Neural networks are now accepted as a viable alternative to more traditional and conventional solution approaches in the area of pattern classification (e.g classification, signal recognition, discriminant analysis). Such networks possess the significant advantage of being nonparametric—as well as being easy to use and understand. Network development is typically accomplished by means of a two-phase approach. In the first, a network architecture is selected. In the second, a training exercise is conducted so as to establish the weights on the network branches. However, as with any tool for analysis, neural-network classifiers have certain drawbacks. Included among these is the quite ad hoc nature of the design and training prcnxss. Another problem is one wmrnon not only to neural networks, but also inherent in most any other methods of classification. This is the fact that such methods generally ‘do not know what they do not know’. As a consequence, even when faced with new cases that are very much unlike anything that they have k e n trained upon, they still do not hesitate to boldly (and often foolishly) produce a classification. In this paper, we present an approach-that simultaneous designs and trains ontogenic neural network classifiers. Ontonenic neural networks. in turn. have two features of articular interest. First, they virtually design and train themselves. Second, they are iesitant to classify objects that appear to be ‘too dissimilar’ from those upon which they were trained. We demonstrate the performance of such an approach on a very specific problem: the classification of firms with regard to their future fiscal well-being.
Copyright (C) 1996, Institute of Mathematics and its Applications
In order to set up a list of libraries that you have access to,
you must first login
or sign up.
Then set up a personal list of libraries from your profile page by
clicking on your user name at the top right of any screen.