A Novel Fast Constructive Algorithm for Neural Classifier

This paper presents a novel fast algorithm to construct feedforward neural networks for pattern classification tasks. The algorithm constructs the neural network by adding elementary neural elements to the first layer, in response to the distribution of the patterns in the training set. Each elementary neural element of the network is trained with different pattern subsets and forms a hyper-plane in the input space. The decision regions of these elementary neural elements are restricted and combined through the second layer of simple fixed weights. This network performs the nonlinear map of input to output and realizes a piecewise linear classifier. The learning rule for the weights is very simple and the network can be very fast constructed and trained. In addition, the algorithm automatically determines the number of hidden nodes.