This article introduces a new architecture and associated algorithms ideal
for implementing the dimensionality reduction of an ni-dimensional manifold
initially residing in an n-dimensional Euclidean space where n >> m. Motiv
ated by Whitney's embedding theorem, the network is capable of training the
identity mapping employing the idea of the graph of a function. In theory,
a reduction to a dimension d that retains the differential structure of th
e original data may be achieved for some d less than or equal to 2m + 1. To
implement this network, we propose the idea of a good-projection, which en
hances the generalization capabilities of the network, and an adaptive seca
nt basis algorithm to achieve it. The effect of noise on this procedure is
also considered. The approach is illustrated with several examples.