Implementation of a Bayesian self-organizing neural network

Presents the implementation of a self-organizing neural network, that extracts invariant features in the input images. The prior probability distribution of the invariant feature is assumed to be known, and is modeled using a Markov random field. The network learns to extract the desired invariant feature by producing output patterns whose probability distribution is as close to the prior probability distribution as possible. In the author's simulations, an input image is an object surface corrupted by (1) noise, and (2) a distracting feature. As a result, the shapes of the objects can not be accurately represented by the input intensities. The network learns to recover the surface shape, which is the invariant feature that the author wishes to extract. Experimental results of the network learning are presented and analyzed.<>