Solving Proportional Analogy Problems using Tensor Product Networks with Random Representations
William H. Wilson, Deborah J. Street, Graeme S. Halford · 2005
This paper describes the use of vectors with randomly generated components for representing concepts in a system for solving proportional analogy problems using a memory based on a tensor product network. Both dense random vectors (all or most components non-zero) and sparse random vectors (most components zero) are used. In both cases systems which are able to solve proportional analogy problems were successfully produced: the degree of success varied with the number of components in the vectors and/or the proportion of non-zero components. 1 Introduction The STAR (Structured Tensor Analogical Reasoning) model [4,5,6] solves analogical problems using a distributed connectionist approach. In this respect it contrasts with ACME [7] and SME [3]. The original versions of STAR relied on distributed representations for concepts which comprised an orthonormal set of vectors. The STAR model does not set out to be faithful model of cognitive processing at the neural level, but this restricti...