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
8 results
Search Results
Item Spectral Methods for Neural Network Designs(2022) Su, Jiahao; Huang, Furong; Electrical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Neural networks are general-purpose function approximators. Given a problem, engineers or scientists select a hypothesis space of functions with specific properties by designing the network architecture. However, mainstream designs are often ad-hoc, which could suffer from numerous undesired properties. Most prominently, the network architectures are gigantic, where most parameters are redundant while consuming computational resources. Furthermore, the learned networks are sensitive to adversarial perturbation and tend to underestimate the predictive uncertainty. We aim to understand and address these problems using spectral methods --- while these undesired properties are hard to interpret from network parameters in the original domain, we could establish their relationship when we represent the parameters in a spectral domain. These relationships allow us to design networks with certified properties via the spectral representation of parameters.Item Microbial Induced Corrosion in Oil Pipelines(2020) Farzaneh, Azadeh; Al-Sheikhly, Mohamad MA; Material Science and Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Crude oil pipeline failure due to corrosion processes is a global issue with detrimental effects on the environment and economy. More than 10,000 oil spills occur in the United States alone, each year. These oil spills are so prevalent that they have become the rule rather than being a 1-time incident. Many of these oil spills happen as a result of pipeline failure due to corrosion. Microbial-Induced Corrosion (MIC) accounts for 20% of the total number of pipeline corrosion incidents. Therefore, the mechanisms involved and especially in the case of microbial corrosion must be studied and elucidated.Sulfate-Reducing Bacteria (SRB) are the main culprits of MIC. The first suggested mechanism in 1930’s related high corrosion rates in buried pipelines to SRB hydrogen utilization and depolarization of the cathodic area on the metal surface. Despite its numerous flaws, it remained the most widely accepted mechanism of MIC. In 2004, a new mechanism called direct electron uptake was suggested for MIC. It related corrosivity of bacteria to direct electron uptake from metallic iron. This mechanism is not fully understood hitherto. Only a few bacteria have been isolated so far that demonstrated direct electron uptake capabilities. Most of the research has been focused on these few isolates. However, if direct-electron uptake is the main MIC mechanism, other SRB strains should possess similar capabilities. This work investigated the possibility of direct electron uptake as the main MIC mechanism for SRB D. bastinii, which has not been studied before, and D. vulgaris, an organotrophic SRB. Both are common bacteria existing in crude oil pipelines. Studies including electrochemical measurements, immersion corrosion testing, metal surface monitoring via scanning electron microscope revealed direct-electron uptake capabilities for both strains. SRB strains were tested under 18 different environmental conditions. Extremely high cathodic current densities were observed in SRB cultures confirming electron transfer from the iron surface to bacteria cells. Finally, based on the large experimental dataset provided in this work, an artificial neural network model was developed to predict MIC. This model demonstrates high correlation coefficients comparable or higher than existing models for general corrosion prediction in the literature. Revealing the predominant mechanisms of MIC along with modeling capabilities enables us to design appropriate measures to eradicate pipe failure due to MIC. Additionally the investigated direct electron uptake ability of the specific SRB strains studied can be used in microbial fuel cells for enhancing the efficiency of biocathodes.Item Nonlinear Sampling Theory and Efficient Signal Recovery(2020) Lin, Kung-Ching; Benedetto, John; Mathematics; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Sampling theory investigates signal recovery from its partial information, and one of the simplest and most well-known sampling schemes is uniform linear sampling, characterized by the celebrated classical sampling theorem. However, the requirements of uniform linear sampling may not always be satisfied, sparking the need for more general sampling theories. In the thesis, we discuss the following three sampling scenarios: signal quantization, compressive sensing, and deep neural networks. In signal quantization theory, the inability of digital devices to perfectly store analog samples leads to distortion when reconstructing the signal from its samples. Different quantization schemes are proposed so as to minimize such distortion. We adapt a quantization scheme used in analog-to-digital conversion called signal decimation to finite dimensional signals. In doing so, we are able to achieve theoretically optimal reconstruction error decay rate. Compressive sensing investigates the possibility to recover high-dimensional signals from incomplete samples. It has been proven feasible as long as the signal is sufficiently sparse. To this point, all of the most successful examples follow from random constructions rather than deterministic ones. Whereas the sparsity of the signal can be almost as large as the ambient dimension for random constructions, current deterministic constructions require the sparsity to be at most the square-root of the ambient dimension. This apparent barrier is the well-known square-root bottleneck. In this thesis, we propose a new explicit sampling scheme as a possible candidate for deterministic compressive sensing. We present a partial result, while the full generality is still work in progress. For deep neural networks, one approximates signals with neural networks. To do so, many samples need to be drawn in order to find an optimal approximating neural network. A common approach is to employ stochastic gradient descent, but it is unclear if the resulting neural network is indeed optimal due to the non-convexity of the optimization scheme. We follow an alternative approach, utilizing the derivatives of the signal