Layered Networks for Unsupervised Learning

Among several models of neural networks(1–4), layered structures are particularly appealing as they lead naturally to a hierarchical representation of the input sets, along with a reduced connectivity between individual cells. In Ref. 3 and 4, it was shown that such layered networks are able to memorize complicated input patterns, such as alphabetic characters, during unsupervised learning. On top of that, the filtering properties of the network can be continuously tuned from very sharp discrimination between similar patterns, to broad class aggregation when the selectivity of the cells is decreased. Unfortunately, it was also shown(4) that these properties are obtained with a reduced stability of the learning (the learning process does not converge for some values of the selectivity).