An Adaptive Cellular Network for Subspace Extraction

The work proposes a novel network structure for the least mean square error reconstruction (LMSER) principle to perform principal subspace analysis (PSA). The LMSER principle allows for an efficient parallel and robust implementation of PSA, where each individual processing cell contains a local adaptation algorithm. Instead of the classical feedforward network topology this work introduces a recursive topology. It is also shown that the fully connected two-layered network can be represented by a network of multiple locally connected processing layers. This locally coupled network closely resembles cellular nonlinear networks (CNN) and is very suitable for a VLSI (very-large-scale-integration) implementation.