Weight estimation for the learning of modular perceptron networks

We propose a weight estimation method for feedforward neural networks. The proposed method includes two steps: (1) weight vector orientation estimation and (2) weight vector length estimation, such that the iteration learning of modular perceptron networks (MPN) can be reduced. We have applied the proposed method to the divide-and-conquer learning (DCL) for two-spiral problems (TPS) on an MPN. The experimental results show that the number of pattern presentations can be reduced 79.05% (from 190,980 to 40,012), and the number of subnets in MPN can also be reduced 20.38% (from 26.5 to 21.1).