results generated by Selenium, we can gain insights into the limitations of Selenium testing, which contributes to improving automated testing in the future work. In this thesis, we also employ OpenAI’s ChatGPT (version 3.5) to assess the testability of each feature. We then compare ChatGPT’s answers with the manual analysis results.

Asthma is a widespread chronic respiratory disease which is poorly understood despite significant research efforts into its etiology. Some of the difficulty involved is due to its heterogeneous nature and the wide variety of factors involved. Developing a predictive model for asthma risk would help us to better model and predict asthma and develop better treatments and protocols for its management. Bayesian networks are a useful tool that have been use to learn relationships embedded in an observed dataset, but causality cannot be inferred from any relationships learned in this way. Bayesian networks also are sensitive to the composition of the given dataset, and relationships present in only subsets of the data may not be identified. We propose a methodology that integrates structural equation modeling with Bayesian networks to facilitate the development of predictive models of asthma to incorporate causal information based on domain knowledge. We also utilize clustering to identify and model unique subsets of the population with potentially unique relationships of various factors relevant towards asthma. We have applied this approach with NHANES data to model the risk of asthma in subpopulations of residents of the United States. We have also validated several steps in our methodology to demonstrate its effectiveness for developing predictive models for asthma risk in specific subpopulations. We believe that this methodology can be extended to many real world complex processes with heterogeneous phenotypes to create robust models and find new insights.

Flapping-Wing Micro Air Vehicles (FW-MAVs) are very small air vehicles that fly in a manner similar to that employed by small winged animals like hummingbirds or bumble bees. More traditional air vehicles tend to discretely allocate responsibility for various whole-craft motions to subsystems that are actuated separately. Jet engines, for example, produce forward thrust while ailerons, elevators, and flaps, introduce differential drags to actuate yaw, pitch, and ultimately, elevation. FW-MAVs tend to have a much smaller number of actuatable effectors (E.G. two flapping wings) that must provide all of these services in an integrated way. Flight control of FW-MAVs is challenging with, perhaps, the first effective control systems appearing in the literature only in the 2010's. Much research has been done in an attempt to understand the physics governing FW-MAVs, but very little of it at the time of this thesis addresses how a vehicle can adjust to ongoing damage and wear to what are, often, relatively fragile wings even though it can be demonstrated that even relatively minor wing damage can prevent a vehicle from achieving desired waypoints in a flight plan. The bulk of the work available that addresses this issue via in-flight adaptation and learning employs Evolutionary Algorithms. Building upon those efforts, this thesis will provide three distinct contributions. First, existing vehicle flight simulators were modified to allow use via a socket interface. The original simulator had a C-language API only and was difficult to use with contemporary machine learning tools written in Python or other languages. The new socket interface enables convenient interfacing to a wide variety of languages and tools and amplifies the techniques we may bring to bear on the learning problem. Second, over the years the existing Evolutionary Algorithm approaches have been adopting progressively more simple internal representations. This work will move the needle a bit more toward simplicity and examine if there are any improvements or benefits to be gained, such as increased performance or simpler implementation. Third, allied researchers have recently been asking if Reinforcement Learning might be applied to this application. This thesis will present some initial explorations in doing so by leveraging the socket-based API to enable the use of standard Reinforcement Learning packages.

The threat of malware is continuously evolving, with each newly designed system developed to detect and process new forms. Antivirus applications started with detecting malicious code via signatures, however, recently they started using Heuristics to detect and stop malware. The effectiveness of heuristics models depends on how representative are the used features, which are usually secondary features, that are sometimes not necessarily a reflection of a malicious code inside the program, thus, leaving a gap for false positives. Examples of such features are files accesses or system calls logs. Many examples can be listed of how successful bioinspired systems can be, given a well-studied biological system with principles that can be copied to solve a problem in a different field. With the discovery of CRISPR and the characterization of its proteins, this study takes inspiration from, specifically CRISPR, it to design a system to detect and process malware. In this work we built an analogy to map the different biological structural units to the matching structural units of a program. Based on it, we propose a model inspired by CRISPR targeting via PAMs, to target the promoters for the malicious functions of a malware. We used Triplet network, with CNN layers, to feed in the data in form of images. We believe that despite not having a fully functional model at this stage, due to the way we fed the data into the network and the current state of the feature extractor. The ideas presented in this work could be a starting point for a new way of looking at the problem of detection of malware and healing the infected files.

This thesis uses Raman spectroscopy and machine learning to forecast the pre-osteoblast proliferation rate. Raman spectra were taken while live preosteoblast cells were cultured for different lengths of time (7, 10, and 14 days). After preprocessingtechniques were used to improve the spectra’s quality, Principal Component Analysis (PCA) was used to find significant spectral features linked to temporal fluctuations. Identification of prospective biomarkers was made possible by the correlation between the strength of these features and the proliferation rate as determined by the biochemical assay, Water-Soluble Tetrazolium (WST-1) assay. Using preosteoblasts-specific Raman spectral characteristics and accompanying proliferation rates, machine-learning models were created. Model performance was judged using accuracy, Mean Square Error (MSE), and R-squared metrics. The algorithms used included random forest, Support Vector Machine (SVM), and ridge regression. Results showed that machine learning could accurately forecast the proliferation of preosteoblasts, with identified Raman peaks acting as potential indications. This study adds to our understanding of preosteoblast cellular dynamics and offers valuable information for tissue engineering and regenerative medicine applications. It also presents a novel method for forecasting cell proliferation rates using Raman spectroscopy along with quantitative biochemical assay data and machine learning,opening the door to individualized cell therapies and better patient outcomes.

With the rapid development of satellite and remote sensing technology, thehuge remote sensing data is being accumulated, which has been widely used in manyfields, such as land resource monitoring, Unmanned Aerial Vehicle (UAV) trafficsurveillance, change detection and so on. Deep learning methods have achieved greatsuccess in the field of computer vision, such as image classification, object detection and so on. Deep learning methods can help remote sensing image analysis inmany ways, for example, representative feature extraction, accuracy improvementcompared to traditional machine learning methods, large-scale parallel image dataprocessing and computation, etc. However, there are still some challenges for deeplearning-based remote sensing image analysis under some complex situations, such asthin cloud, thick cloud, adversarial attack and defense, and camouflaging adversarialattack and so on. Directly and simply applying the existing deep learning models onthese complex scenarios might fail to some extent.Therefore, this dissertation focuses on the deep learning based remote sensingimage analysis under complex scenarios. Specifically, four remote sensing relatedresearch problems are systematically studied in this dissertation. For problem 1, weexplored the effect of thin clouds on the remote sensing scene classification. Forproblem 2, we studied the occlusions by thick clouds on the tasks of remote sensingscene classification, cross-view geolocation, and semantic segmentation. For problem3, we studied the adversarial cloud based hacking attack and defense methods onremote sensing datasets. For problem 4, we investigated camouflaging adversarial attack methods against deep learning based UAV object detection models in realworld.Experimental results on the publicized and synthesized remote sensing datasetsshow the effectiveness and accuracy of the proposed deep learning methods for remotesensing image analysis under complex scenarios.

Hardware Trojans are malicious circuits, hidden in integrated circuits (ICs) which pose a significant threat to security. Detection of hardware Trojans is important to build trust, verify, and make the semiconductor ICs process secure. The existing hardware Trojan detection methods are generally destructive, require intricate comparisons, or require a long time for reverse engineering.In the initial phase of this study, the substitution of supervised hardware Trojan detection methods in ASICs chips is explored with unsupervised approaches, thereby eliminating the dependence on golden references. The Trojan detection uses a ring oscillator (RO) based on NAND as the power monitor. Frequency side channels are computed, and unsupervised classifiers are employed, achieving a commendable accuracy of 93%, comparable to state-of-the-art methods requiring golden data.In the later part of our work, different feature extraction techniques are applied to improve the competence and precision of Trojan detection. Experiments include frequency side-channel extraction with the use of implanted ROs in Basys FPGA, resulting a higher F1 score and accuracy (95\%) in detection of the Trojan. Further, the method has performed equally well for all the Trojan sizes considered (8-bit, 16-bit, and 32-bit) during the performance evaluation with no significant difference and has yielded equal efficiency in detecting even small-sized Trojans.After clustering ICs, a post-processing localization algorithm is proposed to help in identifying and locating hardware Trojan within the chip with reference to the ROs. This algorithm eliminates the requirement of complete reverse engineering of at least one IC from a cluster to figure out if the cluster formed by ML contains Trojan-infected or not 'golden' chips.The proposed algorithm simplifies this process by focusing mainly on the chip portion, where the Trojan is localized.Lastly, the study presents some vulnerability of the clustering model to adversarial attacks in the Trojan detection technique. These attacks reduce significantly the accuracy of chip clustering into the correct cluster up to 50\%. Consequently, we advocate for the integration of adversarial defense mechanisms into the hardware Trojan detection machine learning models, an area we identify as a focus for future research. In conclusion, the presented work enhances unsupervised hardware Trojan detection improvements by presenting new feature extraction techniques and giving a localization algorithm to avoid reverse engineering efforts. The need and importance of adversarial defense mechanisms to make machine learning systems detect Trojans has been highlighted.

