A non-competitive model for unsupervised learning

The learning algorithm presented is noncompetitive, it is related to the principal component analysis rather than to cluster analysis. It is based on `backward inhibition´, i.e., the inhibition of features already discovered in the input, which makes finding further, more subtle features possible. It is shown that the backward-inhibition algorithm is superior to the competitive feature discovery algorithm in feature independence and controllable grain. Moreover, the representation in the feature layer is distributed, and the features define an implicit `classification hierarchy´