Abstract: | The thesis summarizes state-of-the art of the parametric model creation with the Neural Mod-
eling Fields (NMF) approach. The thesis contains a detailed description of the NMF as a
method for parametric model creation including relation to NMF interpretation as model whose
formalisms allows to use the NMF for a modeling mind processes.
The thesis contains proof of equivalence of the NMF and an algorithm known as Expectation-
Maximization (EM). The proved equivalence allowed to take over more easily some relations for
the Gamma, Exponential, Normal, Log-normal, von Mises-Fisher, Wishart and Dirichlet proba-
bility distribution functions for which the algorithm is implemented. Enumerated distributions
are analysed in detail with in relation to parametric model creation. For Exponential, multivari-
ate Log-normal and Dirichlet distributions no relevant resource about parametric model creation
in a mixture of densities was found thus the thesis contains these equations including derivation
of the equations.
Experimental part contains evaluation of the supervised parametric model based on Hierarchical
Mixture of Experts for approximation problems whose learning is based on maximum likelihood
principle. Further the thesis contains experiments of the NMF for Normal distribution in a
hierarchy with Kohonen’s Self Organizing Map, known as MLANS, as a classifier of input images
where the data are based on Feature Integration Theory. The last experiment is region classifier
of images based on feature matrices classified by a mixture of Wishart distribution.
|
---|