Distributed representations of structure: A theory of analogical access and mapping.

Abstract

We describe an integrated theory of analogical access and mapping, instantiated in a computational model called LISA (Learning and Inference with Schemas and Analogies). LISA represents predicates and objects as distributed patterns of activation over units representing semantic primitives. These representations are dynamically bound into propositional structures, thereby achieving the structure-sensitivity of a symbolic system and the flexibility of a connectionist system. LISA also has a number of inherent limitations, including capacity limits and sensitivity to the manner in which a problem is represented. A key theoretical claim is that similar limitations also arise in human reasoning, suggesting that the architecture of LISA can provide computational explanations of properties of the human cognitive architecture. We report LISA's performance in simulating a wide range of empirical phenomena concerning human analogical access and mapping. The model treats both access and mapping as types of guided pattern classification, differing only in that mapping is augmented by a capacity to learn new correspondences. Extensions of the approach to account for analogical inference and schema induction are also discussed.