Complex Cognitive Information Processing: A Computational Architecture (CONPOSIT) With a Connectionist Implementation.

摘要

We developed several novel representational and processing techniques for use in connectionist systems designed for high-level AI-like applications such as common-sense reasoning and natural language understanding. The techniques were used, for instance, in a connectionist system (Composit/SYLL) that implements Johnson-Laird's mental-model theory of human syllogistic reasoning. this theory was chosen as a case study for verifying the power of the techniques, because it was developed independently of the project, contains complex symbolic structures of various types, and requires complex sequences of operations. The resulting connectionist system is probably the most advanced, complex, and complete connectionist rule-based system in existence. It has a more complete scheme for binding rule-variables and for role binding than any other connectionist system. The representational techniques developed in the project were Relative-Position encoding (RPE) and Pattern-Similarity Association (PSA). Rpe allows structure to be encoded by the relative positioning of connectionist activation patterns within a subnetwork. PSA allows structure to be encoded by having different substructures include similar activation subpatterns. These techniques are similar to data structuring techniques used in computer memory (in particular, PSA is similar to associative addressing), but they had not previously been used in any non-trivial way in connectionism. The most distinctive processing technique developed in the project was the Temporal-Winner-Take-All (TWTA) method for selection in connectionist networks. This is more convenient and efficient for some purposes than conventional WTA methods.