Recognition of membership in classes

This paper presents an approach to the general problem of recognition of membership in classes which are known only from a set of their examples. A geometrical approach is taken where membership in classes is regarded measurable by metrics with which a set of points, representing different members of the same class, may be brought "close" to one another. For the case where classes are Gaussian processes, the method described herein and that of decision theory are found to agree. A practical application of the method to the automatically "learned" recognition of spoken numerals is described.