Learning of Mahalanobis Discriminant Functions by a Neural Network

It is known that a neural network can learn a Bayesian discriminant function. Ito et al. (2006) has pointed out that if the inner potential of the output unit of the network is shifted by a constant, the output becomes a Mahalanobis discriminant function. However, it was a heavy task for the network to calculate the constant. Here, we propose a new algorithm with which the network can estimate the constant easily. This method can be extended to higher dimensional classificasions problems without much effort.