Partitioning Strategies for Modular Neural Networks

Abstract-We observe the effects of a variety of splitting strategies for partitioning the input domain in a self-splitting modular neural network applied to the two-spiral classification problem, and assisted by a special-purpose visualization tool. The observations motivate the development of an improved strategy, consisting of a series of binary splits along the boundaries of trained areas, and a particular weight initialization strategy. The work is leading to fewer networks and better generalization for this application, when backpropagation is used. I. SELF-SPLITTING NEURAL NETWORKS HE Self-Splitting Neural Network (SSNN) is a particular version of a modular neural network, and as such divides the input domain into partitions where a separate network is assigned to each partition. In A variety of methods have been used for partitioning the input domain. Chen et al. used divide-and-conquer to carve up the input space heuristically based on features of the problem being solve