Symbolic Processing Techniques in Connectionist Networks and Their Application to High-Level Cognitive Tasks

Abstract

Symbolic and distributed connectionist systems live in distinct processing niches. Symbolic systems supply variables, bindings, logical rules, constituent structure, tokens (versus types), hierarchies and inheritance. In contrast, distributed connectionist systems supply statistically based associations, reconstructive memories, graceful error degradation, automatic category/prototype formation and generalization to novel instances. High-level cognitive tasks, such as language comprehension, require symbolic processing capabilities; however, the brain gains its robustness from its distributed connectionist nature. Clearly, a synthesis is desirable and this paper surveys a range of techniques, being explored at the UCLA Artificial Intelligence Laboratory, for giving symbolic capabilities to connectionist systems. These techniques include: 1) use of signature activation to represent bindings in localist, spreading activation networks, 2) parallel distributed semantic networks (PDS) for integrating distributed connectionist networks with the structure of semantic networks, 3) symbol recirculation methods for automatically forming distributed representations of symbols, and 4) tensor manipulation networks for binding distributed symbols.