Abstract
Layered feed-forward neural networks (LNNs) have been broadly applied to classification, prediction and other modeling problems. There have been so far, however, few studies that have provided a theoretical interpretation for the application of LNN. Most of the conventional studies have been empirical and the LNNs have been applied just like “black box” machines. This paper discusses the application of LNN to image or remotely sensed data classification. It provides a theoretical interpretation for the LNN classifier in comparison with the conventional classification or discriminant methods. The most distinguished part is the derivation of a generalized form of LNN classifier based on the maximum entropy principle. According to the generalized form, this paper discusses the relationship between the familiar type of LNN classifier employing the sigmoidal activation function and the other types of discriminant models such as the Multinomial Logit Model.
