Machine learning despite unknown classes

This paper revisits supervised machine learning for multiclass problems with the assumption that all classes cannot be represented in a training set. This is common in many applications in which there are numerous classes or in which some classes are exceedingly rare. In this paper we propose the use of a decision function to serve in place of the decision boundaries which are used in many machine learning techniques. We demonstrate this technique using Fisher´s iris data and an application to language recognition.