Representation by correspondence: An inadequate conception of knowledge for artificial systems

Abstract

In the Artificial Intelligence community knowledge is normally viewed as structures in the mind (symbols, features, images, etc.) that correspond to structures in the environment. I argue that the standard view is inadequate and that it cannot aid in the construction of truly intelligent systems. Representations by correspondence require prior knowledge of the structure in the mind, the structure in the world, and of the correspondence between them. Unless some other kind of knowledge is already available to the system, it can have no knowledge by correspondence. Various derivatives of this fundamental problem are discussed, including the proliferation of correspondences; the need to posit an observer; the inability to account for error from the system’s point of view; and the radical incompatibility between representation by correspondence and evolutionary or developmental accounts of knowledge.