48440 research outputs found
Sort by
Time Series Prediction Using Deep Learning Methods
No full textThis report examines and compares the behaviors of Long Short-Term Memory (LSTM), Transformer, and Auto-Regressive Fractionally integrated Moving Average (ARFIMA) models in time-series forecasting. The models are applied to predict the BOIL leveraged ETF and analyze data gathered from a microtubule-kinesin active matter experiment. The comparisons revealed the impacts of increasing training data, introducing noise into the training data, and evaluates model convergence, outlier predictions, and model complexity. This report includes conceptual overviews and technical explanations for each model. The complete MATLAB code for the LSTM and Transformer models as well as R code for the ARFIMA models, is provided in the appendix
Deep Learning Analysis of Neuroimaging (MRI) Data
No full textOne of the most challenging and crucial aspects of patient care and clinical research for Alzheimer’s disease is its timely and reliable detection. In this study, we have expanded on using 3D convolutional neural networks to differentiate between subjects who are cognitively normal, have mild cognitive impairment (MCI), and Alzheimer’s dementia using structural MRI images. We use additional data to train the model, and a larger validation and testing set from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). The model is trained and validated using data from ADNI. Compared to a reference model using the volumes and thickness of previously reported brain regions associated with the progression of Alzheimer’s disease and traditional Atlas based methods, using a deep-learning model proves to be much faster. The deep-learning model is expected to be accurate and achieve a higher area-under-the-curve (AUC) than 85 (from the previous study) when distinguishing between subjects who are cognitive normal and have mild cognitive impairment or Alzheimer’s disease. Previously, the model reported an AUC of 62.45 in detecting MCI, which we hope to increase when implementing the model using a larger dataset. When compared to a reference model based on the volumes and thickness of reported brain regions associated with MCI and Alzheimer’s, this model proves to provide quicker results. The model’s ability to detect brain regions impacting MCI specifically allow for the early detection of Alzheimer’s disease. Forecasting the progression of Alzheimer’s disease would play a pivotal role in clinical research to detect Alzheimer's disease early
Leicester Cemetery E Term Project
No full textThe aim of this project is to improve the accessibility, accuracy, and visibility of historically relevant information regarding the Quaker Cemetery. In collaboration with the Computer Science Department, I developed customizable code to extract data from FindAGrave.com, an open-source database of cemetery records. This enabled the identification and correction of inconsistencies in existing burial records. Leveraging these datasets, I created a new Wikipedia page and contributed significant refinements to multiple related articles, including new images of the Quaker Cemetery, ensuring that accurate information is available online. Additionally, I was able to expand and verify a list of Quaker Cemetery interments on-site by comparing observable marked burials with existing records
Stock Market Simulation 2504
No full textThe goal of this IQP project was to gain an understanding of the stock market: its history, purpose, risk management, and trading strategies that can be used to benefit from it (using technical analysis). It compared two trading strategies: the buy and hold and swing trading methods in a 5-week simulation. Seven companies were selected to form the portfolios. The initial principal was $100,000 for each portfolio. The swing trading strategy had 1.7% return while the buy and hold strategy had 10.2% returns. The buy and hold strategy was shown to be the better of the two strategies. The experiences learned from this project will be useful for future investment
Rethinking Livability in Venice: Addressing Overtourism through Island-Level Indicators
No full textWith the surge of tourists continuing to pour into the Historic City of Venice year after year, it is vital to evaluate the livability of the Venetian Islands. To determine this, we have synthesized seven quantifiable indicators, including affordability, resident services, green space, tourism, user density, accessibility, and reachability. This ultimately allowed us to develop a variety of graphs to determine the livability of all islands within the Historic City of Venice, update Venipedia pages on the Inner City Islands, Major Lagoon Islands, and Minor Lagoon Islands, and to develop a website with a quiz to determine the island that is most livable for an individual. Through the analysis of our indicators, we determined that Venice is a city of balances. There is no singular island that can be considered the most livable, as the livability of an island is determined by the personal needs of the individual, which vary for everyone based on their particular circumstances
Walks of Life: Chronicling Pedestrian Experience and Street Purpose in Venice
