A technique for creating and initializing hidden units for neural net pattern classification problems

Summary form only given, as follows. A significant problem associated with the application of the standard backpropagation algorithm for pattern classification is the lack of a rationale for determining the number of hidden units necessary to obtain an acceptable level of performance. A procedure is presented for developing neural networks that estimate from the training data the number of hidden units needed for a given two-class problem and also provide an initial estimate of the weights of all hidden units in the first layer of the network. The method is developed for the general case where the underlying input space is n-dimensional. Computer simulations are given for a neural network with a single layer of hidden units and a two-dimensional pattern space.<>