The CHIR Algorithm: A Generalization for Multiple-Output and Multilayered Networks
Tal Grossman
Department of Electronics, Weizmann Institute of Science,
Rehovot 76100 Israel
Abstract
A new learning algorithm, learning by choice of internal representations (CHIR), was recently introduced. The basic version of this algoriths was developed for a two-layer, single-output, feed-forward network of binary neurons. This paper presents a generalized version of the CHIR algorithm that is capable of training multiple-output networks. A way to adapt the algorithm to multilayered feed-forward networks is also presented. We test the new version on two typical learning tasks: the combined parity--symmetry problem and the random problem (random associations). The dependence of the algorithm performance on the network size and on the learning parameters is studied.
