An Analysis of Representation Shift in Concept Learning

In spite of the importance of representation in learning, little progress has been made toward understanding what makes representations work. This paper describes a framework for knowledge-level analysis of changes in the representation of training examples in concept learning. This a very fundamental sort of representation change; such a change alters the very space over which learning occurs, find hence necessitates selection of a new hypothesis space and (probably) a new learning algorithm. The goals of this paper are first, to provide a framework for analysis of representation shifts; second, to make explicit the assumptions implicit in representation shifts that have actually been used in learning systems; and third, to suggest a procedure for finding the most appropriate representation shift, given some background knowledge about a learning problem. The analytic framework is used to analyze a class of hybrid EBL/SBL systems by characterizing the sorts of domain theories that can be used with these systems.