A perspective on functional-link computing, dimension reduction and signal/image understanding

This paper provides a perspective for understanding the role of neural-net computing in signal and image understanding and also presents new results in the matter of dimension reduction for facilitating that task. Neural-net signal processing is viewed from the perspective of function estimation. In that practice, a critical step is the choice of an (approximate) basis for spanning the function space. The image of the input data in that function space is also an "internal representation" of the data. Once generated, the basis can be improved through regularization, SVD conditioning and so on. This paper describes a nonlinear variance-constrained transformation of the input data which can result in dimension reduction and has characteristics similar to principal component extraction or conditioning. An application of the method to optimization is also described.