Learning the systematic transformation of holographic reduced representations

Abstract

Holographic Reduced Representation is a representational scheme which allows for the representation of variable-sized structures in a distributed manner. It has been shown that these compositional structures can be transformed holistically. However, in order to do so, the transformation vector was constructed by hand. In this paper we present two methods of learning the holistic transformation of Holographic Reduced Representations from examples. We show that the acquired knowledge can be generalised to structures containing unseen elements and to structures more complex than the training examples. These generalisations require a degree of systematicity which to our knowledge has not yet been achieved by other comparable methods.