Recognizing Visual Images through Latent Graphical Models based on the Bayesian Nonparametric Methods

Jinsan Yang and Byoung-Tak Zhang

Abstract

Latent model is a graphical model with one or more layers of latent variables and a layer of input variables. The relationships between the layers could be linear as in the factor analysis model, or could be non-linear for the more general case. When the data comes from several different sources, the corresponding model becomes a mixture of latent variable model. A generative model for this problem is proposed through a two way multi-layer network structure for learning this complicated model. By suitable selection of significant features, the upcoming visual images can be recognized with more accuracies and insights. The advantages of our model over the other learning methods is the flexibility of choosing the size of the latent dimension, the generative feature with nice topographic properties for visulaizations, and finally but not lastly, the clustering ability after suitable settings of model parameters.