Improving Neural Network Performances - Training with Negative Examples

Abstract

The paper describes a new concept for training the neural networks, which can be concretized by using a new modified classical architecture. The idea considered to be the basis of this new architecture was taken over from learning machine, namely that for defining a concept, we need both negative and positive examples. Neural networks are models that are trained only with positive examples and allow the recognition of new examples using the learning machine. Training neural networks with negative examples aims at preventing the development of some specific features of these examples and at obtaining a better recognition of the positive examples. The architecture developed through this method is generic and can be applied to any type of neural network. For simplicity and for the need of obtaining immediate results, a multilayer perceptron was chosen for testing. The results achieved with the help of this network are encouraging and they open new possibilities of study for the future.