Additive neural networks and periodic patterns

In this contribution we discuss weight selection which allows additive neur al networks to represent certain periodic patterns. Given a periodic set of vectors V-l whose components are v(i)(l) = +/- 1 we measure correlation be tween i-th and j-th components of V-l in time l. We show that in the additi ve neural net with weights chosen based on this correlation, almost all tra jectories converge to a periodic orbit, which consecutively visit orthants, determined by the vectors V-l. We also construct two weights selection pro cesses, one discrete in time and one continuous in time, which construct th e desired weights dynamically. (C) 1999 Elsevier Science Ltd. All rights re served.