In this thesis, the deep belief networks and factored 3-way restricted Boltzmann machine are used for modeling the natural image. When dealing with binary visible units from MNIST handwritten digits, the restricted Boltzmann machine maps the visible units to the stochastic binary hidden units which are better representations for the information. The hidden units learnt in a lower layer are used as the visible units for a higher layer, and consequently a deep belief network with improving representation is constructed. When dealing with real valued visible units with strong covariance from Berkeley data set, the model maps the visible units through a 3-way interaction to stochastic binary hidden units. This model converges quickly on both data sets, and provides good performance of reconstruction. Another major advantage of the model is that it learns the filters and the feature activities simultaneously in an unsupervised learning manner.