Researching the feed-orward neural network by using the piecewise linear division (PLD)

The mechanism of conventional feedforward neural networks is discussed from the viewpoint of pattern recognition using piecewise linear division (PLD). The decisive factor of the first hidden layer is emphasized. It is pointed out that in the first hidden layer the neural elements provide a group of hyperplanes to cut the pattern space and the first hidden layer is free from the influence of other layers. The role of the other layers is to organize the subspaces to satisfy the demands for classification. The PLD provides not only a fast method to determine the connecting weight sets, but also can establish the structure for the feedforward neural network according to the complexity of the classification problem. Some examples are presented to explain the PLD method and others are given to compare it with the conventional NN method