Theses and Dissertations from UMD
Permanent URI for this communityhttp://hdl.handle.net/1903/2
New submissions to the thesis/dissertation collections are added automatically as they are received from the Graduate School. Currently, the Graduate School deposits all theses and dissertations from a given semester after the official graduation date. This means that there may be up to a 4 month delay in the appearance of a give thesis/dissertation in DRUM
More information is available at Theses and Dissertations at University of Maryland Libraries.
Browse
5 results
Search Results
Item BAYESIAN INFERENCE OF LATENT SPECTRAL AND TEMPORAL NETWORK ORGANIZATIONS FROM HIGH DIMENSIONAL NEURAL DATA(2022) Rupasinghe, Anuththara; Babadi, Behtash; Electrical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)The field of neuroscience has striven for more than a century to understand how the brain functionally coordinates billions of neurons to perform its many tasks. Recent advancements in neural data acquisition techniques such as multi-electrode arrays, two-photon calcium imaging, and high-speed light-sheet microscopy have significantly contributed to this endeavor's progression by facilitating concurrent observation of spiking activity in large neuronal populations. However, existing methods for network-level inference from these data have several shortcomings: including undermining the non-linear dynamics, ignoring non-stationary brain activity, and causing error propagation by performing inference in a multi-stage fashion. The goal of this dissertation is to close this gap by developing models and methods to directly infer the dynamic spectral and temporal network organizations in the brain, from these ensemble neural data. In the first part of this dissertation, we introduce Bayesian methods to infer dynamic frequency-domain network organizations in neuronal ensembles from spiking observations, by integrating techniques such as point process modeling, state-space estimation, and multitaper spectral estimation. Firstly, we introduce a semi-stationary multitaper multivariate spectral analysis method tailored for neuronal spiking data and establish theoretical bounds on its performance. Building upon this estimator, we then introduce a framework to derive spectrotemporal Granger causal interactions in a population of neurons from spiking data. We demonstrate the validity of these methods through simulations, and applications on real data recorded from cortical neurons of rats during sleep, and human subjects undergoing anesthesia. Finally, we extend these methods to develop a precise frequency-domain inference method to characterize human heart rate variability from electrocardiogram data. The second part introduces a methodology to directly estimate signal and noise correlation networks from two-photon calcium imaging observations. We explicitly model the observation noise, temporal blurring of spiking activities, and other underlying non-linearities in a Bayesian framework, and derive an efficient variational inference method. We demonstrate the validity of the resulting estimators through theoretical analysis and extensive simulations, all of which establish significant gains over existing methods. Applications of our method on real data recorded from the mouse primary auditory cortex reveal novel and distinct spatial patterns in the correlation networks. Finally, we use our methods to investigate how the correlation networks in the auditory cortex change under different stimulus conditions, and during perceptual learning. In the third part, we investigate the respiratory network and the swimming-respiration coordination in larval zebrafish by applying several spectro-temporal analysis techniques, on whole-brain light-sheet microscopy imaging data. Firstly, using multitaper spectrotemporal analysis techniques, we categorize brain regions that are synchronized with the respiratory rhythm based on their distinct phases. Then, we demonstrate that zebrafish swimming is phase-locked to breathing. Next, through the analysis of neural activity and behavior under optogenetic stimulations and two-photon ablations, we identify the brain regions that are key for this swimming-respiration coordination. Finally, using the Izhikevich model for spiking neurons, we develop and simulate a circuit model that replicates this swimming-respiration coupling phenomenon, providing new insights into the possible underlying neural circuitry.Item A Bayesian Framework for Analysis of Pseudo-spatial Models of Comparable Engineered Systems With Application to Spacecraft Anomaly Prediction Based on Precedent Data(2017) Ndu, Obibobi Kamtochukwu; Mosleh, Ali; Modarres, Mohammad; Reliability Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)To ensure that estimates of risk and reliability inform design and resource allocation decisions in the development of complex engineering systems, early engagement in the design life cycle is necessary. An unfortunate constraint on the accuracy of such estimates at this stage of concept development is the limited amount of high fidelity design and failure information available on the actual system under development. Applying the human ability to learn from experience and augment our state of knowledge to evolve better solutions mitigates this limitation. However, the challenge lies in formalizing a methodology that takes this highly abstract, but fundamentally human cognitive, ability and extending it to the field of risk analysis while maintaining the tenets of generalization, Bayesian inference, and probabilistic risk analysis. We introduce an integrated framework for inferring the reliability, or other probabilistic measures of interest, of a new system or a conceptual variant of an existing system. Abstractly, our framework is based on learning from the performance of precedent designs and then applying the acquired knowledge, appropriately adjusted based on degree of relevance, to the inference process. This dissertation presents a method for inferring properties of the conceptual variant using a pseudo-spatial model that describes the spatial configuration of the family of systems to which the concept belongs. Through non-metric multidimensional scaling, we formulate the pseudo-spatial model based on rank-ordered subjective expert perception of design similarity between systems that