A Generalized Feedforward Neural Network Architecture and Its Training Using Two Stochastic Search Methods

Abdesselam Bouzerdoum¹ and Rainer Mueller^2*

¹School of Engineering and Mathematics
Edith Cowan University,
Perth, WA, Australia
a.bouzerdoum@ieee.org

²University of Ulm,
Ulm, Germany

Abstract. Shunting Inhibitory Artificial Neural Networks (SIANNs) are biologically inspired networks in which the synaptic interactions are mediated via a nonlinear mechanism called shunting inhibition, which allows neurons to operate as adaptive nonlinear filters. In this article, The architecture of SIANNs is extended to form a generalized feedforward neural network (GFNN) classifier. Two training algorithms are developed based on stochastic search methods, namely genetic algorithms (GAs) and a randomized search method. The combination of stochastic training with the GFNN is applied to four benchmark classification problems: the XOR problem, the 3-bit even parity problem, a diabetes dataset and a heart disease dataset. Experimental results prove the potential of the proposed combination of GFNN and stochastic search training methods. The GFNN can learn difficult classification tasks with few hidden neurons; it solves perfectly the 3-bit parity problem using only one neuron.

^*R. Mueller was a visiting student at ECU for the period July 2001 to June 2002.

LNCS 2723, p. 742 ff.

Full article in PDF

lncs@springer.de
© Springer-Verlag Berlin Heidelberg 2003