Pattern completion with the random neural network using the RPROP learning algorithm
Catherine Hubert · 2002
A new model of neural networks called the random neural network (RNN) has been introduced by Gelenbe (1989). It provides many analytical properties and in particular the product form of its solution. The pattern completion operation may be performed by an associative single-layer RNN network. For the learning phase, the author retains the local adaptive learning algorithm RPROP which is much faster than pure gradient descent. Performances in learning and pattern completion have been evaluated considering geometrical patterns of various size. Though longer learning times are necessary with the RNN model, the latter globally outperforms the connectionist model introduced by Rumelhart (1986) and is much less sensible to pattern geometry.>