Pattern completion with distributed representation

Pattern completion is used commonly with state vectors composed of fields. A neural net is first trained with a set of complete state vectors and is then given vectors with missing fields, which it fills by completing the vector. Distributed representation is characterized by the absence of fields, as every item of information in the vector is distributed over the entire vector. The paper reviews a conventional pattern-completion task and shows how the same information is represented in spatter code, which is a distributed binary code, and how patterns are completed when this code is used: an incomplete state vector looks like a noisy version of the complete vector with the noise distributed over the entire vector, the noise-free vector is found in a content-addressable memory, and the missing values(the “fields”) are extracted from the noise-free vector with functions called probing and clean-up. Distributed representation allows exceptional flexibility in designing data bases for use by neural nets