A distributed connectionist representation for concept structures

We describe a representation for frame-like concept structures in a neural network called DUCS. Slot names and slot fillers are diffuse patterns of activation spread over a collection of units. Our choice of a distributed representation gives rise to certain useful properties not shared by conventional frame systems. One of these is the ability to encode fine semantic distinctions as subtle variations on the canonical pattern for a slot. DUCS typically maintains several concepts simultaneously in its concept memory; it can retrieve a concept given one or more slots as cues. We show how Hinton's notion of a "reduced description" can be used to make one concept fill a slot in another.