A novel gradient-based system for processing sequential time-varying inputs and outputs is described. With the method it is possible to train a system with time-varying inputs and outputs to use its dynamic links for temporarily binding variable contents to variable names as long as it is necessary for solving a particular task. Various learning methods for nonstationary environments are derived. Two experiments with unknown time delays illustrate the approach. A by-product of this work is the demonstration that a system consisting of two feedforward networks can solve tasks that only dynamic recurrent networks were supposed to solve.
|Title of host publication
|1991 IEEE International Joint Conference on Neural Networks - IJCNN '91
|Publ by IEEEPiscataway
|Number of pages
|Published - Jan 1 1991