A Model for Determining the Number of Negative Examples used in Training a MLP

Abstract

In general, a MLP training uses a training set containing only positive examples, which may change the neural network into an over confident network for solving the problem. A simple solution for this problem is the introduction of negative examples in the training set. Through this procedure, the network will be prepared for the cases it has not been trained for. Unfortunately, up to the present, the number of negative examples that must be used in the training process was not mentioned in the literature. Consequently, the present article aims at finding a general mathematical pattern for training a MLP with negative examples. With that end in view, we have used a regressive analytic technique in order to analyze the data resulted from training three neural networks for a number of three datasets: a dataset for letter recognition, one for the data supplied by a sonar and a last one for the data resulted from the medical tests for determining diabetes. The pattern testing was performed on a new database for confirming its truthfulness.