University of Ioannina
Μπεϋζιανές μέθοδοι για προβλήματα μηχανικής μάθησης και επεξεργασίας εικόνας
Abstract
dc:descriptionIn this thesis, we study the sparse Bayesian linear model for regression and classification tasks and for solving image processing problems. We start with an overview of the Bayesian inference methodology and its application to linear regression. We then develop a computationally efficient training algorithm for sparse Bayesian regression of images. The proposed training algorithm uses operations in the Fourier domain and the conjugate gradients method, in order to allow regression of large images at reasonable computational cost. We then apply this algorithm to detect objects in images, using a variant of the relevance vector machine (RVM), which uses many types of kernels simultaneously and we call the multikernel RVM. Next, we use the sparse Bayesian linear model to estimate the point spread function (PSF) in the blind image deconvolution (BID) problem. We propose a Bayesian model that estimates the support of the blurring PSF, allows reconstruction of image edges and achieves noise robustness. Bayesian inference on this model is performed using the variational approximation. Furthermore, we focus on the problem of selecting appropriate basis functions for the sparse Bayesian linear model, which is crucial in order to achieve good generalization performance. Typically, this problem is tackled using cross-validation technique, but this technique is computationally expensive and it can be used to select the best out of a small number of candidate basis function sets. Instead, we propose an adaptive kernel learning algorithm, which is similar to the RVM but also learns the parameters of the kernels during model training. More specifically, the proposed method estimates different parameter values for each kernel, resulting in a very flexible model. In order to avoid overfitting, a sparsity enforcing prior is imposed that controls the effective number of parameters of the model. The proposed methodology is evaluated on benchmark regression and classification datasets and applied for analysis of functional magnetic resonance images (fMRI). We also propose a modification of this method that performs local feature selection by estimating the parameters of appropriate kernel functions. In order to incorporate local feature selection, anisotropic Gaussian kernels are considered, which use a separate scaling factor (width) for each feature. Feature selection is local, in the sense that different features are assumed to be significant at different regions of the input space. This is achieved because we learn different values for the scaling factors of each kernel. In order to eliminate irrelevant features, we assume a sparsity enforcing prior on the scaling factors of the kernels. The proposed adaptive kernel learning algorithms (with or without local feature selection) are employed to the problem of classifying DNA microarray datasets.
Degree
thesis:*- Grantor dc:publisher
- University of Ioannina
- Year dc:date
- 2009
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Tzikas, Dimitrios
- Τζίκας, Δημήτριος
Subjects
dc:subject × 14Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/17830