Training of a neural network for pattern classification based on an entropy measure

A neural net model for pattern classification is introduced. Unlike models in which the network topology is specified before training, in this model the network expands during training. The proposed model introduces a novel type of unit (neuron) and a standard treelike feedforward network topology. The simplicity of the interconnection pattern is a particular advantage over existing models. Internal representations are formed by separating hyperplanes. Selection of the hyperplanes and expansion of the network is based on an entropy measure which is appropriately defined. The weight vectors of all units with a certain layer are determined in a single presentation of the training set.<>