In an unsupervised learning problem, one is given an unlabelled dataset and hopes to find some hidden structure; the prototypical example is clustering similar data. Such problems often arise in machine learning and statistics, but also in signal processing, theoretical computer science, and any number of quantitative scientific fields. The distinguishing feature of unsupervised learning is that there are no privileged variables or labels which are particularly informative, and thus the greatest challenge is often to differentiate between what is relevant or irrelevant in any particular dataset or problem. In the course of this thesis, we study a number of problems which span the breadth of unsupervised learning. We make progress in Gaussian mixtures, independent component analysis (where we solve the open problem of underdetermined ICA), and we formulate and solve a feature selection/dimension reduction model. Throughout, our goal is to give finite sample complexity bounds for our algorithms -- these are essentially the strongest type of quantitative bound that one can prove for such algorithms. Some of our algorithmic techniques turn out to be very efficient in practice as well. Our major technical tool is tensor spectral decomposition: tensors are generalisations of matrices, and often allow access to the "fine structure" of data. Thus, they are often the right tools for unravelling the hidden structure in an unsupervised learning setting. However, naive generalisations of matrix algorithms to tensors run into NP-hardness results almost immediately, and thus to solve our problems, we are obliged to develop two new tensor decompositions (with robust analyses) from scratch. Both of these decompositions are polynomial time, and can be viewed as efficient generalisations of PCA extended to tensors.

In radiation therapy, a treatment plan is designed to make the delivery of radiation to a target more accurate, effective, and less damaging to surrounding healthy tissues. In lung sites, the tumor is affected by the patient’s respiratory motion. Despite tumor motion, current practice still uses a static delivery plan. Unexpected changes due to coughs and sneezes are not taken into account and as a result, the tumor is not treated accurately and healthy tissues are damaged. In this thesis we detail a framework of using an accelerometer device to detect and forecast coughs. The accelerometer measurements are modeled as a ARMA process to make forecasts. We draw from studies in cough physiology and use amplitudes and durations of the forecasted breathing cycles as features to estimate parameters of Gaussian Mixture Models for cough and normal breathing classes. The system was tested on 10 volunteers, where each data set consisted of one 3-5 minute accelerometer measurements to train the system, and two 1-3 minute accelerometer measurements for testing.

In radiation therapy, a treatment plan is designed to make the delivery of radiation to a target more accurate, effective, and less damaging to surrounding healthy tissues. In lung sites, the tumor is affected by the patient’s respiratory motion. Despite tumor motion, current practice still uses a static delivery plan. Unexpected changes due to coughs and sneezes are not taken into account and as a result, the tumor is not treated accurately and healthy tissues are damaged. In this thesis we detail a framework of using an accelerometer device to detect and forecast coughs. The accelerometer measurements are modeled as a ARMA process to make forecasts. We draw from studies in cough physiology and use amplitudes and durations of the forecasted breathing cycles as features to estimate parameters of Gaussian Mixture Models for cough and normal breathing classes. The system was tested on 10 volunteers, where each data set consisted of one 3-5 minute accelerometer measurements to train the system, and two 1-3 minute accelerometer measurements for testing.

No description available.

Texture analysis is an important problem in machine vision, with applications in many fields including medical imaging, remote sensing (SAR), automated flaw detection in various products, and document analysis to name but a few. Over the last four decades many techniques for the analysis of textured images have been proposed in the literature for the purposes of classification, segmentation, synthesis and compression. Such approaches include analysis the properties of individual texture elements, using statistical features obtained from the grey-level values of the image itself, random field models, and multichannel filtering. The wavelet transform, a unified framework for the multiresolution decomposition of signals, falls into this final category, and allows a texture to be examined in a number of resolutions whilst maintaining spatial resolution. This thesis explores the use of the wavelet transform to the specific task of texture classification and proposes a number of improvements to existing techniques, both in the area of feature extraction and classifier design. By applying a nonlinear transform to the wavelet coefficients, a better characterisation can be obtained for many natural textures, leading to increased classification performance when using first and second order statistics of these coefficients as features. In the area of classifier design, a combination of an optimal discriminate function and a non-parametric Gaussian mixture model classifier is shown to experimentally outperform other classifier configurations. By modelling the relationships between neighbouring bands of the wavelet trans- form, more information regarding a texture can be obtained. Using such a representation, an efficient algorithm for the searching and retrieval of textured images from a database is proposed, as well as a novel set of features for texture classification. These features are experimentally shown to outperform features proposed in the literature, as well as provide increased robustness to small changes in scale. Determining the script and language of a printed document is an important task in the field of document processing. In the final part of this thesis, the use of texture analysis techniques to accomplish these tasks is investigated. Using maximum a posterior (MAP) adaptation, prior information regarding the nature of script images can be used to increase the accuracy of these methods. Novel techniques for estimating the skew of such documents, normalising text block prior to extraction of texture features and accurately classifying multiple fonts are also presented.