Hardware components are becoming prone to threats with increasing technological advances. Malicious modifications to such components are increasing and are known as hardware Trojans. Traditional approaches rely on functional assessments and are not sufficient to detect such malicious actions of Trojans. Machine learning (ML) assisted techniques play a vital role in the overall detection and improvement of Trojan. Our novel approach using various ML models brings an improvement in hardware Trojan identification with power signal side channel analysis. This study brings a paradigm shift in the improvement of Trojan detection in integrated circuits (ICs).In addition to this, our further analysis towards hardware authentication extends towards PUFs (Physical Unclonable Functions). Arbiter PUFs were chosen for this purpose. These are also Vulnerable towards ML attacks. Advanced ML assisted techniques predict the responses and perform attacks which leads to the integrity of PUFs. Our study helps improve ML-assisted hardware authentication for ML attacks. In addition, our study also focused on the defense part with the addition of noise and applying the same attack ML-assisted model.Detection of Trojan in hardware components is achieved by implementing machine learning techniques. For this Purpose, several Machine learning models were chosen. Among them, Random Forest classifier (RFC) and Deep neural network shows the highest accuracy. This analysis plays a vital role in the security aspect of all hardware components and sets a benchmark for the overall security aspects of hardware. Feature extraction process plays major role for the improvement of accuracy and reliability of hardware Trojan classification.Overall, this study brings significant improvement in the field of overall hardware security. Our study shows that RFC performs best in hardware classification with an average of 98. 33% precision of all chips, and deep learning techniques give 93. 16% precision of all chips. Moreover, on the hardware authentication side, RFC performs the best of all other models with the accuracy of 89% in the attack part and 81% in the defense part in 6000 data samples.

Semiconductor microelectronics integrated circuits (ICs) are increasingly integrated into modern life-critical applications, from intelligent infrastructure and consumer electronics to the Internet of Things (IoT) and advanced military and medical systems. Unfortunately, these applications are vulnerable to new hardware security attacks, including microelectronics counterfeits and hardware modification attacks. Physical Unclonable Functions (PUFs) are state-of-the-art hardware security solutions that utilize process variations of integrated circuits for device authentication, secret key generation, and microelectronics counterfeit detection. The negative impact of aging on Static Random \linebreak Access Memory Physical Unclonable Functions (SRAM PUFs) has significant consequences for microelectronics authentication, security, and reliability. This research thoroughly \linebreak examines the effect of aging on the reliability of SRAM PUFs used for secure and trusted microelectronics integrated circuit applications. It initially provides an overview of SRAM PUFs, highlighting their significance and essential features while addressing encountered challenges. The study then covers mitigation techniques, including multi-modal PUFs, that already exist to boost the resilience of SRAM PUFs against aging impacts, highlighting their advantages and the gap in the research addressed in this research. This work proposes a novel AI-enabled security for reliable SRAM PUFs. The proposed approach aims to study and countermeasure the impact of aging on SRAM PUF by analyzing data samples, including Bias Temperature Instability (BTI), Bit Flips, Accelerated aging, and Hot Carrier Injection (HCI) and to study their effects on SRAM PUF cell properties and output. Accelerated aging is a direct result of a change in the environmental temperature and voltage for a few hours. We aim to mitigate the impact of accelerated aging on the reliability authentication and encryption keys of SRAM PUFs. Further, AI-assisted approach is used to analyze SRAM PUF stability under accelerated aging operating conditions. Illegal memory chips from shady companies around the world have compromised the safety and reliability of electronic devices. Detecting these manufacturer details is crucial to detecting illegal and substandard manufacturing prior to their integration into key systems. This paper presents a novel approach designed specifically to analyze the SRAM PUF manufacturers. The suggested technique assigns a distinct process variation to every manufacturer, \linebreak facilitating the classification of the manufacturer, without the need for elaborate registration or verification protocols.The experimental results show the reliability and stability factors of SRAM PUF, including its influence on environmental conditions and the associated effects of aging. Our findings stress the importance of reliable and trustworthy PUF-based hardware security to reliably classify older microelectronic devices from new ones. Using 345 FPGA tested chips, the results show that using different ML models, we can efficiently classify these chips to correctly distinguish the manufacturer, aging impacts, and component, with F1 scores of 96\% and 98\%.

The effectiveness of a deployed knowledge graph is commonly evaluated with defined use-cases from domain experts. This poses challenges during the development cycle in determining how to represent data. Developers of a knowledge graph can optionally include semantics into a knowledge graph by abstracting the data representation in such a way that mirrors information as it exists in the real world. Consequently, the abstraction is represented by additional layers, resulting in performant differences in knowledge graph embedding; such as, the embedded model's ability to infer facts between entities through link predictions. This thesis presents a comprehensive analysis of the performance impact observed across a range of knowledge graph embedding models trained on FB15k-237, a widely recognized benchmark dataset for knowledge graph completion. Additionally, the experiment is performed with augmented versions of FB15k-237, serving to introduce semantics into the knowledge graph.

Knowledge Graphs (KGs) leverage structured data (entities and their relationships) to create a richly interconnected world. However, to fully explore these intricate connections, sophisticated exploration tools are essential as manual exploration can become overwhelming. Applications for KG exploration span many use-cases: social network analysis, corporate intelligence, and medical research.This research improves the InK Browser (Interactive Knowledge Browser), a modular, web-based tool for KG exploration, facilitated by flexible views. The goal is to enhance user understanding and this is tested through a user study.Flexible views are made possible by applying complex constraint definitions against data instances. When data points (and their relations) match a data shape, the flexible view provides an adaptive perspective of that data. The InK Browser already provides a flexible view for geospatial data (a map) and metadata (semantic and type information), as well as search functionality. This research has added a new functionality in Flexible Views, a KG summarization that is utilized within the InK Browser by the dynamic creation of SPARQL queries made from shortcuts of the used schema. This functionality aids the challenge of navigation and comprehension of KGs.

Planetary health recognizes the inextricable link between human health and the health of our planet. Our planet’s growing crises include biodiversity loss, with animal population sizes declining by an average of 70% since 1970, and maternal mortality, with 1 in 49 girls in low-income countries dying from complications in pregnancy or birth. Overcoming these crises requires effectively allocating and managing our limited resources. This dissertation develops data-driven AI decision-making methods to do so, overcoming the messy data ubiquitous in these settings. Here, we present technical advances in multi-armed bandits, reinforcement learning, and causal inference, addressing research questions that emerged from on-the-ground challenges across conservation and maternal health. These methods enable practitioners to take efficient and robust actions necessary for planetary health. Specifically, this work was done in close partnership with NGOs, government agencies, and other stakeholders to deploy AI for on-the-ground intervention efforts in biodiversity conservation, where these methods have been integrated with a system used by 1,200 protected areas worldwide, and maternal health, working with the world's largest mobile health program to reduce maternal mortality in India.

Generative Adversarial Networks (GANs) were initially developed for computer vision tasks and have shown impressive capabilities in enhancing the performance of various AI solutions. In the field of speech signal processing, numerous studies have utilized GAN models for speech recognition and enhancement. However, the effect of GANs on speech separation has not been extensively explored.This thesis proposes a comprehensive framework to study and compare the efficacy and efficiency of various GAN variants for the task of speech separation. Within our framework, we employ Conv-TasNet as the common generator and propose deep Convolutional Neural Networks as the discriminator in different GAN setups. The GAN models are designed with diverse objective functions, including the original GAN, Least Squares GAN (LSGAN), and MetricGAN. Results from the experiments demonstrate the improved performance of our GAN-enhanced Conv-TasNet models. Among all the models evaluated, MetricGAN appears to be the most effective variant, demonstrating significant improvements in Scale-Invariant Signal-to-Noise Ratio (SI-SNRi) over the baseline model.

