Can the classification capability of network be further improved by using quadratic sigmoidal neurons?

摘要

In Ref. [4], by using constructive method, Chiang et al., proved that a three-layer neural network containing k+1 single threshold quadratic sigmoidal hidden neurons and one multithreshold sigmoidal output neuron could learn arbitrary dichotomy defined on a training set of 4k patterns. In this paper the classification capability of the feed forward neural networks containing multiple or single threshold quadratic sigmoidal neurons in the hidden and output layer is evaluated. The degree of improvement on the classification capability of network by using quadratic sigmoidal neurons is analyzed.