abstract: This thesis examines the application of statistical signal processing approaches to data arising from surveys intended to measure psychological and sociological phenomena underpinning human social dynamics. The use of signal processing methods for analysis of signals arising from measurement of social, biological, and other non-traditional phenomena has been an important and growing area of signal processing research over the past decade. Here, we explore the application of statistical modeling and signal processing concepts to data obtained from the Global Group Relations Project, specifically to understand and quantify the effects and interactions of social psychological factors related to intergroup conflicts. We use Bayesian networks to specify prospective models of conditional dependence. Bayesian networks are determined between social psychological factors and conflict variables, and modeled by directed acyclic graphs, while the significant interactions are modeled as conditional probabilities. Since the data are sparse and multi-dimensional, we regress Gaussian mixture models (GMMs) against the data to estimate the conditional probabilities of interest. The parameters of GMMs are estimated using the expectation-maximization (EM) algorithm. However, the EM algorithm may suffer from over-fitting problem due to the high dimensionality and limited observations entailed in this data set. Therefore, the Akaike information criterion (AIC) and the Bayesian information criterion (BIC) are used for GMM order estimation. To assist intuitive understanding of the interactions of social variables and the intergroup conflicts, we introduce a color-based visualization scheme. In this scheme, the intensities of colors are proportional to the conditional probabilities observed. / Dissertation/Thesis / M.S. Electrical Engineering 2012

A contribução original desta tese é um novo algoritmo que integra um aproximador de funções com alta eficiência amostral com aprendizagem por reforço em espaços de estados contínuos. A pesquisa completa inclui o desenvolvimento de um algoritmo online e incremental capaz de aprender por meio de uma única passada sobre os dados. Este algoritmo, chamado de Fast Incremental Gaussian Mixture Network (FIGMN) foi empregado como um aproximador de funções eficiente para o espaço de estados de tarefas contínuas de aprendizagem por reforço, que, combinado com Q-learning linear, resulta em performance competitiva. Então, este mesmo aproximador de funções foi empregado para modelar o espaço conjunto de estados e valores Q, todos em uma única FIGMN, resultando em um algoritmo conciso e com alta eficiência amostral, i.e., um algoritmo de aprendizagem por reforço capaz de aprender por meio de pouquíssimas interações com o ambiente. Um único episódio é suficiente para aprender as tarefas investigadas na maioria dos experimentos. Os resultados são analisados a fim de explicar as propriedades do algoritmo obtido, e é observado que o uso da FIGMN como aproximador de funções oferece algumas importantes vantagens para aprendizagem por reforço em relação a redes neurais convencionais. / This thesis’ original contribution is a novel algorithm which integrates a data-efficient function approximator with reinforcement learning in continuous state spaces. The complete research includes the development of a scalable online and incremental algorithm capable of learning from a single pass through data. This algorithm, called Fast Incremental Gaussian Mixture Network (FIGMN), was employed as a sample-efficient function approximator for the state space of continuous reinforcement learning tasks, which, combined with linear Q-learning, results in competitive performance. Then, this same function approximator was employed to model the joint state and Q-values space, all in a single FIGMN, resulting in a concise and data-efficient algorithm, i.e., a reinforcement learning algorithm that learns from very few interactions with the environment. A single episode is enough to learn the investigated tasks in most trials. Results are analysed in order to explain the properties of the obtained algorithm, and it is observed that the use of the FIGMN function approximator brings some important advantages to reinforcement learning in relation to conventional neural networks.

Classificação de fluxos contínuos de dados possui muitos desafios para a comunidade de mineração de dados quando o ambiente não é estacionário. Um dos maiores desafios para a aprendizagem em fluxos contínuos de dados está relacionado com a adaptação às mudanças de conceito, as quais ocorrem como resultado da evolução dos dados ao longo do tempo. Duas formas principais de desenvolver abordagens adaptativas são os métodos baseados em conjunto de classificadores e os algoritmos incrementais. Métodos baseados em conjunto de classificadores desempenham um papel importante devido à sua modularidade, o que proporciona uma maneira natural de se adaptar a mudanças de conceito. Os algoritmos incrementais são mais rápidos e possuem uma melhor capacidade anti-ruído do que os conjuntos de classificadores, mas têm mais restrições sobre os fluxos de dados. Assim, é um desafio combinar a flexibilidade e a adaptação de um conjunto de classificadores na presença de mudança de conceito, com a simplicidade de uso encontrada em um único classificador com aprendizado incremental. Com essa motivação, nesta dissertação, propomos um algoritmo incremental, online e probabilístico para a classificação em problemas que envolvem mudança de conceito. O algoritmo é chamado IGMN-NSE e é uma adaptação do algoritmo IGMN. As duas principais contribuições da IGMN-NSE em relação à IGMN são: melhoria de poder preditivo para tarefas de classificação e a adaptação para alcançar um bom desempenho em cenários não estacionários. Estudos extensivos em bases de dados sintéticas e do mundo real demonstram que o algoritmo proposto pode rastrear os ambientes em mudança de forma muito próxima, independentemente do tipo de mudança de conceito. / Data stream classification poses many challenges for the data mining community when the environment is non-stationary. The greatest challenge in learning classifiers from data stream relates to adaptation to the concept drifts, which occur as a result of changes in the underlying concepts. Two main ways to develop adaptive approaches are ensemble methods and incremental algorithms. Ensemble method plays an important role due to its modularity, which provides a natural way of adapting to change. Incremental algorithms are faster and have better anti-noise capacity than ensemble algorithms, but have more restrictions on concept drifting data streams. Thus, it is a challenge to combine the flexibility and adaptation of an ensemble classifier in the presence of concept drift, with the simplicity of use found in a single classifier with incremental learning. With this motivation, in this dissertation we propose an incremental, online and probabilistic algorithm for classification as an effort of tackling concept drifting. The algorithm is called IGMN-NSE and is an adaptation of the IGMN algorithm. The two main contributions of IGMN-NSE in relation to the IGMN are: predictive power improvement for classification tasks and adaptation to achieve a good performance in non-stationary environments. Extensive studies on both synthetic and real-world data demonstrate that the proposed algorithm can track the changing environments very closely, regardless of the type of concept drift.

