A minimum classification error, maximum likelihood, neural network
H. Gish · 1992
The authors present a method for training neural networks to minimize classification errors. The method is based on a maximum likelihood (ML) training algorithm. The ML criterion is interpreted as a distance measure of the data points to the decision boundary. This view leads to a modified network that will minimize classification errors when trained with the ML criterion. The robustness properties of the minimum error network are discussed and illustrated.>