A Mean Field Approximation of BCPNN
Anders Lansner, Christopher Johansson · 2005
In this report we study a mean field (MF) approximation of the Bayesian Confidence Propagating Neural Network (BCPNN) for which we derive the energy function. This MF approximation is compared with the original formulation of the network in a number of different tasks in order to establish the similarities and dissimilarities. We investigate the effect of different updating strategies on the storage capacity. Three different ways of modulating the attractor size are experimentally tested. We apply the networks to prototype extraction. Finally, we investigate how the networks cluster data. These experiments show that there are some differences between BCPNN and its MF approximation. Furthermore, the experiments provide some new knowledge on the clustering of memories in a BCPNN.