No full textVenice’s street network, composed of over 3,100 streets spanning over 330 km, contends with unprecedented amounts of congestion caused by high amounts of tourist traffic. By pursuing this project, we seek to understand how the physical state of streets interacts with pedestrian experience. We use ArcGIS to compile geometric information about the historic center’s streets and plot locations of amenities across six of Venice’s largest campi (public squares). In addition, we conduct pedestrian counts at several main arterial streets at multiple times of day using manual and AI-automated methods, determining the amounts of local and tourist pedestrians as well as the amounts of pedestrians heading in each direction. To account for different street sizes, we normalize pedestrian counts with a pedestrian flow intensity metric. We find that key amenities that serve the public are few in number, while commercially-oriented amenities occupy a significant amount of walkable space in campi. Pedestrian count data indicate that locals’ peak hours on streets are in the morning and evening and tourists’ are in the afternoon, and that pedestrians migrate toward tourist hotspots. Applying the normalizing metric, we observe that flow intensity tends to be far greater on narrower streets. We recommend that future endeavors be dedicated to expanding the scope of amenity plotting on both campi and traditional streets, as well as expanding upon our body of pedestrian count data at more streets and utilizing video cameras and AI-automated software on a larger scale. The implementation of “blue streets” that bar access to tour groups and more robust signage throughout the city seems to be an idea with promise, but further targeted research efforts are needed
High-Precision Robotic Assembly and Packing in the Presence of Uncertainties and Unknowns
No full textHigh-precision robotic manipulation tasks, such as robotic assembly tasks or dense packing tasks, are contact-rich and have to overcome the effects of various uncertainties stemming from inaccuracies in modeling, sensing, perception, and robotic motion control. Efficient and robust robotic assembly and dense packing of complexshaped parts are critical, open challenges in advanced manufacturing and logistics, since existing solutions are often limited to very specific objects or tasks. This dissertation presents a perception-based, online framework that enables autonomous robotic and human-robot collaborative (HRC) systems to perform tightclearance assembly under six-dimensional uncertainties. Additionally, it introduces a perception-driven, online approach for a robot manipulator to autonomously and densely pack unknown objects that arrive continually at a packing site into a container to maximize space utilization and overcome the effects of various uncertainties. The proposed methods are extensively validated through experimental evaluations on complex-shaped multi-peg-in-hole assembly tasks and dense packing problems involving unknown objects of varied shapes with tight clearances. The contributions of this research include designing, implementing, and testing the following frameworks and systems: - A robust and general autonomous robotic assembly approach for assembly involving arbitrary, complex-shaped parts, including multi-peg-in-hole tasks, under significant pose uncertainties exceeding ten times the task clearances. - A general, geometry-invariant representation of parts and novel optimization algorithm for online contact configuration estimation. - A flexible robotic assembly framework that enables arbitrary picking and insertion directions for assembly within the robot’s workspace. - A human-robot collaboration framework that combines intuitive human guidance with autonomous robotic precision for tight-clearance assembly, without requiring operator expertise in robotics. - A real-time, perception-driven framework for concurrent object detection, modeling, grasp planning, packing planning, and motion execution under uncertainty. - A novel online 4-DoF dense packing planning method that continuously optimizes object placement sequences and poses for high packing density. By addressing the challenges of uncertainty in high-precision robotic manipulation tasks of complex objects, this dissertation aims to push the envelope in robotics and automation towards broader applications of robotic systems in complex, real-world environments
Vanity - A Live Coding Language for Unreal Engine
No full textGames engines present new opportunities for digital arts practitioners. In the performance practice of live coding, performers create audiovisual works by programming them live on stage, and projecting source code for the audience to follow. This research presents Vanity, a new programming language and environment that brings the graphical capabilities of Unreal Engine 5 to live coding performance; to the best of our knowledge it is the first live coding system to run inside of Unreal Engine. We discuss the design of the Vanity language, the implementation of the associated live coding environment in Unreal Engine, and the results of a user study exploring the usability and creative potential of our system
Predictive and Causal Models of the Impact of Pandemics on Cardiovascular Disease Patient Biomarkers
