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
2 results
Search Results
Item Scalable learning for geostatistics and speaker recognition(2011) Srinivasan, Balaji Vasan; Duraiswami, Ramani; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)With improved data acquisition methods, the amount of data that is being collected has increased severalfold. One of the objectives in data collection is to learn useful underlying patterns. In order to work with data at this scale, the methods not only need to be effective with the underlying data, but also have to be scalable to handle larger data collections. This thesis focuses on developing scalable and effective methods targeted towards different domains, geostatistics and speaker recognition in particular. Initially we focus on kernel based learning methods and develop a GPU based parallel framework for this class of problems. An improved numerical algorithm that utilizes the GPU parallelization to further enhance the computational performance of kernel regression is proposed. These methods are then demonstrated on problems arising in geostatistics and speaker recognition. In geostatistics, data is often collected at scattered locations and factors like instrument malfunctioning lead to missing observations. Applications often require the ability interpolate this scattered spatiotemporal data on to a regular grid continuously over time. This problem can be formulated as a regression problem, and one of the most popular geostatistical interpolation techniques, kriging is analogous to a standard kernel method: Gaussian process regression. Kriging is computationally expensive and needs major modifications and accelerations in order to be used practically. The GPU framework developed for kernel methods is extended to kriging and further the GPU's texture memory is better utilized for enhanced computational performance. Speaker recognition deals with the task of verifying a person's identity based on samples of his/her speech - "utterances". This thesis focuses on text-independent framework and three new recognition frameworks were developed for this problem. We proposed a kernelized Renyi distance based similarity scoring for speaker recognition. While its performance is promising, it does not generalize well for limited training data and therefore does not compare well to state-of-the-art recognition systems. These systems compensate for the variability in the speech data due to the message, channel variability, noise and reverberation. State-of-the-art systems model each speaker as a mixture of Gaussians (GMM) and compensate for the variability (termed "nuisance"). We propose a novel discriminative framework using a latent variable technique, partial least squares (PLS), for improved recognition. The kernelized version of this algorithm is used to achieve a state of the art speaker ID system, that shows results competitive with the best systems reported on in NIST's 2010 Speaker Recognition Evaluation.Item Gaussian Process Regression for Model Estimation(2008) Srinivasan, Balaji Vasan; Duraiswami, Ramani; Electrical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)State estimation techniques using Kalman filter and Particle filters are used in a number of applications like tracking, econometrics, weather data assimilation, etc. These techniques aim at estimating the state of the system using the system characteristics. System characteristics include the definition of system's dynamical model and the observation model. While the Kalman filter uses these models explicitly, particle filter based estimation techniques use these models as part of sampling and assigning weights to the particles. If the state transition and observation models are not available, an approximate model is used based on the knowledge of the system. However, if the system is a total black box, it is possible that the approximate models are not the correct representation of the system and hence will lead to poor estimation. This thesis proposes a method to deal with such situations by estimating the models and the states simultaneously. The thesis concentrates on estimating the system's dynamical model and the states, given the observation model and the noisy observations. A Gaussian process regression based method is developed for estimating the model. The regression method is sped up from O(N2) to O(N) using an data-dependent online approach for fast Gaussian summations. A relevance vector machine based data selection scheme is used to propagate the model over iterations. The proposed method is tested on a Local Ensemble Kalman Filter based estimation for the highly non-linear Lorenz-96 model.