There are numerous software testing methods available today in the software developmentprocess, each with its own set of benefits and drawbacks. Regression testing is a typeof software testing that is used to ensure that recent changes to a program or codehave not adversely affected existing functions. The human-driven stages of traditionalregression testing are eliminated through the use of automation-based tools. There aretwo types of software testing: parallel testing and sequential testing. Parallel testing isthe execution of multiple automated tests at the same time. As a result, the executiontime is decreased. There are numerous trustworthy and efficient tools available for paralleltesting in automated regression. The aim of this thesis is to examine concurrent regressiontesting in a production environment using docker containers in the GitLab environment andautomating the software testing process using GitLab CI/CD. Automating test executionconcurrently across multiple docker containers leads to time and effort savings. Theimplementation of Delay-Tolerant Networking (DTN) known as NASA’s InterplanetaryOverlay Network (ION) software distribution serves as a solution for network connectivityin interplanetary communication. In this thesis, various operating systems are used asthe base image for the docker container to measure the parallel execution performanceof ION-DTN regression tests. The performance evaluation involves analyzing differentparameters, including test execution duration, test failure occurrences, and system memoryand CPU usage. In addition to parallel testing, the approach proposed in the thesis alsoprovides support for sequential testing in operating systems that have limited virtualization capabilities. Additionally, the approach outlined in this thesis incorporates an automaticretry mechanism for failed tests.

The world is currently experiencing a shortage of semiconductor chips. This shortage is affecting different industries that rely on electronic components that involve semiconductor chips to manufacture their products. Due to the shortage of chips, manufacturers are unable to complete the final assembly of their products, resulting in a delay in delivering the finished products to their customers. To address this issue, the US Congress passed the "Creating Helpful Incentives to Produce Semiconductors (CHIPS) and Science Act of 2022" on 9th August, 2022. This act aims to improve the competitiveness, innovation, and national security of the US.This dissertation focuses on addressing the chip shortage through the reduction of chronic semiconductor manufacturing process quality problems caused by wafer map surface defects. The proposed solution involves detecting mixed-type wafer map surface defect patterns using Ensemble Deformable Convolutional Neural Networks. The framework for defect detection proposed in this dissertation outperforms other machine learning models from literature, such as Conv-Pool-CNN, All-CNN, NIN-CNN, DCNN-v1, and DCNN-v2, in terms of F1-score. The proposed framework uses an industrial wafer map dataset (MixedWM38) from a semiconductor wafer manufacturing process to train the base models for the ensemble method. The results show that the proposed framework accurately identifies multi-pattern defects from the surface of wafer maps. This dissertation will contribute to advancing academic literature for the new field of detecting mixed-type defect patterns from the surface of wafer maps. Defects are indicators of process problems, and preventing quality defects in advance is the best approach to achieving positive yield. The efficient and accurate detection of wafer map mixed-type surface defect patterns is important for addressing chronic manufacturing process quality problems. The proposed framework can be used by semiconductor manufacturers for real-time defect detection on the manufacturing floor, enabling the identification of the relevant manufacturing process from where the defect originates and resolving process problems quickly. This dissertation will contribute to the advancement of semiconductor chip manufacturing by improving defect identification methods and reducing chronic manufacturing process quality issues that will lead to better yield outcomes, minimizing waste and damage. With the world facing chip shortages, improving the yield will contribute to solving the chip shortage crisis globally.

Aerial Vehicle (UAS) remote and autonomous operation is a growing industry with applications in defense and the commercial space. UAS's can use the Global Positioning System (GPS) to determine their position when signals are available. Operational environments like urban canyons, forested areas, and indoors can block the UAS from receiving GPS signals and prevent the receiver from determining a position, which can hinder the vehicle's ability to successfully complete the mission. Unreliable positioning from GPS data requires UAS's to have an alternative means of localization, which can be accomplished by utilizing ultra-wideband ranging to landmarks and other aerial systems. Machine learning can be used to train models to utilize raw data from intervehicle distance sensors on the agent and anchors to determine the location of the agent without access to GPS data. This work will explore using machine learning as an adaptable localization system using neural networks and gradient descent methods. Fully trained neural networks will be capable of learning specific noise models and performing cooperative localization using unfiltered intervehicle ranges and anchor positions.

In recent years, there has been significant research and development in Deep Learning (DL) due to its efficiency and extensive applicability across diverse domains, including Computer Vision and Large Language Models. However, the architecture of large Deep Learning models, containing dense layers, makes them compute and memory intensive. Distributed Deep Learning (Distributed DL) is the successful adaption to accelerate and enable training and inference for large-scale DL models, where it also deals with various parallel approaches, inference and training techniques, and communication optimization strategies to enhance performance.In this thesis, we focus on accelerated and memory-efficient techniques to optimize distributed training and inference. It is broadly categorized into three different approaches: 1. Inference for scaled images using quantization, achieving a speedup of 6.5x with integer- only quantization and 1.58x with half-precision, with less than 1% accuracy degradation. 2. MPI4DL: Distributed Deep Learning Parallelism framework encompassing various parallelism techniques with integral components such as Spatial Parallelism, Bidirectional Parallelism, and Hybrid Parallelism 3. Communication optimization by leveraging MCR- DL: A distributed module for DL frameworks with support for mixed-backend communication, dynamic selection of the optimal backend, and communication optimization enhancements such as compression and tensor fusion.

In this work, we study how electrical stimulation of the nerve can be used to restore the sensation of touch to individuals with hand injuries. To recreate the sensation of a real mechanical touch, it is necessary to understand how it is encoded by the population of neurons on a finger. We use a simulator capable of modeling different types of individual receptor behavior on the finger and generate data for a wide variety of touch stimuli. Then, we modify online dictionary learning methods to learn embeddings of these datasets that are sparse (use few receptors) and local (the receptors are grouped locally on the finger). To do this we develop a novel locality penalty function for dictionary learning. Experiments show that these embeddings can reconstruct novel neural responses more accurately than manually curated features representing the state of the art. We then develop neural network architectures that can learn optimal stimulation parameters using such embedded data. Our experiments show that one of our architectures can recover the optimal stimulation parameters for desired novel neural responses accurately about 85% of the time.

In the field of link prediction, determining an accurate understanding of how well algorithms are performing is important to make informed decisions about how the algorithms can best be utilized. Existing metrics are reliant on using a setting’s known links to judge the algorithm’s accuracy, typically using a train-test split. However, if assessing link prediction in a setting for which the ground truth is difficult to determine or subject to sample bias, this can lead to errors that are not easily detected by standard link prediction validation. As a solution to this, we propose using model-based validation (MBV) as an alternative assessment for link prediction algorithms, where the algorithm is graded based on how well it reflects a model underlying the data. Additionally, we test a type of network refinement to see if degree bias as a mechanism for sample bias can be addressed by changing the network’s degree distribution directly.

Graph Neural Networks (GNNs) have risen as powerful approaches for learning graph representations. Nevertheless, their performance is closely tied to the quality of the graph data they utilize. The integrity of graph data, particularly when compromised or subjected to adversarial attacks, can drastically impact the success of such learning algorithms. The pivotal question then is: How can we distill valuable knowledge from graphs riddled with noise? Addressing this requires the development of robust methods tailored to learn graph representations efectively, despite the presence of noise in the data.On the other hand, it is quite typical to encounter prevalent inaccuracies, inconsistencies, and incompleteness in real-world graph data. This situation demands the development of sophisticated techniques to address these errors, which is critical for applications in felds such as social network analysis, bioinformatics, and recommendation systems. The intricate nature of graph data means errors can propagate, afecting the entire structure, thereby highlighting the importance of advanced techniques to extract insights, enhance data quality, and boost machine learning model performance of downstream tasks. This leads us to another important question: How can we employ advanced graph learning strategies to improve the quality of the graph?This thesis underscores the importance of data-centric approaches in graph representation learning, prioritizing dataset enhancement to address the root causes of performance degradation. It begins by presenting two novel frameworks, GEDet and GALE, which are designed to overcome the obstacles of employing graph learning techniques to enhance graph quality. GEDet utilizes an adversarially learned, few-shot approach to detect erroneous nodes, enriching training data through a graph augmentation module and enhancing detection accuracy with an adversarially learned module. GALE extends this concept by combining active learning with a graph generative adversarial model, efciently using limited labeled examples to improve error detection signifcantly. On the other way, this thesis investigates the development of robust graph representation learning strategies under the assumption that the underlying graph may be compromised by attacks. RoboGNN further contributes by robustifying GNNs against link perturbations, employing a co-learning strategy to refne both the graph structure and the GNN model, ensuring model reliability under adversarial conditions. Our round-trip framework showcases its efectiveness in downstream applications, like steadfast prediction of time-series data within spatial-temporal graphs.

