Made available in DSpace on 2016-08-29T15:33:12Z (GMT). No. of bitstreams: 1 tese_3521_.pdf: 2782962 bytes, checksum: d4b2294e5ee9ab86b7a35aec083af692 (MD5) Previous issue date: 2010-02-12 / Os Modelos de Predição implementados pelos algoritmos de Aprendizagem de Máquina advindos como linha de pesquisa da Inteligência Computacional são resultantes de pesquisas e investigações empíricas em dados do mundo real. Neste contexto; estes modelos são extraídos para comparação de duas grandes técnicas de aprendizagem de máquina Redes Neuro-Fuzzy e de Regressão aplicadas no intuito de estimar um parâmetro de qualidade do produto em um ambiente industrial sob processo contínuo. Heuristicamente; esses Modelos de Predição são aplicados e comparados em um mesmo ambiente de simulação com intuito de mensurar os níveis de adequação dos mesmos, o poder de desempenho e generalização dos dados empíricos que compõem este cenário (ambiente industrial de mineração).

abstract: Gas turbine engine for aircraft propulsion represents one of the most physics-complex and safety-critical systems in the world. Its failure diagnostic is challenging due to the complexity of the model system, difficulty involved in practical testing and the infeasibility of creating homogeneous diagnostic performance evaluation criteria for the diverse engine makes. NASA has designed and publicized a standard benchmark problem for propulsion engine gas path diagnostic that enables comparisons among different engine diagnostic approaches. Some traditional model-based approaches and novel purely data-driven approaches such as machine learning, have been applied to this problem. This study focuses on a different machine learning approach to the diagnostic problem. Some most common machine learning techniques, such as support vector machine, multi-layer perceptron, and self-organizing map are used to help gain insight into the different engine failure modes from the perspective of big data. They are organically integrated to achieve good performance based on a good understanding of the complex dataset. The study presents a new hierarchical machine learning structure to enhance classification accuracy in NASA's engine diagnostic benchmark problem. The designed hierarchical structure produces an average diagnostic accuracy of 73.6%, which outperforms comparable studies that were most recently published. / Dissertation/Thesis / Masters Thesis Electrical Engineering 2015

Ingeniero Civil Matemático / El suicidio en Chile se ha convertido en uno de los problemas más necesarios de hacer frente en salud pública, más aún, si consideramos que la enorme mayoría de las personas que mueren por suicidio presentan algún diagnóstico psiquiátrico y han consultado a un especialista los meses antes de cometer suicidio. Esto, motiva la creación de indicadores y alertas para detectar de forma eficaz y oportuna cuando una persona ingresa a una zona de riesgo suicida. En el presente trabajo se aborda este problema, definiendo una zona o espectro de riesgo suicida, y generando modelos matemáticos y estadísticos para la detección de pacientes en esta zona de riesgo. Para esto, se utiliza una base de datos de 707 pacientes, consultantes de salud mental, de tres centros de salud distintos de la región metropolitana. La base de datos a su vez contempla 343 variables, incluyendo tanto información sociodemográfica de cada paciente, como también sus respuestas en siete instrumentos clínicos utilizados habitualmente en salud mental (DEQ, STAXI, OQ, RFL, APGAR, PBI Madre y PBI Padre). Inicialmente la base de datos es depurada eliminando aquellos campos y/o registros con gran porcentaje de valores nulos, mientras que la imputación de valores perdidos se realiza mediante técnicas tradicionales y en algunos casos según el criterio experto, donde se utiliza un método de imputación según valor de subescala para los distintos instrumentos clínicos. Posteriormente, se realiza una reducción de atributos mediante el uso de herramientas estadísticas y provenientes del machine learning. Con esta información, se generan cinco modelos utilizando distintas técnicas y herramientas del ámbito de la minería de datos y machine learning mediante aprendizaje supervisado. Los modelos son generados y calibrados usando el lenguaje estadístico R, y se comparan sus resultados mediante cuatro métricas distintas: precisión (o accuracy), sensibilidad, especificidad, y mediante su representación en el espacio ROC. El modelo o clasificador finalmente propuesto corresponde a un modelo de support vector machine, que permite discriminar cuando un paciente se encuentra en una zona de riesgo suicida. El modelo fue entrenado utilizando un kernel de tipo RBF, y utiliza tan sólo 22 variables predictoras, entregando una precisión aproximada del $78%, calculada mediante k-validación cruzada de n-folds con k=100 y n=10.

