In this paper, a novel neural network architecture designed for pattern classification pruposes is presented. The classifier is a two-layer neural network. The first layer classifies the input vectors into a number of clusters using a stochastic competitive learning algorithm. The output of this layer is a Gibbs probability distribution for the association of the input with these clusters. The output of the first layer is used as input to the second layer that implements a classifier similar to the Bayes minimum risk classifier. A new complementary Hebbian learning algorithm is proposed to train the second layer. Computer simulations have been performed and thd results demonstrate that the new classifier consistently provides high correct recognition rates and is competitive to other similar systems.
展开▼