Since its discovery in 1995 resting state functional connectivity derived from functional MRI data has become a popular neuroimaging method for study psychiatric disorders. Current methods for analyzing resting state functional connectivity in disease involve thousands of univariate tests, and the specification of regions of interests to employ in the analysis. There are several drawbacks to these methods. First the mass univariate tests employed are insensitive to the information present in distributed networks of functional connectivity. Second, the null hypothesis testing employed to select functional connectivity dierences between groups does not evaluate the predictive power of identified functional connectivities. Third, the specification of regions of interests is confounded by experimentor bias in terms of which regions should be modeled and experimental error in terms of the size and location of these regions of interests. The objective of this dissertation is to improve the methods for functional connectivity analysis using multivariate predictive modeling, feature selection, and whole brain parcellation. A method of applying Support vector classification (SVC) to resting state functional connectivity data was developed in the context of a neuroimaging study of depression. The interpretability of the obtained classifier was optimized using feature selection techniques that incorporate reliability information. The problem of selecting regions of interests for whole brain functional connectivity analysis was addressed by clustering whole brain functional connectivity data to parcellate the brain into contiguous functionally homogenous regions. This newly developed famework was applied to derive a classifier capable of correctly seperating the functional connectivity patterns of patients with depression from those of healthy controls 90% of the time. The features most relevant to the obtain classifier match those previously identified in previous studies, but also include several regions not previously implicated in the functional networks underlying depression.

Speech communication encompasses diverse types of information, including phonetics, affective state, voice quality, and speaker identity. From a speech production standpoint, the acoustic speech signal can be mainly divided into glottal source and vocal tract components, which play distinct roles in rendering the various types of information it contains. Most deployed speech analysis systems, however, do not explicitly represent these two components as distinct entities, as their joint estimation from the acoustic speech signal becomes an ill-defined blind deconvolution problem. Nevertheless, because of the desire to understand glottal behavior and how it relates to perceived voice quality, there has been continued interest in explicitly estimating the glottal component of the speech signal. To this end, several inverse filtering (IF) algorithms have been proposed, but they are unreliable in practice because of the blind formulation of the separation problem. In an effort to develop a method that can bypass the challenging IF process, this thesis proposes a new glottal source information extraction method that relies on supervised machine learning to transform smoothed spectral representations of speech, which are already used in some of the most widely deployed and successful speech analysis applications, into a set of glottal source features. A transformation method based on Gaussian mixture regression (GMR) is presented and compared to current IF methods in terms of feature similarity, reliability, and speaker discrimination capability on a large speech corpus, and potential representations of the spectral envelope of speech are investigated for their ability represent glottal source variation in a predictable manner. The proposed system was found to produce glottal source features that reasonably matched their IF counterparts in many cases, while being less susceptible to spurious errors. The development of the proposed method entailed a study into the aspects of glottal source information that are already contained within the spectral features commonly used in speech analysis, yielding an objective assessment regarding the expected advantages of explicitly using glottal information extracted from the speech signal via currently available IF methods, versus the alternative of relying on the glottal source information that is implicitly contained in spectral envelope representations.

m-Estimation represents a broad class of estimators, including least-squares and maximum likelihood, and is a widely used tool for statistical inference. Its successful application however, often requires negotiating physical resources for desired levels of accuracy. These limiting factors, which we abstractly refer as costs, may be computational, such as time-limited cluster access for parameter learning, or they may be financial, such as purchasing human-labeled training data under a fixed budget. This thesis explores these accuracy- cost tradeoffs by proposing a family of estimators that maximizes a stochastic variation of the traditional m-estimator. Such "stochastic m-estimators" (SMEs) are constructed by stitching together different m-estimators, at random. Each such instantiation resolves the accuracy-cost tradeoff differently, and taken together they span a continuous spectrum of accuracy-cost tradeoff resolutions. We prove the consistency of the estimators and provide formulas for their asymptotic variance and statistical robustness. We also assess their cost for two concerns typical to machine learning: computational complexity and labeling expense. For the sake of concreteness, we discuss experimental results in the context of a variety of discriminative and generative Markov random fields, including Boltzmann machines, conditional random fields, model mixtures, etc. The theoretical and experimental studies demonstrate the effectiveness of the estimators when computational resources are insufficient or when obtaining additional labeled samples is necessary. We also demonstrate that in some cases the stochastic m-estimator is associated with robustness thereby increasing its statistical accuracy and representing a win-win.

