A novel approach for simplifying neural networks by identifying decoupling inputs

Authors:
Sanggil Kang;Wonil Kim
Affiliations:
Department of Computer, College of Information Engineering, The University of Suwon, Suwon, Gyeonggi-do, Korea;Dept of Digital Contents, College of Electronics and Information Engineering, Sejong University, Seoul, Korea
Venue:
AI'04 Proceedings of the 17th Australian joint conference on Advances in Artificial Intelligence
Year:
2004

Citing 3
Cited 0

Optimal brain damage

Advances in neural information processing systems 2
Original Contribution: Improving model selection by nonconvergent methods

Neural Networks
Nonlinear black-box modeling in system identification: a unified overview

Automatica (Journal of IFAC) - Special issue on trends in system identification

Quantified Score

Hi-index	0.00

Visualization

Abstract

This paper proposes a novel approach for modeling partially connected feedforward neural networks (PCFNNs) by identifying input type which refers to whether an input is coupled or uncoupled with other inputs The identification of input type is done by analyzing input sensitivity changes by varying the magnitude of input In the PCFNNs, each input is linked to the neurons in the hidden layer in a different way according to its input type Each uncoupled input does not share the neurons with other inputs in order to contribute to output in an independent manner The simulation results show that PCFNNs outperform fully connected feedforward neural networks with simple network structure.