Solute clustering is known to play a significant role in the embrittlement of reactor pressure vessel (RPV) steels. When precipitates form they impede the movement of dislocations, causing an increase in hardness and a shift in the ductile-brittle transition temperature. Over time this can cause the steel to become brittle and more susceptible to fracture. Thus, understanding precipitate formation is of great importance to the nuclear industry. The first part of this thesis aims to isolate and better understand the thermal aging component of embrittlement in low copper, model RPV steels. Currently, relatively little is known about the effects of Ni and Mn in a low copper environment. Therefore, it is of interest to determine if Ni and Mn form precipitates under these conditions. To this end, hardness measurements and atom probe tomography were utilized to link the mechanical properties to the microstructure. After 11,690 hours of thermal aging a statistically significant decrease in hardening was observed. Consistent with hardness measurements, no precipitates were present within the matrix of the thermally aged RPV steels. The local chemistry method was then applied to investigate the very early stages of solute clustering. Association was found to be statistically significant in both the thermally aged and as-received model RPV steels. Therefore, no apparent trends regarding the changes in solute association between the as-received and thermally aged RPV steels were identified. Small, non-random clusters were observed at heterogeneous nucleation sites, such as carbide/matrix interfaces and grain boundaries, within the thermally aged material. The clusters found at the carbide/matrix interfaces were all rich in Mn and approximately 90-150 atoms in size. The clusters located along the observed low-angle grain boundary, however, were significantly larger (on the order of hundreds of atoms) and rich in Ni. Lastly, copper-rich precipitates (CRPs) and Mn- and Ni-rich precipitates (MNPs) were observed within the cementite phase of a high copper and low copper RPV steel, respectively, following long term thermal aging. APT was used to characterize these precipitates and obtain more detailed chemical information. The presence of such precipitates indicates that a range of precipitation can take place within the cementite phase of thermally aged RPV steels. The second part of this thesis aims to investigate the effects of ion irradiation on the microstructure of low copper RPV steels via APT. These steels were ion irradiated with 6.4 MeV Fe³⁺ ions with a dose rate of 1.5 x 10^-4 dpa/s at 290°C. MNPs were observed in all five of the RPV steels analyzed. These precipitates were found to have nucleated within the matrix as well as at dislocations and grain boundaries. Using the maximum separation method these MNPs were extracted and characterized. Precipitate composition, size, volume fraction, and number density were determined for each of the five samples. Lastly, several grain boundaries were characterized. Several emerging trends were observed within the samples: Ni content within the precipitates did not vary significantly once a threshold between 30-50% was reached; bulk Mn content appeared to dictate Si and Mn content within the precipitates; and samples low in bulk Ni content were characterized by a higher number density of smaller precipitates. Additionally, by regressing precipitate volume fraction against the interaction of Ni and Mn, a linear relationship was found to be statistically significant.

abstract: Parkinson's disease is a neurodegenerative condition diagnosed on patients with clinical history and motor signs of tremor, rigidity and bradykinesia, and the estimated number of patients living with Parkinson's disease around the world is seven to ten million. Deep brain stimulation (DBS) provides substantial relief of the motor signs of Parkinson's disease patients. It is an advanced surgical technique that is used when drug therapy is no longer sufficient for Parkinson's disease patients. DBS alleviates the motor symptoms of Parkinson's disease by targeting the subthalamic nucleus using high-frequency electrical stimulation. This work proposes a behavior recognition model for patients with Parkinson's disease. In particular, an adaptive learning method is proposed to classify behavioral tasks of Parkinson's disease patients using local field potential and electrocorticography signals that are collected during DBS implantation surgeries. Unique patterns exhibited between these signals in a matched feature space would lead to distinction between motor and language behavioral tasks. Unique features are first extracted from deep brain signals in the time-frequency space using the matching pursuit decomposition algorithm. The Dirichlet process Gaussian mixture model uses the extracted features to cluster the different behavioral signal patterns, without training or any prior information. The performance of the method is then compared with other machine learning methods and the advantages of each method is discussed under different conditions. / Dissertation/Thesis / Masters Thesis Electrical Engineering 2015