A geometrical approach to the design of an efficient neural-network supervised learning scheme

A simple, one-step, supervised learning scheme is studied from the geometrical point of view in N-dimensional space. The theories are simple, and the implementation with conventional electronic circuits is feasible. In this scheme, learning new input-output mappings does not destroy old mappings already learned. Some mappings cannot be learned at all and the legality of a given mapping to be learned must be checked. Learning digital input-output mappings allows the machine to do analogue pattern classifications. The learning capacity of this scheme is much higher than that of conventional learning schemes