for stable reconstruction. In this thesis, we focus on non-smooth signals, and using weak differentiation, it is easy to obtain stable reconstruction for one-layer neural networks. We are currently working on the two-layer case, and our approach is outlined in this thesis.Item Adversarial Robustness and Robust Meta-Learning for Neural Networks(2020) Goldblum, Micah; Czaja, Wojciech; Mathematics; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Despite the overwhelming success of neural networks for pattern recognition, these models behave categorically different from humans. Adversarial examples, small perturbations which are often undetectable to the human eye, easily fool neural networks, demonstrating that neural networks lack the robustness of human classifiers. This thesis comprises a sequence of three parts. First, we motivate the study of defense against adversarial examples with a case study on algorithmic trading in which robustness may be critical for security reasons. Second, we develop methods for hardening neural networks against an adversary, especially in the low-data regime, where meta-learning methods achieve state-of-the-art results. Finally, we discuss several properties of the neural network models we use. These properties are of interest beyond robustness to adversarial examples, and they extend to the broad setting of deep learning.Item Temporal Context Modeling for Text Streams(2018) Rao, Jinfeng; Lin, Jimmy; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)There is increasing recognition that time plays an essential role in many information seeking tasks. This dissertation explores temporal models on evolving streams of text and the role that such models play in improving information access. I consider two cases: a stream of social media posts by many users for tweet search and a stream of queries by an individual user for voice search. My work explores the relationship between temporal models and context models: for tweet search, the evolution of an event serves as the context of clustering relevant tweets; for voice search, the user's history of queries provides the context for helping understand her true information need. First, I tackle the tweet search problem by modeling the temporal contexts of the underlying collection. The intuition is that an information need in Twitter usually correlates with a breaking news event, thus tweets posted during that event are more likely to be relevant. I explore techniques to model two different types of temporal signals: pseudo trend and query trend. The pseudo trend is estimated through the distribution of timestamps from an initial list of retrieved documents given a query, which I model through continuous hidden Markov approach as well as neural network-based methods for relevance ranking and sequence modeling. As an alternative, the query trend, is directly estimated from the temporal statistics of query terms, obviating the need for an initial retrieval. I propose two different approaches to exploit query trends: a linear feature-based ranking model and a regression-based model that recover the distribution of relevant documents directly from query trends. Extensive experiments on standard Twitter collections demonstrate the superior effectivenesses of my proposed techniques. Second, I introduce the novel problem of voice search on an entertainment platform, where users interact with a voice-enabled remote controller through voice requests to search for TV programs. Such queries range from specific program navigation (i.e., watch a movie) to requests with vague intents and even queries that have nothing to do with watching TV. I present successively richer neural network architectures to tackle this challenge based on two key insights: The first is that session context can be exploited to disambiguate queries and recover from ASR errors, which I operationalize with hierarchical recurrent neural networks. The second insight is that query understanding requires evidence integration across multiple related tasks, which I identify as program prediction, intent classification, and query tagging. I present a novel multi-task neural architecture that jointly learns to accomplish all three tasks. The first model, already deployed in production, serves millions of queries daily with an improved customer experience. The multi-task learning model is evaluated on carefully-controlled laboratory experiments, which demonstrates further gains in effectiveness and increased system capabilities. This work now serves as the core technology in Comcast Xfinity X1 entertainment platform, which won an Emmy award in 2017 for the technical contribution in advancing television technologies. This dissertation presents families of techniques for modeling temporal information as contexts to assist applications with streaming inputs, such as tweet search and voice search. My models not only establish the state-of-the-art effectivenesses on many related tasks, but also reveal insights of how various temporal patterns could impact real information-seeking processes.Item The Influence of Collective Working Memory Strategies on Agent Teams(2007-08-03) Winder, Ransom Kershaw; Reggia, James A.; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Past self-organizing models of collectively moving "particles" (simulated bird flocks, fish schools, etc.) typically have been based on purely reflexive agents that have no significant memory of past movements or environmental obstacles. These agent collectives usually operate in abstract environments, but as these domains take on a greater realism, the collective requires behaviors use not only presently observed stimuli but also remembered information. It is hypothesized that the addition of a limited working memory of the environment, distributed among the collective's individuals can improve efficiency in performing tasks. This is first approached in a more traditional particle system in an abstract environment. Then it is explored for a single agent, and finally a team of agents, operating in a simulated 3-dimensional environment of greater realism. In the abstract environment, a limited distributed working memory produced a significant improvement in travel between locations, in some cases improving performance over time, while in others surprisingly achieving an immediate