A multi-network architecture for high generalization in pattern recognition with backpropagation neural network modules

Backpropagation networks are the most popular multi-layer networks, used for either function approximation or pattern classification. They are trained and tested using two disjoint sets of patterns drawn randomly from the pattern space. In many cases, the overtraining phenomenon occurs i.e. the network learns to produce the proper output for the patterns to which it has been trained but it produces meaningless outputs for unforeseen patterns. In this paper, the overtraining phenomenon is analyzed in depth, and an alternative architecture with increased generalization ability is proposed.