Nonmonotonic reasoning is a core problem in AI. An example of nonmonotonic reasoning is the type of default reasoning which occurs with inheritance structures which allow exceptions. This paper describes a connectionist model of a hierarchical inheritance structure with exceptions. Existing symbolic and related connectionist research are described, and their limitations summarized. The requirements for an adaptable connectionist model are laid out, and a representational architecture is constructed. The architecture requires relations (i.e. links between objects) to be bi-directional or directional, where the former is meant to capture those relations for which it is useful to have the inverse relation (e.g. `isa', `part-of'). The general assumption is that inferential distances are best captured by relying on representational similarities in the semantic features of tokens and types. Both the encoding mechanism and the decoding mechanism (for checking the uniqueness of the distributed representations) are described in detail. The representational architecture is implemented in recursive autoassociative memory. The model is successful, and future adaptation or handling multiple inheritance with exceptions is briefly explored.
HS-IDA-TR-92-003. Annotation: Published in Selected papers from the first Swedish Conference on Connectionism -- 1992: Connectionism in a broad perspective, Ellis Horwood.