benefit from the influence of memory. When strategies for accumulating and manipulating memory were subsequently explored for a more realistic single agent in the 3-dimensional environment, if the agent kept a local or a cumulative working memory, its performance improved on different tasks, both when navigating nearby obstacles and, in the case of cumulative memory, when covering previously traversed terrain. When investigating a team of these agents engaged in a pursuit scenario, it was determined that a communicating and coordinating team still benefited from a working memory of the environment distributed among the agents, even with limited memory capacity. This demonstrates that a limited distributed working memory in a multi-agent system improves performance on tasks in domains of increasing complexity. This is true even though individual agents know only a fraction of the collective's entire memory, using this partial memory and interactions with others in the team to perform tasks. These results may prove useful in improving existing methodologies for control of collective movements for robotic teams, computer graphics, particle swarm optimization, and computer games, and in interpreting future experimental research on group movements in biological populations.Item Neural Network Generation of Temporal Sequences from Single Static Vector Inputs using Varying Length Distal Target Sequences(2007-04-10) Gittens, Shaun; Reggia, James; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Training an agent to operate in an environment whose mappings are largely unknown is generally recognized to be exceptionally difficult. Further, granting such a learning agent the ability to produce an appropriate sequence of actions entirely from a single input stimulus remains a key problem. Various reinforcement learning techniques have been utilized to handle such learning tasks, but convergence to optimal policies is not guaranteed for many of these methods. Traditional supervised learning methods hold more assurances of convergence, but these methods are not well suited for tasks where desired actions in the output space of the learner, termed proximal actions, are not available for training. Rather, target outputs from the environment are distal from where the learning takes place. For example, a child acquiring language skill who makes speech errors must learn to correct them based on heard information that reaches his/her auditory cortex, which is distant from the motor cortical regions that control speech output. While distal supervised learning techniques for neural networks have been devised, it remains to be established how they can be trained to produce sequences of proximal actions from only a single static input. The architecture demonstrated here incorporates recurrent multi-layered neural networks, each maintaining some manner of memory in the form of a context vector, into the distal supervised learning framework. This enables it to train learners capable of generating correct proximal sequences from single static input stimuli. This is in contrast to existing distal learning methods designed for non-recurrent neural network learners that utilize no concept of memory of their prior behavior. Also, a technique known as teacher forcing was adapted for use in distal sequential learning settings which is shown to result in more efficient usage of the recurrent neural network's context layer. The effectiveness of this approach is demonstrated by applying it in training recurrent learners to acquire phoneme sequence generating behavior using only previously heard and stored auditory phoneme sequences. The results indicate that recurrent networks can be integrated with distal learning methods to create effective sequence generators even when constantly updating current state information is unavailable.Item Robust Optimization Model for Bus Priority under Arterial Progression(2005-10-05) Vasudevan, Meenakshy; Chang, Gang-Len; Civil Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)The purpose of this study is to design a real-time robust arterial signal control system that gives priority to buses while simultaneously maximizing progression bandwidths and optimizing signal timing plans at each intersection along the arterial. The system architecture is divided into three levels. At the progression control level bandwidths are maximized. Existing progression strategies do not use real-time traffic data or use simple mathematical models to estimate traffic evolution. The proposed model eliminates this drawback by using real-time data to develop a neural network model for predicting traffic flows. Rather than using pre-specified values, queue clearance and minimum green times are computed as functions of the predicted queues. To eliminate uncertainty in the prediction due to the long time horizon, robust discrete optimization technique is used to determine the progression bands. At the intersection control level, signal timing plans are optimized subject to bandwidth constraints to allow for uninterrupted arterial flow, and minimum green constraints for driver safety and to discharge average waiting queues. At the bus priority control level, whenever a bus is detected and is a candidate for priority it is granted priority based on a performance index that is a function of bus schedule delay, automobile and bus passenger delays, and vehicle delays, subject to bandwidth and minimum green constraints. Minimum green constraints ensure that other traffic users are not unduly penalized. Bandwidth constraints allow for uninterrupted arterial flow despite a preferential treatment of buses. The performance of the proposed system is evaluated through a case study conducted in a laboratory environment using CORSIM. Results show that the models developed at the three levels are superior to the signal control implemented in the field, and the alternatives that use the off-line MULTIBAND model for progression for all traffic scenarios. Robust optimization was highly effective in reducing control delays, stop times, queues, and bus delays, and increasing throughput and speeds, when traffic volumes were high. The model that integrated bus priority with robust arterial signal control produced the most reductions in bus delays while not causing significant delays to automobiles.