Selective weight update rule for hybrid neural network

Abstract

VSF-Network,Vibration Synchronizing Function Network, is a hybrid neural network combining a chaos neural network with a hierarchical network. It is a neural network model which learns symbols. In this paper, the two theoretical backgrounds of VSF-Network are described. The first one is the incremental learning by CNN and the second background is ensemble learning. VSF-Network finds unknown parts of input data by comparing to learned pattern and it learns the unknown parts using unused part of the network. By the ensemble learning, the capability of VSF-network for recognizing combined patterns that are learned by every sub-network of VSF-network can be explained. Through the experiments, we show that VSF-network can recognize combined patterns only if it has learned parts of the patterns and show factors for affecting performance of the learning. © 2012 Springer-Verlag.