A wide variety of text analysis applications are based on statistical machine learning techniques. The success of those applications is critically affected by how we represent a document. Learning an efficient document representation has two major challenges: sparsity and sequentiality. The sparsity often causes high estimation error, and text's sequential nature, interdependency between words, causes even more complication. This thesis presents novel document representations to overcome the two challenges. First, I employ label characteristics to estimate a compact document representation. Because label attributes implicitly describe the geometry of dense subspace that has substantial impact, I can effectively resolve the sparsity issue while only focusing the compact subspace. Second, while modeling a document as a joint or conditional distribution between words and their sequential information, I can efficiently reflect sequential nature of text in my document representations. Lastly, the thesis is concluded with a document representation that employs both labels and sequential information in a unified formulation. The following four criteria are utilized to evaluate the goodness of representations: how close a representation is to its original data, how strongly a representation can be distinguished from each other, how easy to interpret a representation by a human, and how much computational effort is needed for a representation. While pursuing those good representation criteria, I was able to obtain document representations that are closer to the original data, stronger in discrimination, and easier to be understood than traditional document representations. Efficient computation algorithms make the proposed approaches largely scalable. This thesis examines emotion prediction, temporal emotion analysis, modeling documents with edit histories, locally coherent topic modeling, and text categorization tasks for possible applications.

The term “Google” has become a verb for most of us. Search engines, however, have certain limitations. For example ask it for the impact of the current global financial crisis in different parts of the world, and you can expect to sift through thousands of results for the answer. This motivates the research in complex question answering where the purpose is to create summaries of large volumes of information as answers to complex questions, rather than simply offering a listing of sources. Unlike simple questions, complex questions cannot be answered easily as they often require inferencing and synthesizing information from multiple documents. Hence, this task is accomplished by the query-focused multidocument summarization systems. In this thesis we apply different supervised learning techniques to confront the complex question answering problem. To run our experiments, we consider the DUC-2007 main task. A huge amount of labeled data is a prerequisite for supervised training. It is expensive and time consuming when humans perform the labeling task manually. Automatic labeling can be a good remedy to this problem. We employ five different automatic annotation techniques to build extracts from human abstracts using ROUGE, Basic Element (BE) overlap, syntactic similarity measure, semantic similarity measure and Extended String Subsequence Kernel (ESSK). The representative supervised methods we use are Support Vector Machines (SVM), Conditional Random Fields (CRF), Hidden Markov Models (HMM) and Maximum Entropy (MaxEnt). We annotate DUC-2006 data and use them to train our systems, whereas 25 topics of DUC-2007 data set are used as test data. The evaluation results reveal the impact of automatic labeling methods on the performance of the supervised approaches to complex question answering. We also experiment with two ensemble-based approaches that show promising results for this problem domain. / x, 108 leaves : ill. ; 29 cm

Active Learning arises as an important issue in several supervised learning scenarios where obtaining data is cheap, but labeling is costly. In general, this consists in a query strategy, a greedy heuristic based on some selection criterion, which searches for the potentially most informative observations to be labeled in order to form a training set. A query strategy is therefore a biased sampling procedure since it systematically favors some observations by generating biased training sets, instead of making independent and identically distributed draws. The main hypothesis of this thesis lies in the reduction of the bias inherited from the selection criterion. The general proposal consists in reducing the bias by selecting the minimal training set from which the estimated probability distribution is as close as possible to the underlying distribution of overall observations. For that, a novel general active learning query strategy has been developed using an Information-Theoretic framework. Several experiments have been performed in order to evaluate the performance of the proposed strategy. The obtained results confirm the hypothesis about the bias, showing that the proposal outperforms the baselines in different datasets.

The aim of the study is to apply and compare the performance of two different types of neural networks on the Quick, Draw! dataset and from this determine whether interpreting the sketches as sequences gives a higher accuracy than interpreting them as pixels. The two types of networks constructed were a recurrent neural network (RNN) and a convolutional neural network (CNN). The networks were optimised and the final architectures included five layers. The final evaluation accuracy achieved was 94.2% and 92.3% respectively, leading to the conclusion that the sequential interpretation of the Quick, Draw! dataset is favourable.

