Constraint Programming Technique for the Automated Combination of Networks of Functional Groupings of Spiking Neurons

Existing training techniques for spiking neural networks tend to be monolithic in nature and scale poorly to larger networks.This thesis aims to address this issue by providing an automated technique for combining certain types of functional groupings of spiking neurons into composite functional networks. The technique takes the form of an algorithm which uses constraint programming to ensure that four axioms hold true in the network; the axioms having been designed to ensure that signals arrive simultaneously to component groupings. A number of experiments were conducted in which the algorithm was used to combine component groupings into more complex composite networks; these experiments show the practical utility of the algorithm and reinforce by demonstration the correctness of the axioms. Networks designed for the experiments include memory registers which store binary values, binary counters, and a robot-control network for navigating a variant of the maze described by the T-Maze problem.