We present a novel method for learning Support Vector Machines (SVMs) in the online setting. Our method is generally applicable in that it handles the online learning of the binary, multiclass, and structural SVMs in a unified view. The SVM learning problem consists of optimizing a convex objective function that is composed of two parts: the hinge loss and quadratic regularization. To date, the predominant family of approaches for online SVM learning has been gradient-based methods, such as Stochastic Gradient Descent (SGD). Unfortunately, we note that there are two drawbacks in such approaches: first, gradient-based methods are based on a local linear approximation to the function being optimized, but since the hinge loss is piecewise-linear and nonsmooth, this approximation can be ill-behaved. Second, existing online SVM learning approaches share the same problem formulation with batch SVM learning methods, and they all need to tune a fixed global regularization parameter by cross validation. On the one hand, global regularization is ineffective in handling local irregularities encountered in the online setting; on the other hand, even though the learning problem for a particular global regularization parameter value may be efficiently solved, repeatedly solving for a wide range of values can be costly. We intend to tackle these two problems with our approach. To address the first problem, we propose to perform implicit online update steps to optimize the hinge loss, as opposed to explicit (or gradient-based) updates that utilize subgradients to perform local linearization. Regarding the second problem, we propose to enforce local regularization that is applied to individual classifier update steps, rather than having a fixed global regularization term. Our theoretical analysis suggests that our classifier update steps progressively optimize the structured hinge loss, with the rate controlled by a sequence of regularization parameters; setting these parameters is analogous to setting the stepsizes in gradient-based methods. In addition, we give sufficient conditions for the algorithm's convergence. Experimentally, our online algorithm can match optimal classification performances given by other state-of-the-art online SVM learning methods, as well as batch learning methods, after only one or two passes over the training data. More importantly, our algorithm can attain these results without doing cross validation, while all other methods must perform time-consuming cross validation to determine the optimal choice of the global regularization parameter.

No description available.

No description available.

Yes / Burns are one of the obnoxious injuries subjecting thousands to loss of life and physical defacement each year. Both high income and Third World countries face major evaluation challenges including but not limited to inadequate workforce, poor diagnostic facilities, inefficient diagnosis and high operational cost. As such, there is need to develop an automatic machine learning algorithm to noninvasively identify skin burns. This will operate with little or no human intervention, thereby acting as an affordable substitute to human expertise. We leverage the weights of pretrained deep neural networks for image description and, subsequently, the extracted image features are fed into the support vector machine for classification. To the best of our knowledge, this is the first study that investigates black African skins. Interestingly, the proposed algorithm achieves state-of-the-art classification accuracy on both Caucasian and African datasets.

Emotion dysregulation plays a central role in psychopathology (B. Bradley et al., 2011) and has been linked to aberrant activation of neural circuitry involved in emotion regulation (Beauregard, Paquette, & Lévesque, 2006; Etkin & Schatzberg, 2011). In recent years, technological advances in neuroimaging methods coupled with developments in machine learning have allowed for the non-invasive measurement and prediction of brain states in real-time, which can be used to provide feedback to facilitate regulation of brain states (LaConte, 2011). Real-time functional magnetic resonance imaging (rt-fMRI)-guided neurofeedback, has promise as a novel therapeutic method in which individuals are provided with tailored feedback to improve regulation of emotional responses (Stoeckel et al., 2014). However, effective use of this technology for such purposes likely entails the development of (a) a normative model of emotion processing to provide feedback for individuals with emotion processing difficulties; and (b) best practices concerning how these types of group models are designed and translated for use in a rt-fMRI environment (Ruiz, Buyukturkoglu, Rana, Birbaumer, & Sitaram, 2014). To this end, the present study utilized fMRI data from a standard emotion elicitation paradigm to examine the impact of several design decisions made during the development of a whole-brain model of affective processing. Using support vector machine (SVM) learning, we developed a group model that reliably classified brain states associated with passive viewing of positive, negative, and neutral images. After validating the group whole-brain model, we adapted this model for use in an rt-fMRI experiment, and using a second imaging dataset along with our group model, we simulated rt-fMRI predictions and tested options for providing feedback. / Master of Science

