Hebbian learning and self-association in nonlinear neural networks

A self-organizing feature map, based on a Hebbian paradigm, is proposed as a universal adaptive memory. The learning paradigm, can be applied to arbitrary network topologies containing the standard sigmoidal nonlinearities at their nodes. The system generalizes the linear principal components by mapping the input space into a set of orthogonal nonlinear projections. Only localized learning rules are necessary for the adaptation. The size of the system is related to the desired accuracy and to the density of the examples.<>