Le credit scoring est généralement considéré comme une méthode d’évaluation du niveau du risque associé à un dossier de crédit potentiel. Cette méthode implique l'utilisation de différentes techniques statistiques pour aboutir à un modèle de scoring basé sur les caractéristiques du client.Le modèle de scoring estime le risque de crédit en prévoyant la solvabilité du demandeur de crédit. Les institutions financières utilisent ce modèle pour estimer la probabilité de défaut qui va être utilisée pour affecter chaque client à la catégorie qui lui correspond le mieux: bon payeur ou mauvais payeur. Les seules données disponibles pour construire le modèle de scoring sont les dossiers acceptés dont la variable à prédire est connue. Ce modèle ne tient pas compte des demandeurs de crédit rejetés dès le départ ce qui implique qu'on ne pourra pas estimer leurs probabilités de défaut, ce qui engendre un biais de sélection causé par la non-représentativité de l'échantillon. Nous essayons dans ce travail en utilisant l'inférence des refusés de remédier à ce biais, par la réintégration des dossiers refusés dans le processus d'octroi de crédit. Nous utilisons et comparons différentes méthodes de traitement des refusés classiques et semi supervisées, nous adaptons certaines à notre problème et montrons sur un jeu de données réel, en utilisant les courbes ROC confirmé par simulation, que les méthodes semi-supervisé donnent de bons résultats qui sont meilleurs que ceux des méthodes classiques. / Credit scoring is generally considered as a method of evaluation of a risk associated with a potential loan applicant. This method involves the use of different statistical techniques to determine a scoring model. Like any statistical model, scoring model is based on historical data to help predict the creditworthiness of applicants. Financial institutions use this model to assign each applicant to the appropriate category : Good payer or Bad payer. The only data used to build the scoring model are related to the accepted applicants in which the predicted variable is known. The method has the drawback of not estimating the probability of default for refused applicants which means that the results are biased when the model is build on only the accepted data set. We try, in this work using the reject inference, to solve the problem of selection bias, by reintegrate reject applicants in the process of granting credit. We use and compare different methods of reject inference, classical methods and semi supervised methods, we adapt some of them to our problem and show, on a real dataset, using ROC curves, that the semi-supervised methods give good results and are better than classical methods. We confirmed our results by simulation.

La démocratisation de la « vidéosurveillance intelligente » nécessite le développement d’outils automatiques et temps réel d’analyse vidéo. Parmi ceux-ci, la détection de piétons joue un rôle majeur car de nombreux systèmes reposent sur cette technologie. Les approches classiques de détection de piétons utilisent la reconnaissance de formes et l’apprentissage statistique. Elles souffrent donc d’une dégradation des performances quand l’apparence des piétons ou des éléments de la scène est trop différente de celle étudiée lors de l’apprentissage. Pour y remédier, une solution appelée « contextualisation du détecteur » est étudiée lorsque la caméra est fixe. L’idée est d’enrichir le système à l’aide d’informations provenant de la scène afin de l’adapter aux situations qu’il risque de fréquemment rencontrer. Ce travail a été réalisé en deux temps. Tout d’abord, l’architecture d’un détecteur et les différents outils utiles à sa construction sont présentés dans un état de l’art. Puis la problématique de la contextualisation est abordée au travers de diverses expériences validant ou non les pistes d’amélioration envisagées. L’objectif est d’identifier toutes les briques du système pouvant bénéficier de cet apport afin de contextualiser complètement le détecteur. Pour faciliter l’exploitation d’un tel système, la contextualisation a été entièrement automatisée et s’appuie sur des algorithmes d’apprentissage semi-supervisé. Une première phase consiste à collecter le maximum d’informations sur la scène. Différents oracles sont proposés afin d’extraire l’apparence des piétons et des éléments du fond pour former une base d’apprentissage dite contextualisée. La géométrie de la scène, influant sur la taille et l’orientation des piétons, peut ensuite être analysée pour définir des régions, dans lesquelles les piétons, tout comme le fond, restent visuellement proches. Dans la deuxième phase, toutes ces connaissances sont intégrées dans le détecteur. Pour chaque région, un classifieur est construit à l’aide de la base contextualisée et fonctionne indépendamment des autres. Ainsi chaque classifieur est entraîné avec des données ayant la même apparence que les piétons qu’il devra détecter. Cela simplifie le problème de l’apprentissage et augmente significativement les performances du système. / With the rise of videosurveillance systems comes a logical need for automatic and real-time processes to analyze the huge amount of generated data. Among these tools, pedestrian detection algorithms are essential, because in videosurveillance locating people is often the first step leading to more complex behavioral analyses. Classical pedestrian detection approaches are based on machine learning and pattern recognition algorithms. Thus they generally underperform when the pedestrians’ appearance observed by a camera tends to differ too much from the one in the generic training dataset. This thesis studies the concept of the contextualization of such a detector. This consists in introducing scene information into a generic pedestrian detector. The main objective is to adapt it to the most frequent situations and so to improve its overall performances. The key hypothesis made here is that the camera is static, which is common in videosurveillance scenarios.This work is split into two parts. First a state of the art introduces the architecture of a pedestrian detector and the different algorithms involved in its building. Then the problem of the contextualization is tackled and a series of experiments validates or not the explored leads. The goal is to identify every part of the detector which can benefit from the approach in order to fully contextualize it. To make the contextualization process easier, our method is completely automatic and is based on semi-supervised learning methods. First of all, data coming from the scene are gathered. We propose different oracles to detect some pedestrians in order to catch their appearance and to form a contextualized training dataset. Then, we analyze the scene geometry, which influences the size and the orientation of the pedestrians and we divide the scene into different regions. In each region, pedestrians as well as background elements share a similar appearance.In the second step, all this information is used to build the final detector which is composed of several classifiers, one by region. Each classifier independently scans its dedicated piece of image. Thus, it is only trained with a region-specific contextualized dataset, containing less appearance variability than a global one. Consequently, the training stage is easier and the overall detection results on the scene are improved.

