Bidirectional activation-based neural network learning algorithm

Abstract

We present a model of a bidirectional three-layer neural network with sigmoidal units, which can be trained to learn arbitrary mappings. We introduce a bidirectional activation-based learning algorithm (BAL), inspired by O'Reilly's supervised Generalized Recirculation (GeneRec) algorithm that has been designed as a biologically plausible alternative to standard error backpropagation. BAL shares several features with GeneRec, but differs from it by being completely bidirectional regarding the activation propagation and the weight updates. In pilot experiments, we test the learning properties of BAL using three artificial data sets with binary patterns of increasing complexity. © 2013 Springer-Verlag Berlin Heidelberg.