A new type of Multilayer network including certain class of Radial Basis Units (RBU), whose kernels are implemented at the synaptic level, is compared through simulations with the Multi-Layer Perceptron (MLP) in a classification problem with a high interference of class distributions. The simulations show that the new network gives error rates in the classification near those of the Optimum Bayesian Classifier (OBC), while MLP presents an inherent weakness for these classification tasks.