Novel training algorithm for limited connected networks

It is argued that in many neural net learning algorithms (e.g. standard backward propagation) the algorithms themselves do not learn, so that they continue adjusting all weights at late stages in learning, and any new information requires a complete re-learning. Binary data are considered. A net is designed with elements modelled as binary units, even though an analog circuit may produce the desired response. The fan-in is limited but an architecture with many layers is assumed. The concern is with the development of a training algorithm for such a system. This algorithm overcomes some of the difficulties