Submitted by Lucas Guerreiro null (lucasg@rc.unesp.br) on 2017-10-16T22:03:24Z No. of bitstreams: 1 LucasGuerreiro_dissertacao.pdf: 2072249 bytes, checksum: 03cb08b42175616dd567a364cf201bcd (MD5) / Approved for entry into archive by Monique Sasaki (sayumi_sasaki@hotmail.com) on 2017-10-18T18:42:00Z (GMT) No. of bitstreams: 1 guerreiro_l_me_sjrp.pdf: 2072249 bytes, checksum: 03cb08b42175616dd567a364cf201bcd (MD5) / Made available in DSpace on 2017-10-18T18:42:00Z (GMT). No. of bitstreams: 1 guerreiro_l_me_sjrp.pdf: 2072249 bytes, checksum: 03cb08b42175616dd567a364cf201bcd (MD5) Previous issue date: 2017-09-01 / O Aprendizado de Máquina é uma área que vem crescendo nos últimos anos e é um dos destaques dentro do campo de Inteligência Artificial. Atualmente, uma das subáreas mais estudadas é o Aprendizado Semi-Supervisionado, principalmente pela sua característica de ter um menor custo na rotulação de dados de exemplo. A categoria de modelos baseados em grafos é a mais ativa nesta subárea, fazendo uso de estruturas de redes complexas. O algoritmo de competição e cooperação entre partículas é uma das técnicas deste domínio. O algoritmo provê acurácia de classificação compatível com a de algoritmos do estado da arte, e oferece um custo computacional inferior à maioria dos métodos da mesma categoria. Neste trabalho é apresentado um estudo sobre Aprendizado Semi-Supervisionado, com ênfase em modelos baseados em grafos e, em particular, no Algoritmo de Competição e Cooperação entre Partículas (PCC). O objetivo deste trabalho é propor um novo algoritmo de competição e cooperação entre partículas baseado neste modelo, com mudanças na caminhada pelo grafo, com informações de dominância sendo mantidas nas arestas ao invés dos nós; as quais possam melhorar a acurácia de classificação ou ainda o tempo de execução em alguns cenários. É proposta também uma metodologia de avaliação da rede obtida com o modelo de competição e cooperação entre partículas, para se identificar a melhor métrica de distância a ser aplicada em cada caso. Nos experimentos apresentados neste trabalho, pode ser visto que o algoritmo proposto teve melhor acurácia do que o PCC em algumas bases de dados, enquanto o método de avaliação de métricas de distância atingiu também bom nível de precisão na maioria dos casos. / Machine Learning is an increasing area over the last few years and it is one of the highlights in Artificial Intelligence area. Nowadays, one of the most studied areas is Semi-supervised learning, mainly due to its characteristic of lower cost in labeling sample data. The most active category in this subarea is that of graph-based models, using complex networks concepts. The Particle Competition and Cooperation in Networks algorithm (PCC) is one of the techniques in this field. The algorithm provides accuracy compatible with state of the art algorithms, and it presents a lower computational cost when compared to most techniques in the same category. In this project, it is presented a research about semi-supervised learning, with focus on graphbased models and, in special, the Particle Competition and Cooperation in Networks algorithm. The objective of this study is to base proposals of new particle competition and cooperation algorithms based on this model, with new dynamics on the graph walking, keeping dominance information on the edges instead of the nodes; which may improve the accuracy classification or yet the runtime in some situations. It is also proposed a method of evaluation of the network built with the Particle Competition and Cooperation model, in order to infer the best distance metric to be used in each case. In the experiments presented in this work, it can be seen that the proposed algorithm presented better accuracy when compared to the PCC for some datasets, while the proposed distance metrics evaluation achieved a high precision level in most cases.