On The Design of Supra - Classi ers for

| We have recently introduced a framework for the reuse of knowledge from previously trained classiiers to improve performance in a current, possibly related classi-cation task. This framework requires the use of a supra-classiier, which makes a classiication decision based on the outputs of a large number of previously trained diverse classiiers. We discuss the performance requirements of a good supra-classiier and introduce several possible supra-classiier architectures. We make performance comparisons of these architectures using public domain data sets for the problem of inadequate training data and compare their scal-ability in the number of simultaneously reused classiiers.