Bidirectional neural networks and class prototypes

Among the numerous existing neural network models, there are only a few bidirectional neural networks. None of these bidirectional models are based on multilayer perceptrons. These models also have limitations which have prevented their popularity. The authors have designed bidirectional neural networks (BDNNs) based on multilayer perceptrons trained by a generalised form of the error backpropagation algorithm. They can be trained as either associative memories or cluster centroid finders. This gives the network new abilities and enables the authors to design powerful data representation techniques which are a key factor in reducing network generalisation error. The authors demonstrate applications of this approach to extracting meaning from neural networks, and finding the centers of clusters. This work could be a step towards simulating via the behaviour of artificial neural networks clustering methods closer to that of the biological brain. BDNNs may also be used as a simulation tool for evaluating some major cognitive psychology theories such as prototypes and dual-code theory. The successful and promising results of applying BDNNs on two real world problems (including missing and noisy attributes) as well as some artificial data sets is reported