elucidate the psychological space of the family. By a novel extension of Kriging methods for analysis of geospatial data to our "pseudo-space of comparable engineered systems", we develop a Bayesian inference model that allows prediction of the probabilistic measure of interest.Item Choice Modeling Perspectives on Social Networks, Social Influence, and Social Capital in Activity and Travel Behavior(2015) Maness, Michael; Cirillo, Cinzia; Civil Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Understanding the determinants of activities and travel is critical for transportation policymakers, planners, and engineers to design and manage transportation systems. These systems, and their externalities, are interwoven with social systems in communities, cities, regions, and societies. But discrete choice models - the predominant modeling tool for researching travel behavior and planning transportation systems - are grounded in theories of individual decision-making. This dissertation expands knowledge about the incorporation of social interactions into activity-travel choice models in the areas of social capital and social network indicators; social influence motivations and informational conformity; and misspecification errors from social network data collection. Incorporating social capital into activity choice models involves using social capital indicators from surveys. Using a position generator question type, the role of social network occupational diversity in activity participation was explored and the performance of models using name generator and position generator data was compared. Access to the resources embedded in diverse networks (extensity) was found to positively correlate with leisure activity participation. Compared to core network indicators from name generators, position generator indicators were typically better at predicting activity participation in a cross-validation study. Current models of social influence in travel do not account for varying motivations for social influence such as for accuracy, affiliation, and self-concept. To test for an accuracy motivation, a latent class discrete choice model was formulated that places individuals into classes based on information exposure. Contrasting with existing work, this model showed that "more informed" households are more likely to own bicycles due to preference changes causing less sensitivity to smaller home footprints and limited incomes. A Bayesian prediction procedure was used to derive distributions of local-level equilibria and social influence elasticity. The effect of errors in social network data collection using name and position generators is not fully understood for choice models. In a case study, the social network occupational diversity measure was robust to varying position generator lengths. Simulation experiments tested the implications of social network structure, misspecification, and small samples on social influence choice models where sample size, social influence strength, and degree of misspecification had the greatest impact.Item IMPROVED PROBABILISTIC LIFE ESTIMATION IN ENGINEERING STRUCTURES: MODELING MULTI-SITE FATIGUE CRACKING(2014) Al Tamimi, Abdallah Moh'd AR. Sh. M.; Modarres, Mohammad; Reliability Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)The purpose of this paper is to investigate the effect of fatigue, in the presence of neighboring cracks, and to integrate that into a probabilistic physics of failure based model that could be used to predict crack growth. A total of 20 fatigue experiments were performed at different loading conditions using dog-bone samples of API-5L grade B carbon steel containing neighboring cracks. The fatigue testing was conducted to generate the data needed for the probabilistic fatigue life prediction model development. Moreover, these experiments have investigated the impact of both neighboring cracks dimensional variability and the loading conditions on cracks interaction, coalescence and growth. The experiment layout was designed to improve some of the existing experimental layouts presented in the literature. Moreover, a new approach for measuring the neighboring cracks depth and the associated number of cycles in dog-bone shaped samples using different microscopy tools and image-processing techniques was proposed. On the other hand, simulation efforts were also performed to assess the Stress Intensity Factor (SIF) around neighboring cracks. Models discussing how the SIF of single semi-elliptical crack could be corrected to account for the neighboring cracks interaction were discussed in order to better understand the fatigue behavior. A combination of these models was integrated to compute the SIF values necessary for the probabilistic life prediction modeling purposes. Also, a new strategy for investigating ligament failure by detecting when it occurs rather than how it occurs was developed in this work. A demonstration of an improved understanding of the impact of different loading conditions on the ligament failure phenomena both using experiments and simulation was also discussed. Finally, a multi-site fatigue crack growth rate model was developed and its parameters including their uncertainties were estimated. A Bayesian approach was adopted to perform uncertainty characterization and model validation.Item An Integrated Methodology for Assessing Fire Simulation Code Uncertainty(2010) Ontiveros, Victor Luis; Milke, James A.; Fire Protection Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Fire simulation codes are powerful tools for use in risk-informed and performance-based approaches for risk assessment. Given increasing use of fire simulation code results, accounting for the uncertainty inherent in fire simulation codes is becoming more important than ever. This research presents a "white-box" methodology with the goal of accounting for uncertainties resulting from simulation code. Uncertainties associated with the input variables used in the codes as well as the uncertainties associated with the sub-models and correlations used inside the simulation code are accounted for. A Bayesian estimation approach is used to integrate all evidence available and arrive at an estimate of the uncertainties associated with a parameter of interest being estimated by the simulation code. Two example applications of this methodology are presented.