Simple Auto-Associative Networks Succeed at Universal Generalization of the Identity Function and Reduplication Rule

Abstract

It has become widely accepted that standard connectionist models are unable to show identity-based relational reasoning that requires universal generalization. The purpose of this brief report is to show how one of the simplest forms of such models, feed-forward auto-associative networks, satisfies two of the most well-known challenges: universal generalization of the identity function and the reduplication rule. Given the simplicity of the modeling account provided, along with the clarity of the evidence, these demonstrations invite a shift in this high-profile debate over the nature of cognitive architecture and point to a way to bridge some of the presumed gulf between characteristically symbolic forms of reasoning and connectionist mechanisms.