A self-organizing neural network for multidimensional approximation

A neural network based on the combination of a feature map (memory) and linear filters is proposed as a generalized adaptive processor for multidimensional nonlinear mapping. The self-organizing part of the system provides a progressively finer embedding of the input space as more units are added to the network. The linear filters, which tap from the memory, provide the function approximations. Learning is achieved with simple rules of the Hebb´s type with no backpropagation needed. The author reports some preliminary results on two-dimensional patterns that show the potential of this approach