A Bidirectional Method for Recognising and Attaining Optimal Nosiy Fits

摘要

Conventional supervised learning methods in feedforward networks are shown in this paper to have a significant degree of uncertainty in recognising the optimality of fit to noisy training data. It is argued here that this is due to the methods being unidirectional in their search and unable to eliminate alternatives ahead of their stopping points. The approach developed here, Neural BiDirectional Convergence (NBDC). instead converges towards a solution from dual pairs of directions. The pairs attempt to provide a greater degree of certainty through each dual member eliminating alternatives ahead of the other members' stopping point.