As Quantum Computers continue to increase in size, throughput has not increased proportionally. Researchers have begun exploring ways of parallelizing circuit execution. Due to the noisiness of quantum computers, this requires new algorithms for efficient resource allocation, qubit mapping, and scheduling. We improve on existing greedy algorithms by formulating the mapping search as a Binary Integer Non-Linear Programming (BINLP) problem. We model practical constraints and propose new heuristics for determining the goodness of a mapping. We observe similar fidelity compared to Qiskit’s transpiler for circuit cutting and throughput benchmarks. We observe greater fidelity over Qiskit’s transpiler for dense QAOA ansatzes. We find that parallel circuit cutting provides greater fidelity than full-circuit execution. We also propose a scheduling algorithm for parallelizing circuits of different lengths and shot counts. Our algorithm achieves a lower makespan for time and number-of-shots for diverse workloads as well as lower per-job runtimes in all cases.

Patients hope and clinicians expect medical imaging to provide the best scans possible on that day given their symptoms. For individuals trusting MRI for diagnosis and treatment, nothing is "standard" about their clinical reality, and therefore nothing should be "standard" about their imaging protocol. This ideal underpins the following dissertation detailing the initial development of technology and infrastructure allowing Magnetic Resonance Fingerprinting (MRF) to provide personalized acquisitions, reconstructions, and reporting to every patient, every time. To achieve personalized acquisition, reconstruction and reporting of quantitative MRI data, a fundamental shift is needed in the scalability, flexibility and interoperability of the underlying infrastructure. Central to this effort, this dissertation details the development of scalable online reconstruction pipelines, innovative visualization and communication techniques, and the mrftools framework, which collectively facilitate immediate clinical decision-making, enhance diagnostic accuracy, and streamline clinical qMRI workflows. Key contributions of this research include efficient acquisition and reconstruction of MRF data at clinical scale via cloud computing, automated visualization, report generation, and quality control systems. This research shows that 3D-MRF is highly repeatable in-phantom and in-vivo across scanners, software versions, and subjects, warranting the initial deployment of a comprehensive MRF protocol at the University Hospitals Cleveland Medical Center. Notable success is shown in integrating MRF into standard clinical procedures, demonstrating the practical feasibility of MRF to establish quantitative baselines across diverse patient cohorts, informing future clinical decision-making. Looking ahead, this dissertation lays the groundwork for further enhancing MRF technology, including the development of automated population-scale quantitative biomarkers and the realization of a closed-loop protocol optimization system. These future directions aim to leverage the full potential of MRF, making it a cornerstone of quantitative and precise diagnostics tailored to each patient's unique pathology and medical history.By addressing technical and operational challenges and demonstrating the potential for significant improvements in patient experience, this dissertation paves the way for a future where diagnostic processes are not only more efficient and precise but also personalized to the unique needs of each patient, fulfilling the promise of personalized medicine in the realm of medical imaging.

CpG sites occur at an irregularly high frequency at regions of the genome called CpG islands; in normal samples, CpG islands are typically unmethylated. When a number of these CpG islands are instead hypermethylated in a cancer sample the sample is referred to as possessing CpG island methylator phenotype (CIMP). CIMP has been well-defined (and was first reported) in colorectal cancers, but research has also been performed on numerous cancer types such as gliomas, melanomas, and leukemia. The goal of the work presented here is to analyze the differences in gene mutation and expression between samples with CIMP (CIMP+) and samples without CIMP (CIMP-) across multiple cancer types. First we perform a coverage analysis to find the sets of genes and mutations which best correlate with CIMP+ samples while minimizing correlation with CIMP- samples. The cancer types used in these analyses are colorectal cancer (COADREAD), stomach cancer (STAD), and uterine corpus endometrial carcinoma (UCEC). The results from each tumor type are combined to create a synthesized set of mutations and mutated genes which correlate with CIMP+ samples. We then perform differential expression analysis on RNA-Seq data from The Cancer Genome Atlas (TCGA) to determine genes that are differentially expressed in CIMP+ samples when compared to CIMP- samples. We find 890 mutations and mutated genes with strong positive correlation to CIMP, many of which are also differentially expressed between CIMP+ and CIMP- samples. Among these mutated and differentially expressed genes are numerous genes in the MSigDB KRAS down-signaling gene set. This analysis furthers our understanding of CIMP and cancer in general and may give more insight into the differences between the genomic characteristics of CIMP-positive and CIMP-negative cancers.

The evolution of blockchain technology has revolutionized various industries, offering decentralized and secure systems for digital transactions and data management. Central to blockchain’s functionality is its reliance on a distributed network of nodes collaborating to validate, process, and record transactions. However, scalability remains a fundamental challenge, with issues such as transaction processing delays, rising fees, and network congestion hindering widespread adoption. Addressing scalability demands innovative approaches, prompting the design of efficient peer-to-peer topologies to optimize network performance. Our research explores various facets of network optimization, aiming to enhance user satisfaction and system efficiency within decentralized ecosystems.We delve into the optimization of bandwidth and latency, treating the network as a societal entity to reduce costs and propagation latency. Through protocols like Perigee, Goldfish, and Topiary, we introduce decentralized solutions aimed at minimizing block prop- agation time and enhancing neighbor selection efficiency. Additionally, our work extends to federated Byzantine agreement systems, developing algorithms to optimize bandwidth costs while considering propagation latency requirements and unique peer connection rules.Furthermore, we address the optimization of individual rewards within Proof-of-Work (PoW) and Proof-of-Stake (PoS) networks. In PoW networks, strategies like Less is More analyze reward distribution bias and propose topology strategies to optimize miners’ rewards. For PoS networks, we introduce Flashback, focusing on equitable balance among participants and leveraging pre-known validator information to facilitate early communication and optimize rewards.Our contributions lie in the design of efficient protocols for minimizing propagation latency, strengthening networks with considerations of topics and bandwidth, and addressing reward bias in PoW networks. Additionally, our work rethinks optimal builder bidding strategies in PoS networks, introducing auction mechanisms that outperform current methods at equilibrium. By synthesizing these efforts, our thesis aims to advance understanding and implementation within decentralized networks, fostering innovation and progress in the blockchain domain. Through our contributions, we strive to pave the way for a more efficient decentralized networks.

Graph Neural Networks (GNNs) have emerged as a powerful class of machine learning techniques capable of processing graph-structured data, showing immense promise in molecular property prediction. This thesis compares the performance of two specific GNNs—Graph Attention Networks (GATs) and Attentive FP models with Graph Convolutional Networks (GCNs) in predicting the biological activities of protein targets across several protein families. Our findings indicate that while GCNs are highly effective for molecular property prediction, GATs and Attentive FP models also offer competitive performance, with GATs showing particular promise for enzymes and transporters. The experiments suggest that the choice of model should be tailored to specific families of target proteins, highlighting the need to consider the particular protein family when selecting GNNs for predictive modeling in drug discovery.

The NASA Astrophysics Data System (ADS) is a critical resource for researchers and students in astronomy, astrophysics, and beyond. ADS indexes a vast collection of papers and scholarly literature that researchers can search through using the ADS website or API. ADS’s database is powered by Apache Solr, enabling users to formulate highly expressive and precise search queries from the more than 50 allowable search fields. However, the sophistication of ADS’s search capabilities comes at the cost of usability, necessitating users to familiarize themselves with Solr and ADS’s documentation to fully exploit its features. This thesis proposes a solution to enhance the accessibility of ADS by creating a chat application where users make requests for papers by asking for them in natural language rather than by constructing Solr queries. This application works by leveraging SOTA transformer-based large language models (LLMs) to translate natural language requests into Solr queries, thereby simplifying user interaction with the ADS database without compromising on the precision of search results. In this work, we use in-context learning (ICL) with retrieval augmented generation (RAG) in order to enhance the translation capabilities of the LLM, leading to significant improvement in translation performance.

