Abstract

Lately, the enormous generation of databases in almost every aspect of life has created a great demand for new, powerful tools for turning data into useful information. Therefore, researchers were encouraged to explore and develop new machine learning ideas and methods. Mixture models are one of the machine learning techniques receiving considerable attention due to their ability to
handle efficiently and effectively multidimensional data. Generally, four critical issues have to be addressed when adopting mixture models in high dimensional spaces: (1) choice of the probability density functions, (2) estimation of the mixture parameters, (3) automatic determination of the number of components M in the mixture, and (4) determination of what features best discriminate among the different components. The main goal of this thesis is to summarize all these challenging
interrelated problems in one unified model.
In most of the applications, the Gaussian density is used in mixture modeling of data. Although a Gaussian mixture may provide a reasonable approximation to many real-world distributions, it is certainly not always the best approximation especially in computer vision and image processing applications where we often deal with non-Gaussian data. Therefore, we propose to use three
highly flexible distributions: the generalized Gaussian distribution (GGD), the asymmetric Gaussian distribution (AGD), and the asymmetric generalized Gaussian distribution (AGGD). We are motivated by the fact that these distributions are able to fit many distributional shapes and then can be considered as a useful class of flexible models to address several problems and applications involving measurements and features having well-known marked deviation from the Gaussian shape.
Recently, researches have shown that model selection and parameter learning are highly dependent and should be performed simultaneously. For this purpose, many approaches have been suggested. The vast majority of these approaches can be classified, from a computational point of view, into two classes: deterministic and stochastic methods. Deterministic methods estimate
the model parameters for a set of candidate models using the Expectation-Maximization (EM) framework, then choose the model that maximizes a model selection criterion. Stochastic methods such as Markov chain Monte Carlo (MCMC) can be used in order to sample from the full a posteriori distribution with M considered unknown. Hence, in this thesis, we propose three learning techniques capable of automatically determining model complexity while learning its parameters. First, we incorporate a Minimum Message Length (MML) penalty in the model learning step performed using the EM algorithm. Our second approach employs the Rival Penalized EM (RPEM)
algorithm which is able to select an appropriate number of densities by fading out the redundant densities from a density mixture. Last but not least, we incorporate the nonparametric aspect of mixture models by assuming a countably infinite number of components and using Markov Chain Monte Carlo (MCMC) simulations for the estimation of the posterior distributions. Hence, the difficulty of choosing the appropriate number of clusters is sidestepped by assuming that there are an infinite number of mixture components.
Another essential issue in the case of statistical modeling in general and finite mixtures in particular is feature selection (i.e. identification of the relevant or discriminative features describing the data) especially in the case of high-dimensional data. Indeed, feature selection has been shown to be a crucial step in several image processing, computer vision and pattern recognition
applications not only because it speeds up learning but also because it improves model accuracy and generalization. Moreover, the learning of the mixture parameters ( i.e. both model selection and parameters estimation) is greatly affected by the quality of the features used. Hence, in this thesis, we are trying to solve the feature selection problem in unsupervised learning by casting it as an estimation problem, thus avoiding any combinatorial search. Finally, the effectiveness of our approaches is evaluated by applying them to different computer vision and image processing
applications.