No full textPandemic-induced disruptions to routine healthcare and lifestyle changes in cardiovascular diseases (CVDs) patients triggered changes in critical CVD biomarkers (measurable parameters of the body that can indicate health or illness). Prior work has overlooked models for predicting these biomarker trajectories or modeling causality during pandemics. Utilizing a first-of-a-kind Electronic Health Record (EHR) dataset of over 400,000 patients treated at the UMass Memorial hospital before and during the Covid pandemic, this doctoral dissertation created ML predictive and causal models for evaluating the COVID-19 pandemic’s impact on CVD patient biomarkers. To create these models, this dissertation explored three methodological approaches. The first explored traditional ML models on EHR data attributes, to predict the impact of the COVID-19 pandemic on CVD patient biomarker (BP, LDL cholesterol, HbA1c and BMI) trajectories and ML causal analysis using the Debiased ML for Difference-in-Differences approach. To interpret model results, SHAP values were calculated on selected ML models. Study results revealed CATBoost and XGBoost showed the best performance for predicting LDL cholesterol and HbA1c, with scores of 0.13 and 0.10. Random Forest performed best for BMI and blood pressure, with values of 0.192 and 0.071. Key features influencing biomarker changes included age, socioeconomic status, and race/ethnicity, underscoring the impact of social determinants of health. causal analysis revealed a significant rise in BMI and systolic BP (p < 0.05) among CVD patients during the COVID-19 pandemic, while HbA1c and LDL cholesterol improved, indicating varied pandemic effects on different biomarkers. However, as evidenced by the Low values, traditional ML models were limited in effectively capturing pandemic impact. This was due to the pervasiveness of categorical input variables and high variability in numerical biomarker levels. To address limitations of the first work including low predictive capacity of traditional ML models (low values, high MSE values), we explored the Genetic Algorithm Neural Architecture Search (GANAS) for automated, large CVD DL model design, and to enhance the low predictive capacity of traditional ML models. NAS is an automated approach for DL model design and determining optimal hyperparameter values, obviating the need for expertise. For the first time, we applied GA, an optimization technique inspired by the principles of natural selection and evolution, to optimize NAS and generate a CVD biomarker DL model. GANAS outperformed existing automated DL model design approaches such as ENAS, DARTS, and traditional ML models in predicting both HbA1c and LDL cholesterol levels. For HbA1c, it achieved 0.9739 accuracy, 0.9485 precision, 0.9739 recall, and a 0.9610 F1 score. For LDL cholesterol, it maintained strong results with 0.9117 accuracy, 0.9143 precision, 0.9100 recall, and a 0.9120 F1 score. SHAP analysis identified socioeconomic status—especially middle-class—and the pre/post-pandemic period as key predictors. The final study aimed at providing more robustness in design by addressing the lack of uncertainty estimation, exploiting temporal data relationships and capturing biomarker interdependency by via multi-target predictions. To do this, we created BMT-CB, a Bayesian multi-target Transformer model (BMT) that combines Bayesian variational inference (BVI) and multi-target prediction with ClinicalBERT (CB), a pretrained BERT-based transformer architecture to jointly predict multiple CVD biomarkers at the first pandemic-onset clinical visit. The model leverages BVI to estimate uncertainties, embeddings to capture temporal and patient characteristics and a DeepMTR model to capture biomarker interdependencies. BMT-CB outperformed other models by having the lowest mean value with an MAE of 0.00887, RMSE of 0.0135 and MSE of 0.00027. Uncertainty estimation revealed the model was able to effectively capture uncertainty and patient interdependency during prediction This research will advance predictive and causal modeling of pandemic impacts on CVD patients and facilitate tools for early detection of health deterioration, uncovering latent trends, and forecasting pandemic-induced health risks. Such tools could be utilized by healthcare professionals, epidemiologists, and public health policymakers for more informed decision-making and targeted interventions during pandemics
Design and Testing of Rotational Speed Sensor with Zero Latency for Unmanned Ground Vehicle
No full textThis report presents the design and testing of a novel, zero-latency Hall-effect-based wheel speed sensor aimed at enhancing agile vehicle dynamics in unpredictable terrain conditions. Agile vehicle dynamic systems rely on quick sensor feedback to detect and respond to changes, such as wheel slippage. Traditional wheel speed sensors, including those utilizing the Hall effect, face limitations in their response time from mechanical constraints, with latencies that can delay corrective actions in real-time. The focus is to improve the sensor's mechanical design by exploring an innovative spiral-shaped reluctor ring, which reduces signal acquisition delays compared to conventional systems. This report details the sensor's design, simulation using Ansys Maxwell, and the integration of mechanical components for real-world testing. Key aspects covered include sensor design, rapid prototyping, experimental setup, and performance analysis, highlighting the potential for this technology to provide faster, more accurate data for measuring wheel speed for an unmanned ground vehicle