Within the game development industry, consumers often are presented with two general types of products: AAA and Indie. At present, the AAA landscape is growing bloated and homogenized while independent studios continue to push the bounds of creativity despite smaller teams and budgets. That said, within the sphere of games, much of the monetary attention is on the AAA space; the games on offer demonstrate unprecedented scale, fidelity, and shine. Despite Indies being more creative and nuanced, for these reasons, they can be undervalued and overlooked. In today's modern age, however, developers are presented with tools capable of closing that gap. Through procedural asset creation and placement tools, the visual scale of Indie games can increase. With the introduction of AI (Artificial Intelligence) assisted development, iteration loops can be cycled through faster resulting in more polished and detailed systems. These tools can significantly impact independent studios in the scale of their projects and their monetary foothold in video games.The following thesis is an exploration of these tools and their effectiveness. Specifically, Houdini: Indie, Houdini Engine, Chat GPT 4.0, and Unreal Engine 5 will all be explored. Houdini will serve as the visual mechanic providing procedural asset creation. ChatGPT will serve as the AI-driven assistance for development. Unreal Engine will serve as the testing ground. The research is divided into two parts. The first half concerns development in Unreal utilizing Chat GPT. The second pertains to creating visual tools in Houdini that can be further used to create environments in Unreal. Throughout the project, Chat logs with ChatGPT have been saved, and a weekly record of progress has been kept. The goal of this process is to evaluate the tools' effectiveness and come to a reasonable conclusion as to the ramifications of the introduction of AI like ChatGPT and whether there may be roadblocks toward this goal, as well as an exploration into the current state of Houdini's introduction into the Indie game sphere.

Drawing inspiration from lateral thinking, synergetics, psychology, creativity, and business, this research project employs an interdisciplinary approach to investigate the research process which drives innovation in the field of artificial intelligence. This research project explores methods for harnessing the synergy present in the latest, neuro-symbolic paradigm of artificial intelligence, while noting similarities between the first two waves of AI and dual process theory. It attempts to integrate unconventional, yet potentially promising interdisciplinary ideas into a proof of concept, including creative tools and techniques like the Six Thinking Hats, methods of psychotherapy, including cognitive behavioral therapy and internal family systems, as well as principles related to conflict resolution and ``tensegrity". The proof of concept is a hybrid semantic search system for research papers in computer science, constructed using a process of rapid prototyping and iteration, with special consideration for evaluating how more modular, interpretable, and human-centric approaches to system design can help narrow the gap between cutting-edge AI research and ethical, practical application in business. This research is conducted with the hope of opening the research field to greater creative possibility, as well as deliberate action towards creating more sustainable and human-centric artificial intelligence systems.

The national conversation about STEM education continues. While math and science have been a constant in K–12 and higher education, curriculum in technology and engineering have not been consistently part of the tapestry of American education. As such, there is a dearth of qualified candidates for the ever-growing number of computer science and engineering career opportunities. Prevailing stereotypes that depict the typical workforce in these industries as White and male contribute to the lack of representation of other groups. This study focused on the representation of girls and young women in computer science between the years of 2005–2018. Data tracking on Advanced Placement (AP) computer science enrollment started in 2005, making this an optimal year to start data collection. The researcher ended collection of articles in 2018. Using 47 articles from one of the country’s newspapers with the highest readership and with wide-reaching influence, The New York Times, a variety of concepts were coded and analyzed in a mixed-method design. Overarching themes emerged and formed a narrative about girls and young women in computer science. Within the text of the articles, causes for the disparity and solutions were explored. There was little conversation about policy changes. Various stakeholders who influence education pathways can use the findings of this study to craft solutions to what The New York Times chronicled as a national crisis.

Driven by significant advancements in algorithmic techniques and the emergence of new multimodal generative applications, deep learning has entered the era of "large-scale AI". As leading models dramatically increase in size and complexity, the hardware and software requirements also become significantly more demanding. If efficient solutions are not developed in a timely manner, model exploration will grind to a halt and at-scale serving will be infeasible. End-to-end co-design solutions must address three key themes: the unique technical challenges posed by the large-scale nature of these models, the distinct requirements for training versus inference, and the critical need for efficiency. This dissertation presents three case studies for navigating the complexities of large-scale AI. The first case involves a cross-stack characterization of large-scale models, identifying performance bottlenecks and potential avenues for optimization across different system layers. The second case study explores redesigning embedding-centric models through data and hardware-aware observations, aiming for substantial improvements from novel embedding representations. The third case study develops tools that help researchers gain better insights into mapping of increasingly complex models onto physical infrastructures, addressing the logistical and operational challenges of deploying large-scale AI systems in data centers. Looking forward, the dissertation identifies areas for future research, including co-design strategies tailored for embedding-driven, multimodal AI models and the role of reliability versus resiliency in data center-scale training environments. Collectively, this work contributes to the foundational understanding and practical advancement of large-scale AI technology, setting a course for future innovations in the field.

While major advancements have been achieved in many fields utilizing artificial intelligence for a variety of tasks, some specialized areas remain difficult to tackle. The legal domain is one such area. It is often said that legal language is a dialect of English and one that requires a Law degree to be fluent in. In this work we examined parallels between software engineering and legal drafting to develop definitions for contract smells, quick indications for potential issues with legal contracts. We created an auto labelled dataset of these contracts smells using engineered prompts and demonstrated how even a small set of human labels can significantly improve auto labelling results with few shots techniques. We demonstrated using bi-directional deep learning models that these contract smells can indeed be successfully detected automatically with high accuracy after further fine tuning BERT as well as LegalBert. This work underscores the feasibility of applying advanced NLP techniques to automate aspects of legal contract review and provides a strong foundation to further develop models for this purpose.

Biomimicry is a technique for solving complex technical challenges by applying biologically inspired processes. When considering flight through challenging conditions, bats (order chiroptera) have demonstrated a keen ability to use non-visual senses to navigate surroundings, avoiding obstacles, and create a mental map of the environment in which they live and operate. Unmanned aerial drones are controlled in flight by an intelligent agent, be it a human pilot or artificial intelligence (AI) based autopilot. Simultaneous Localization and Mapping (SLAM) is a technique for building a map of an environment while keeping track of a vehicle’s position within that environment. This thesis shows that is possible to use bats as the inspiration for the development of an AI autopilot, Chiroptera, that can conduct SLAM operations to navigate a drone through an unknown environment. A bespoke 3D drone flight simulator, Bat Cave, will be developed concurrently within which Chiroptera’s SLAM capabilities can be tested and its behavior refined.

The field of coding education is rapidly evolving, with emerging technologies playing a pivotal role in transforming traditional learning methodologies. This thesis introduces Dr. Lego, an innovative framework designed to revolutionize the assessment and understanding of block-based coding through the integration of sophisticated deeplearning models. Dr. Lego combines cutting-edge technologies such as MobileNetV3 (Howard, 2019), for visual recognition and BERT (Devlin et al., 2018), and XLNet (Yang et al., 2019) for natural language processing to offer a comprehensive approach to evaluating coding proficiency.The research methodology involves the meticulous curation of a diverse dataset comprising projects from the LEGO SPIKE app (LEGO Education, 2022), ensuring that the models are subjected to a broad range of coding scenarios. Leveraging the dynamic educational environment provided by the LEGO SPIKE app (LEGO Education, 2022), Dr. Lego empowers users to design and implement various coding projects, fostering hands-on learning experiences.This thesis delves into methodologies aimed at enhancing coding education by exploring model integration, data generation, and fine-tuning of pre-trained models. Dr. Lego not only evaluates coding proficiency but also provides cohesive and insightful feedback, enhancing the learning experience for users. The adaptability of the framework highlights its potential to shape the future of coding education, paving the way for a new era of interactive and engaging learning experiences.

