Improving neural networks through modularization

• Main Author: Ong, Thian Hock.
• Other Authors: Chen, Yan Qiu
• Format: Theses and Dissertations
• Published: 2008
• Subjects: DRNTU::Engineering::Electrical and electronic engineering::Computer hardware, software and systems
• Online Access: http://hdl.handle.net/10356/5011
• Institution: Nanyang Technological University

This thesis explores and investigates various possibilities to modularize neural net-works to improve its performance and enhance its capability for pattern classifica-tion and function approximation. We have developed a novel Class-Modularized Neural Network, which utilizes one subnetwork for each class of samples. Where conventional approach generates only the class decision, the CMNN is also capable of providing an indication of the confidence of the class decision. For each class of samples (or subnetwork), it performs the task in two stages: (1) estimation of the class conditional probability density function from the training samples; (2) approximation of the probability density function in (1).