A growing cell neural network structure with backpropagation learning algorithm

A classical model used for pattern recognition is the Multilayer Perceptron Artificial Neural Network (ANN), with backpropagation learning algorithm. However the recognition performance of this ANN strongly depends on the number of neurons used in the hidden layer, is a function of the particular problem to be solved. Then, in most cases this number is unknown in advance. A possible solution for this problem may be the use of growing cells structures, such as those used in the solution of some classification problems with auto organizing ANN. Using a similar idea, we propose a growing cell multilayer ANN in which a modified backpropagation algorithm is used to optimize the ANN weights matrix, as well as the number of neurons in the hidden layer. A proposed approach also reduces the computational complexity of conventional multilyer perceptron with backpropagation learning algorithms during the training stage, using at the beginning a reduced number of neurons in the hidden layer to minimize the identification error. This number is then increased until reach its optimal size. The proposed structure was evaluated with some benchmarks problems such as XOR, four different classes classification and ZIGZAG problems.