It is well known that a neural network can learn Bayesian discriminant functions. In the two-category normal-distribution case, a shift by a constant of the logit transform of the network output approximates a corresponding Mahalanobis discriminant function [7]. In [10], we have proposed an algorithm for estimating the constant, but it requires the network to be trained twice, in one of which the teacher signals must be shifted by the mean vectors. In this paper, we propose a more efficient algorithm for estimating the constant with which the network is trained only once.
展开▼
