A PCA-based Optimal Design of Network for Learning

Description

Traditional studies on multi-layered neural networks (MNN) are mainly restricted to regular topology, with less consideration to the structural effects. The recent progress of complex networks could provide a novel view on the relation between the network structure and its function. Unfortunately, this work has not yet been applied to MNN.This project is a preliminary study to investigate the effects of non-regular structure and explore a possible methodology to design a near optimal structure of MNN for the given learning task. The concept of principal component analysis (PCA) will be used to develop a network prune method to eliminate the redundancy in the network. Topologic analysis of the complex network will be used to introduce the global short cut connection to analyze the structural effects. Finally a PCA-based iterative network design method will be proposed to explore the optimal structure for the network. The project aims to design a proper network structure for learning by using hybrid knowledge from traditional neural learning, PCA techniques, and topological analysis of the complex network.