The Breeding Bird Survey (BBS) is a critical North American biodiversity research dataset. However, its raw form presents challenges in geospatial applications. This dissertation unveils the SaveBirds.app, an innovative web tool designed to transform BBS data into actionable insights for conservationists and ecologists. By calculating biodiversity metrics tailored to the specific geospatial areas of interest, SaveBirds.app streamlines the analytical process, enabling researchers to focus on their core expertise without the bottleneck of data manipulation.The application enhances BBS data use by providing user-friendly services to compute essential metrics such as species-specific Rarity Scores and geospatially contextualized biodiversity indices. SaveBirds.app mitigates reliance on data specialists and democratizes access to these metrics, fostering independent and inclusive research across the ecological community.Addressing critical gaps in the field, the SaveBirds app equips users with tools that complement the BBS Online Retrieval System (ORS), simplifies the complexity of raw data analysis, and allows for the rapid calculation of biodiversity metrics based on user-defined geospatial areas. The tool also resolves the need for more pre-fabricated atlases essential for biodiversity studies, often limited to static images, by enabling the swift generation and visualization of dynamic bird distribution atlases.Furthermore, the SaveBirds.app introduces an innovative ArcGIS model that expedites the atlas creation process from months to moments, delivering a suite of atlas types tailored to varying research needs. This capability is a substantial leap forward, conserving time and resources for ecologists.In summary, the SaveBirds.app aims to become a valuable resource for ecological research by facilitating a deeper understanding of avian biodiversity and fostering informed conservation decisions. Its launch paves the way for future integrations, such as climate change variables, exemplifying its potential to reshape ecological research and conservation strategies.

This thesis presents a novel method to control an industrial robotic arm using multiplesensors. This system consists of a hybrid brain activity and vision sensors that convey ahuman being’s intention and visual perception. We fuse and analyze the data from thosesensors using a machine learning-based approach to automatically guide the manipulatorto a designated location. We believe that this Brain–Machine–Interface (BMI) can greatlyalleviate the burdensome traditional method used to program a robot (greatly aids theend-user). We experiment with different modular configurations for the brain activityinformation, i.e., parallelized models and what we refer to as a global model for fusing theinformation. We explore various machine learning and pattern recognition techniques aswell as existing feature selection methods. Our experimental results show that the subjectcan control the robot to a destination of interest using a machine—robot–interface. Weattain accuracy in the order of 99.6% when it comes to the desired motion and 99.8%for the case of deducing the desired characteristic (color) of the targeted object. Theseresults outperform any similar existing approaches that we have researched. Moreover, incomparison to those similar operational systems, we present a unique modular configurationfor brain activity interpretation and object detection mechanism that yields an overallsystem that is highly computationally efficient. Although, in this work, we implementedand demoed our approach using a simple pick and place demo, our work presents the basicstructure underlying a system that can be efficiently used to benefit people with restrictedability to function physically (tetraplegic patients), and allowing them to perform complexand robotics related duties in an industrial setting.

The objective of this thesis is to contribute to the advancement of road surface material identification, which has significant implications for applications like enhanced navigation, traction and stability, predictive maintenance, safety considerations, transportation and infrastructure management, and autonomous car driving. In particular, we study how to accurately identify the various materials used in road surfaces, including asphalt, bricks, chip seal, cobblestone, gravel, or concrete. To achieve this goal, we first compiled a comprehensive database of photos acquired from dashcam videos. The data collection process involves obtaining at least 100 photos for each distinct material class. Then, we carefully annotate the surface material for each dashcam photo. After collecting the extensive set of ground truth images, we focus on applying segmentation techniques to effectively isolate the road surfaces from their surrounding contexts such as buildings, vehicles, or pedestrians. Furthermore, we apply the diffusion method to enrich the training data for all surface material classes. We conducted experiments with different deep-learning models. Our results indicate that our proposed work can recognize the road surface material with a high accuracy rate.

This study aims to incorporate some advantages of computed tomographic data into the chest X-ray deep lung segmentation paradigm. We do this by training a deep convolutional neural network on chest radiographs (a.k.a. X-rays) with manually drawn ground truth and an identical network on radiographs digitally reconstructed from computed tomographic data with ground truth generated for the given computed tomographic image using an automated morphological 3D lung segmentation algorithm. The resulting twin-network ensemble generates pairs of lung image segmentation labels for chest X-rays: 1) a “traditional” segmentation of the lungs encompassing the apparently low-density tissue and 2) a novel, “full” lung segmentation encompassing an expanded view of the lungs’ positionin a chest X-ray including those regions obscured by the heart, ribs, and viscera, in essence, a 2D projection of any portion of the 3D lung. These networks perform consistently, with mean Intersection-Over-Union scores of > 90% and > 95%, respectively, across five trials.By subjective analysis, the proposed lung segmentation approach shows satisfactoryability to generalize onto genuine check X-ray images. The proposed technique’s high performanceand robustness establish a precedent for applying computed tomographic data to automatic chest X-ray segmentation and present an opportunity to further refine existingcomputer-aided detection and diagnostic tools by considering the full lung.

Image is used everywhere since it conveys a story, a fact, or an imagination without any words. Thus, it can substitute the sentences because the human brain can extract the knowledge from images faster than words. However, creating an image from scratch is not only time-consuming, but also a tedious task that requires skills. Creating an image is not a trivial task since it contains rich features and fine-grained details, such as colors, brightness, saturation, luminance, texture, shadow, and so on. Thus, in order to generate an image in less time and without any artistic skills, sketch-to-image synthesis can be used. The reason is that hand sketches are much easier to produce, where only the key structural information is contained. Moreover, it can be drawn without skills and in less time. In fact, since sketches are often simple and rough black and white and sometimes imperfect, converting a sketch into an image is not a trivial problem. Hence, it has attracted the researchers' attention to solve this challenging problem; therefore, much research has been conducted in this field to generate photorealistic images. However, the generated images still suffer from issues, such as the un-naturality, the ambiguity, the distortion, and most importantly, the difficulty in generating images from complex input with multiple objects. Most of these problems are due to converting a sketch into an image directly in one-shot. To this end, in this dissertation, we propose a new framework that divides the problem into sub-problems, leading to generating high-quality photorealistic images even with complicated sketches. Instead of directly mapping the input sketch into an image, we map the sketch into an intermediate result, namely, mask map, through an instance segmentation and semantic segmentation in two levels: background segmentation and foreground segmentation. Background segmentation is formed based on the context of the existing foreground objects. Various natural scenes are implemented for both indoor and outdoor scenes. Then, a foreground segmentation process is commenced, where each detected object is sequentially and semantically added into the constructed segmented background. Next, the mask map is converted into an image through image-to-image translation model. Following this step, a post-processing stage is implemented to enhance the synthetic image further via background improvement and human face refinement. This leads to not only generating better results but also being able to generate images from complicated sketches with multiple objects. We further improve our framework by implementing scene and size sensing. As for size awareness feature, in the instance segmentation stage, the objects' sizes might be modified based on the surrounding environment and their respective size prior to reflect reality and produce more realistic and naturalistic images. Moreover, to implement scene awareness feature in the background improvement step, after the scene is initially defined based on the context and then classified based on a scene classifier, a scene image is first selected. Then, the generated objects are placed on the chosen scene image and based on a pre-defined snapping point to place the objects in their proper location and maintain realism.Furthermore, since the generated images have been improved over time regardless of the input modality, it sometimes becomes hard to distinguish between the synthetic images and genuine ones. Of course, this improves the content and the media, but it is considered a serious threat regarding legitimacy, authenticity, and security. Thus, an automatic detection system of AI-generated images is a legitim need. This system also can be used for image synthesis models as an evaluation tool despite the input modality. Indeed, AI-generated images usually bear explicit or implicit artifacts that result during the generation process. Prior research work on detecting the synthetic images generated by one specific model or similar models with similar architecture. Hence, a generalization problem has arisen. To tackle this problem, we propose to fine-tune a pre-trained Convolutional Neural Network (CNN) model on a special newly collected dataset. This dataset consists of AI-generated images from different image synthesis architectures and different input modalities, i.e., text, sketch, and other sources (another image or mask) to help in the generalization ability across various tasks and architectures. Our contribution in general is two-fold. We first generate high-quality realistic images from simple, rough, black and white sketches, where a newly collected dataset of sketch-like images is compiled for training purposes. Second, since artificial images would have advantages and disadvantages in the real world, we create an automated system that is able to detect and localize synthetic images from genuine ones, where a large dataset of generated and real images is collected to train a CNN model.

