Chaos- b a se d Learning
Paul F. M. J. Verschure · 1991
It is demonstr ated th at t he chaotic prop erti es of neu ral networks (in t his case networks defined by th e generalized delt a procedur e) can be used to improve th eir learning perform ance. By adapt ively varying t he learnin g-rate parameter an annealing mecha nism can be intr odu ced t hat is founded in chaos. T he prop osed mecha nism, chaos-based learning, provides faster convergence t han standard back-propagation and also seems to provide a computat ionally less in tensive alternative to ot her back-propagation accelera ting techniques by using adapt ive step-size control.