3 Learning distributed representations of concepts

There have been many different proposals for how conceptual information may be represented in neural networks. These range from extreme localist theories in which each concept is represented by a single neural unit (Barlow 1972) to extreme distributed theories in which a concept corresponds to a pattern of activity over a large part of the cortex. These two extremes are the natural implementations of two different theories of semantics. In the structuralist approach, concepts are defined by their relationships to other concepts rather than by some internal essence. The natural expression of this approach in a neural net is to make each concept be a single unit with no internal structure and to use the connections between units to encode the relationships between concepts. In the componential approach each concept is simply a set offeatures and so a neural net can be made to implement a set of concepts by assigning a unit to each feature and setting the strengths of the connections between units so that each concept corresponds to a stable pattern of activity distributed over the whole network (Hopfield 1982; Kohonen 1977; Willshaw, Buneman, and Longuet-Higgins 1969). The network can then perform concept completion (i.e. retrieve the whole concept from a sufficient subset of its features). The problem with componential theories is that they have little to say about how concepts are used for structured reasoning. They are primarily concerned with the similarities between concepts or with pairwise associations. They provide no obvious way of representing articulated structures composed of a number of concepts playing different roles within the structure.