Land cover classification using Synthetic Aperture Radar (SAR) data has been a topic of great interest in recent literature. Food commodities output prediction through crop identification, environmental monitoring, and forest regrowth tracking are some of the many problems that can be aided by land cover classification methods. The need for fast and automated classification methods is apparent in a variety of applications involving vast amounts of SAR data. One fundamental step in any supervised learning classification algorithm is the selection and/or extraction of features present in the dataset to be used for class discrimination. A popular method that has been proposed for feature extraction from polarimetric data is to decompose the data into the underlying scattering mechanisms. In this research, the Freeman and Durden scattering model is applied to ALOS PALSAR fully polarimetric data for feature extraction. Efficient methods for solving the complex system of equations present in the scattering model are developed and compared. Using the features from the Freeman and Durden work, the classification capability of the model is assessed on amazon rainforest land cover types using a supervised Support Vector Machine (SVM) classification algorithm. The quantity of land cover types that can be discriminated using the model is also determined. Additionally, the performance of the median as a robust estimator in noisy environments for multi-pixel windowing is also characterized. / Master of Science / Land type classification using Radar data has been a topic of great interest in recent literature. Food commodities output prediction through crop identification, environmental monitoring, and forest regrowth tracking are some of the many problems that can be aided by land cover classification methods. The need for fast and automated classification methods is apparent in a variety of applications involving vast amounts of Radar data. One fundamental step in any classification algorithm is the selection and/or extraction of discriminating features present in the dataset to be used for class discrimination. A popular method that has been proposed for feature extraction from polarized Radar data is to decompose the data into the underlying scatter components. In this research, a scattering model is applied to real world data for feature extraction. Efficient methods for solving the complex system of equations present in the scattering model are developed and compared. Using the features from the scattering model, the classification capability of the model is assessed on amazon rainforest land types using a Support Vector Machine (SVM) classification algorithm. The quantity of land cover types that can be discriminated using the model is also determined and compared using different estimators.

No / Background: Slowness of movement, known as bradykinesia, is the core clinical sign of Parkinson's and fundamental to its diagnosis. Clinicians commonly assess bradykinesia by making a visual judgement of the patient tapping finger and thumb together repetitively. However, inter-rater agreement of expert assessments has been shown to be only moderate, at best. Aim: We propose a low-cost, contactless system using smartphone videos to automatically determine the presence of bradykinesia. Methods: We collected 70 videos of finger-tap assessments in a clinical setting (40 Parkinson's hands, 30 control hands). Two clinical experts in Parkinson's, blinded to the diagnosis, evaluated the videos to give a grade of bradykinesia severity between 0 and 4 using the Unified Pakinson's Disease Rating Scale (UPDRS). We developed a computer vision approach that identifies regions related to hand motion and extracts clinically-relevant features. Dimensionality reduction was undertaken using principal component analysis before input to classification models (Naïve Bayes, Logistic Regression, Support Vector Machine) to predict no/slight bradykinesia (UPDRS = 0–1) or mild/moderate/severe bradykinesia (UPDRS = 2–4), and presence or absence of Parkinson's diagnosis. Results: A Support Vector Machine with radial basis function kernels predicted presence of mild/moderate/severe bradykinesia with an estimated test accuracy of 0.8. A Naïve Bayes model predicted the presence of Parkinson's disease with estimated test accuracy 0.67. Conclusion: The method described here presents an approach for predicting bradykinesia from videos of finger-tapping tests. The method is robust to lighting conditions and camera positioning. On a set of pilot data, accuracy of bradykinesia prediction is comparable to that recorded by blinded human experts.