A Learning Algorithm for the Dynamics of CNN with Nonlinear Templates.pdfVIP

A Learning Algorithm for the Dynamics of CNN with Nonlinear Templates Part II: Continuous-Time Case F. Puffer, R. Tetzlaff and D. Wolf Institut f¨ur Angewandte Physik, Universit¨at Frankfurt Robert Mayer-Straße 2–4, 60054 Frankfurt a. M., Germany puffer@apx00.physik.uni-frankfurt.de ABSTRACT: A gradient-based learning algorithm for the dynamics of continuous-time CNN with nonlinear templates is presented. It is applied in order to find the parameters of CNN that model the dynamics of certain multidimensional nonlinear systems, which are characterized by partial differential equations (PDE). The efficiency of the algorithm is compared to that of a non gradient-based learning procedure we have previously developed. Results for model- ing two systems, whose dynamics are determined by nonlinear Klein-Gordon-equations, are discussed in detail. 1. Introduction During the past few years an increasing number of researchers have started to investigate not only the well defined steady states but also the dynamics of CNN [1, 4]. The possibility of applying learning algorithms in this context, i.e. methods which adjust the parameters of a CNN in order to obtain a specific dynamical behaviour of the cell states, has hardly been considered so far. The problem addressed in this paper is the modeling of autonomous spatio-temporal systems without an exact knowledge of the underlying PDE. Some values of the system’s state variables for a specific initial state are taken as the training pattern. A learning algorithm is applied in order to find the nonlinear templates of a CNN, which minimize the deviation of the cell states from t