LEARNING AND RECOGNITION IN EXCITABLE CHEMICAL REACTOR NETWORKS

In further work on recognition and learning we present a reactor netwo rk consisting of four electrically coupled chemical reactors that are connected via Pt working electrodes in the fashion of a Hopfield netwo rk. Each reactor can assume either a periodic (P) or a nodal (N) state in the Belousov-Zhabotinsky (BZ) reaction. Two out of 16 (2(4)) dynam ical patterns are encoded by local coupling. The encoded patterns have been chosen such that their Hopfield matrix shows both positive and n egative coupling strengths. To successfully recognize all remaining (1 4) patterns, an averaging procedure for all amplitudes was introduced. Numerical simulations using the seven-variable Gyorgyi-Field model fo r the BZ reaction are in good agreement with the recognition experimen ts. We also simulate an iterative learning method to build up the syna ptic strengths from a random Hopfield matrix without any back-propagat ion of errors. Recognition occurs abruptly at a certain number of iter ations in the absence of any noise reminiscent of a phase transition. The inclusion of parameter noise is found to always broaden the recogn ition probability. Parameter noise enhances the recognition of pattern s in the early iteration stages, while the recognition probability is drastically reduced in the later stages of iterative learning.