This thesis investigates novel approaches to raise the precision and dependability of forecasting models while delving into the intricate realm of bitcoin price forecasting. It proposes a novel method combining trend analysis and manual analysis, with a focus on Bitcoin, a trailblazing cryptocurrency. The process converts daily price fluctuations into binary numbers that represent the mood of the market. After applying manual analysis to these trends, the shared coefficient is obtained that indicates the weighted influence of market patterns on future pricing. Then, a forecasting model that projects Bitcoin values for the eleventh day uses the derivative coefficient. This technique offers a comprehensive view of the dynamics of the bitcoin market by bridging the gap between qualitative trend analysis and quantitative modeling. This thesis not only compares and improves upon the suggested method but also tests it against conventional windowing strategies. The findings provide insight into the fundamental patterns of cryptocurrency pricing and demonstrate the efficacy of the trend-based linear regression and manual analysis technique. It advances the developing area of bitcoin analysis by laying the groundwork for more precise forecasting models and offering insightful information about market behavior.

Sparsity manifests itself in a multitude of modern High-Performance Computing applications including graph neural networks, network analytics, and scientific computing. In sparse matrices, the majority of values are zeros. Traditional methods of storing and processing dense data are unsuitable for the new nature of sparse data, as they end up wasting storage and compute on zeros. Hence, a variety of sparse data formats that store only the non-zero elements were proposed in literature to provide a compact representation of sparse data. Performance of operations on sparse data mainly depends on the sparse data format used for storing the data, as the algorithm needs to closely match the sparse data format. However, choosing the optimal sparse data format for the input sparse matrix is non-trivial, as the optimal format depends on the sparsity pattern of the input sparse matrix. For example, in sparse matrix-vector multiplication (SpMV), for the same input sparse matrix, using different sparse data formats can yield highly variant performance. The best format being the one that closely matches how the non-zeros are arranged within the matrix. Additionally, performance prediction for operations involving sparse matrices is not as straightforward as it used to be for the dense case. For dense computations, dimensions and strides suffice for performance predictions as they provide a sense of the number of floating point operations (FLOPs) to be performed, and how this number compares to the architecture properties (peak FLOPs, number of processing elements, etc.). On the other hand, sparse matrix dimensions do not directly convey useful information about the total number of operations to be performed, since the majority of elements are zeros and do not contribute to the total number of FLOPs. Moreover, existing work on sparse operations optimizations mainly depends on a discrete set of matrices, limiting the ability to generalize observations. To address these challenges, we identify the sparsity pattern as the main driving factor for performance. First, we propose an extensible classifier framework to automatically identify the sparsity pattern of the input sparse matrix. This framework uses graph neural networks (GNNs) by representing the input sparse matrix as a graph, and then learning the structural relationship between nodes in the matrix (graph). Our framework achieves up to 98% classification accuracy on full graphs, the same accuracy for scrambled matrices, and 92% accuracy for small random subsamples taken out of original graphs. Second, we use graph models as a proxy to generate large-scale synthetic sparse matrices. We propose another modular framework to study the correlation between the graph model parameters, and the structure of the resulting graph and its tolerance to noise during the generation step. Third, we also use graph models for a performance evaluation framework that can assist in finding the best sparse format for a given graph model on a given architecture, utilizing the graph model parameters as a representative set of features to predict performance, and providing a more robust way of visualizing sparse matrix operations performance. This framework also takes into consideration noise in the matrix generation step, and evaluates the extent of noise to which performance can still be predictable based on the graph model parameters. Our results show that sparse computations need richer models to categorize sparse matrices in terms of structure, study the sensitivity of such models even for one structure, and tie performance to a more descriptive set of parameters.

In a complex and information-rich world, how do humans and other animals learn and make decisions with limited cognitive resources? This thesis aims to answer this question by presenting a novel theoretical framework for understanding how capacity-limited agents should allocate their cognitive resources to balance the pursuit of reward with the burden of cognitive cost. Specifically, I formalize the idea of policy compression, or the simplification of action policies to reduce their representational cost. Policy compression marries reinforcement learning with information theory by representing policies as capacity-limited channels that transmit information about state to guide action selection. The upper bound on capacity results in a resource-rational trade-off between reward and the complexity cost of representing one's action policies. In this body of work, I employ computational modeling, behavioral experiments, and evidence from lesion studies to explore the implications and applications of policy compression to illuminate how biological agents act with limited cognitive resources. First, I introduce the theory and show how policy compression can account for a wide range of behavioral phenomena, from stochasticity and perseveration to state and action chunking. Next, I provide experimental evidence that human decision making indeed reflects a balance between reward maximization and policy compression, resulting in systematic choice biases towards frequently used actions. I show that this behavioral pattern is uniquely explained by the policy compression model, which outperforms competing models that incorporate memory constraints in instrumental learning. I then introduce the notion of conditional policy compression, which expands the theory to include reductions in cognitive cost by conditioning on additional information. Conditional compression explains why action chunking reduces policy complexity without sacrificing reward. I confirm these predictions experimentally, showing that people employ chunking as an adaptive strategy to reduce cognitive load by exploiting temporal structure in the environment. Finally, I search for potential neural substrates that implement cost-sensitive action selection. Using evidence from lesion studies in motor sequence learning, I propose a computational division of labor between the dorsolateral striatum (DLS) and dorsomedial striatum (DMS), two areas of the basal ganglia that balance automatic, history-dependent action selection with costly, flexible behavior. By using a default policy learned by DLS to “regularize” the reward-maximizing policies learned by DMS, the brain achieves a balance between robustness and flexibility in complex behaviors such as motor learning. By taking a resource-rational approach to decision making---that is, by specifying how capacity-limited agents should allocate their computational resources---policy compression offers a new normative lens through which to interpret the peculiarities of behavior and the functional organization of brain areas.

Probabilistic models provide a principled way to model observed data, perform statistical inferences, and express uncertainty over latent variables. However, it remains computationally challenging to fit high-dimensional probabilistic models to data. In this dissertation, we provide a suite of approaches for reducing this computational burden for various probabilistic models of interest. We examine models along the entire spectrum of structured to flexible, tackling diverse examples such as sparse Bayesian learning, latent Gaussian models, log-concave densities, and mixture models. Our methodologies are a rich combination of mathematical developments that leverage tools spanning optimization and statistics, as well as computational advances such as parallel computing and automatic differentiation.

Urban residents can be exposed to environmental harms that can contribute to chronic illness and/or a decreased quality of life. These harms are currently most typically measured by geographically sparse regulatory monitors, making it difficult to determine the areas and residents most affected given the fine-grained heterogeneity of these harms. Simulation and modelling are frequently used to address this issue, but these approaches result in residual uncertainty due to simplifications and estimates in proxies used (such as omitting highway elevations and making assumptions about pollutants in residential versus industrial areas). Dense low-cost sensor network deployments can provide localized data to improve citywide environmental monitoring models, identify urban hotspots, and promote environmental justice. However, the successful leveraging of these networks is dependent on strategically selecting where to place sensor nodes, a step that is often skipped in urban deployments and remains an open question for network designers. Knowing how to strategically place sensors requires a metric against which to measure success. In Computer Science, node placement strategies often focus on optimizing for area coverage, but this does not accommodate for the three-dimensionality or social nature of cities. This thesis attacks this challenge via the development of quality metrics that account for the three-dimensional nature of cities, using open data to estimate how well nodes are distributed based on different parameters. The primary goal is to develop metrics that focus on key aspects needed for successful network deployments based on the constraints, criteria, sensor placement locations, and design strategies used in prior real-world urban sensor network deployments. The findings from these deployments point to three main themes for new quality metrics that are explored in this thesis: network reliability, "representativeness", and social equity. This thesis first presents quality metrics for network reliability, using data from a low-cost, large-scale urban sensor network deployment to build models that predict connectivity and power issues at single node locations. The best performing models have 75% accuracy and 77% accuracy for connectivity and power, respectively, showing promise for a metric that can accurately predict locations without reliability issues. The thesis next lays the foundation for a quality metric that estimates "data representativeness", or sensor reading utility, by incorporating data from a large-scale urban sensor network, field work experiments, and open data about urban form and land use. The findings show that sensors on certain road types can potentially provide useful data up to 750 meters away but other road types may need more dense sensor placement to account for variability. Finally, to determine how well sensor networks can be designed to achieve social equity goals, the thesis examines differences in commonly used inequality metrics by combining US census data with simulated node placements. The results show that inequality metrics based on the minimum distance to a sensor node are not well suited for urban sensor network design because they show similar levels of inequality regardless of the number of sensors deployed. Thus, new metrics or parameters must be used to evaluate sensor networks for equity goals. This thesis asks the critical question of how low-cost urban environmental sensor networks should be designed and evaluated to determine whether they are well designed for the people they aim to serve. Through the development of novel analyses and proposed quality metrics, this work highlights the power of using open data sources about various urban features in laying the foundation to answer that question. By combining open data with findings from field experiments and a real-world sensor network deployment, methods and findings are provided to allow network designers, city employees, and residents to consider and evaluate network designs in new ways. Thus this thesis will aid in the deployment of long-term, reliable, equitable urban environmental sensor networks, contributing to the future development of smart, healthy, sustainable cities.

