A unified approach to laterally-connected neural NETS

The aim of this paper is to present a new unified approach to real- and complex-valued Principal Component Extractor with laterally-connected neural nets as the APEX (Kung-Diamantaras) and the cAPEX (Chen-Hou) based on an optimization theory specialized for such architectures. We firstly propose an optimization formulation of the problem and study how to recursively determine solutions by means of gradient-based algorithms. In this way we find a class of learning rules called ψ-cAPEX containing, as a special containing as a special case, a cAPEX-like one. Through simulations we finally compare the convergence speed and numerical precision at equilibrium of cAPEX and some members of ψ-cAPEX.