Simulations of Conposit, a supra-connectionist architecture for commonsense reasoning

A computation architecture called `Conposit´ is outlined. Composit manipulates very-short-term complex symbolic data structures of types that are useful in high-level cognitive tasks such as commonsense reasoning, planning, and natural language understanding. Conposit´s data structures are, essentially, temporary configurations of symbolic occurrences in a two-dimensional array of registers. Each register is implementable as a neural subnetwork whose activation pattern realizes the symbol occurrence. The data structures are manipulated by condition-action rules that are realizable as further neural subnetworks attached to the array. In simulations, Conposit performs symbolic processing of types previously found difficult for connectionist/neural networks. A version of Conposit, simulated on the massively parallel processor, embodying core aspects of P. Johnson-Laird´s mental model theory (1983) of human syllogistic reasoning is concentrated on. This version illustrates Conposit´s power and flexibility, which arises from two unusual data-structure encoding techniques: relative-position encoding and pattern-similarity association