Abstract
We split the rule extraction task in to a sub-symbolic and a symbolic phase and present a set of neural networks for filling the former. Under the two general commitments of: i) having a learning algorithm that is sensitive to feedback signals coming from the latter phase, and ii) extracting Boolean variables whose meaning is determined by the further symbolic processing, we introduce three unsupervised learning algorithms and show related numerical examples for a multilayer perceptron, recurrent neural networks, and a specially devised vector quantizer.