While the major bodies that make up the solar system, including planets, comets, the Moon, and the Sun, have been known since antiquity, our knowledge of asteroids is relatively recent. The discovery of Ceres in 1801 both revolutionized our understanding of the solar system and launched a new era in the study of its dynamics. Since then, asteroids have been a source of fascination for both amateur and professional astronomers. Indeed, the study and characterization of asteroids, especially those in the main belt, has been extraordinarily scientifically productive, yielding insights into the dynamical properties of the solar system, the histories of planetary bodies, and compositional gradients in the young protoplanetary disk. Most recently, with the advent of new astronomical surveys that exhibit greater astrometric precision, phase coverage, and an increased number of observed asteroids, combined with advancements in dynamical methods and improvements in machine learning and computation, we can reanalyze the asteroid population. This offers a new, modern perspective on the properties and dynamics of the solar system’s smallest members. This thesis presents several such analyses. We first examine the potential for distant solar system bodies to be detected gravitationally through observations of small bodies, planets, and spacecraft. We then turn our attention to the modeling of visual albedo across the main belt—we take advantage of the correlation of the albedo with asteroid proper orbital elements to make predictions about the albedos of asteroids yet to be observed, extending existing catalogs by nearly a factor of five. Third, we examine the prospects of measuring asteroid masses by taking advantage of Gaia’s precise astrometry and modeling dynamical scattering between asteroids in the main belt. Finally, we consider the relationship of an asteroid’s taxonomy to its phase curve and find that certain inferences about asteroid spectra can be made solely using well-sampled phase curves.

Depositing and borrowing in a currency other than the local currency is a well- documented phenomenon. The associated exchange rate and monetary risks, pivotal during the Asian Financial Crisis of 1997, prompted an academic discourse. Financial dollarization literature explains why individuals and corporations domiciled in emerg- ing markets deposit and borrow in hard currencies, mainly in U.S. dollars. This thesis proposes a model for cryptocurrency lending and assesses model predictions with a data set containing more than one million Ethereum transactions. The computational modelling and statistical analysis show that the main theoretical explanations provided by financial dollarization literature may not be directly transferable to cryptocurrencies. The results suggest that the popularity of depositing and borrowing in cryptocurrencies must be largely motivated by other factors.

Differentiation, the process of individualizing education, remains a difficult endeavor that can overwork teachers and negatively affect students. Due to the tremendous upside of differentiation, this thesis examines the possibility of predicting student performance using Cross-Network LSTMs with additional data sources. After modifying the attention mechanism, the model proposed showed an improved performance in prediction with the inclusion of additional measurements already in the Junyi Academy Online Learning Activity Dataset. The modifications also showed the ability to separate problems into course units based on their similarity scores, and quickly make long term predictions. The development of this network as recommender system for differentiation is discussed as well as a working example using the AP Calculus BC curriculum.

Text-to-image models, often associated with Generative AI techniques, have faced several challenges that have become synonymous with the broader field of Computer Vision and Machine Learning. Generating realistic images from textual descriptions is a complex task. While recent models have made significant progress, achieving photo realism remains a challenge. Understanding and translating natural language descriptions into meaningful visual representations is inherently difficult. Many text-to-image models struggle to produce images with sufficient resolution and clarity, especially for complex scenes or objects. Assessing the quality of these generated images remains a challenging task due to the perception of image quality. Furthermore, generated images lack clear ground truth and higher details for human attention. To address this challenge in the study of artificially generated images, we are introducing a novel approach with a dataset, Generative Artificial Image Assessment (GAIA) trained with architecture to define a definitive reference point. This comprises images from eight popular text-to-image AI models as well as user rankings from a crowd-sourced annotation. This dataset is evaluated and predicted by pre-trained state-of-the-art networks with ranking classes and diverse regression techniques in a comparative study to analyze the images. The approach is a combination of objective evaluation metrics, subjective human judgment, benchmark datasets with diverse ground truth annotations, and advancements in multimodal learning techniques giving us a way forward to advancing the field of text-to-image generation.

Reactive programming has emerged as a pivotal paradigm in modern mobile application development, enabling developers to handle asynchronous events and data streams seamlessly. The choice of reactive programming framework can significantly impact the performance, maintainability, and scalability of native mobile applications. In this research, we aim to conduct a comprehensive comparative study of reactive programming choices in native mobile development. Our research reveals significant performance disparities among various reactive programming frameworks in mobile application development. Specifically, Combine outperforms RxSwift in iOS environments, offering more efficient load times. Like- wise, Coroutines demonstrate a marked performance edge over RxJava in Android development, excelling in load time across multiple test scenarios. This consistent advantage positions Coroutines as a highly favorable choice for native Android app development, due to its efficiency and effectiveness. An intriguing aspect of our study is the performance of the Delegate Pattern in iOS development, which, despite its simplicity and lack of a reactive programming framework, surpasses both Combine and RxSwift in load time, memory consumption, and CPU usage.

The increasing utilization of intelligent agents has led to a growing demand for policy frame- works to govern their behavior. AOPL (Authority Obligation Policy Language) stands as a well- established policy framework within the logic-based AI (Artificial Intelligence) community. This thesis evaluated the suitability and effectiveness of AOPL by comparing it against other prominent policy frameworks. The evaluation assessed their capacity to effectively express policies, including role-based policies, and their ability to analyze policies. The methodology comprised a series of literature policy scenarios designed to test the scalability and efficiency of the policy frameworks. Additionally, the AOPL framework was extended with missing features identified through comparison with EPA (Expressive Policy Analysis) and other logic-based policy frameworks. The outcome of this research contributed to a better understanding of the strengths and limitations of these policy frameworks, providing valuable insights for the development of future policy-based intelligent agent systems.

Autonomous machines such as vehicles, drones, and robotic manipulators promise to transform the world by unleashing humans from repetitive, dangerous, and labor-intensive tasks. However, their widespread deployment requires advances in safety, real-time performance, and resilience. This thesis tackles these key challenges through end-to-end system optimization that maintains safety while improving performance and fault tolerance. We optimize the entire Perception-Planning-Control (PPC) computing pipeline that takes in sensor readings and output control commands. First, this thesis develops a model to quantify the perception processing rate requirements for safe autonomous driving in complex scenarios, connecting the varying real-time latency requirements with the operating scenarios. Second, we accelerate the time-consuming 3D mapping in perception. A specialized accelerator is designed that achieves substantially higher throughput and energy efficiency over a CPU, enabling real-time perception for 3D mapping. Third, we accelerate the computationally expensive optimization-based motion planning algorithms with a variable precision search method that reduces memory bandwidth pressure without sacrificing positional and orientational precision. Lastly, we assess autonomous machines' fault tolerance characteristics against real-world noises and errors to generate reliable control commands. We propose a fault characterization framework that evaluates the impact of silent data corruptions (SDCs) on application-level metrics. To mitigate SDCs, we propose lightweight anomaly detection techniques to recover failures in the computing pipeline with insignificant overhead. This dissertation enables the development of safe, real-time, and resilient autonomous machines. The contributions chart a path toward robust deployment of autonomous machines that can transform society.