The properties and implementation of the nonlinear vector space connectionist model

The nonlinear vector space expansion connectionist model is shown to have a number of advantages over currently applied networks for application to general connectionist problems. Firstly, it converges to its optimal solution far faster, by several orders of magnitude, than previously proposed networks, and this convergence time does not rise as quickly as a function of the size of the pattern vectors. Secondly, it is unimodal and consequently does not suffer from the problems of local minima such as needing to be reset, getting stuck or uncertainty about having converged. It may thus be left in adaptive mode while being `on the job´ as well as in training. Thirdly, the complete mathematical derivation of the network allows a much fuller understanding of its operation and allows direct application of adaptive filter knowledge. This allows easy parameter tuning and application of other adaptive filter update methods. It also allows the effects of approximation and implementation to be considered. The ability of the network to function well for random connection also allows the possibility of the implementation of the network by use of less exact technologies then electronic digital methods