Classification with Softmax Neural Networks

A. Vesely (Czech Republic)

Keywords

Neural networks, neural network with softmax neurons, classification of mutually excluded classes, sumofsquare error function, crossentropy error function.

Abstract

In a classification task different multilayer neural network architectures can be used and different error functions can be during network training minimized. For mutually excluded classes the architecture with softmax output neurons minimizing the cross-entropy function is supposed to be the best choice. In the paper we show on a real medical diagnostic problem, taken from the field of mammography, that in some applications the same decision error can be obtained using more traditional multilayer architectures trained to minimize sum-of square error function.

Important Links:



Go Back