Neural network objective functions for detection problems
David M. Weber, Johannes Breitenbach · 2002
We examine the effects of the choice of neural network objective (criterion) functions on the ability of the neural network to perform detection. The experiments are performed using a multilayer perceptron with mean square error, classification figure of merit (CFM), maximally flat CFM and modified perceptron error objective functions. We develop a thresholding scheme for the outputs of the neural network in order to obtain receiver operating characteristic (ROC) curves for the various objective functions. We perform preliminary tests on a breast cancer cell detection problem.