A principal component network for generalized eigen-decomposition
Dongxin Xu, José Carlos Príncipe, Hsiao‐Chun Wu · 2002
This paper presents a novel principal component network with simple online local learning rules to obtain generalized eigenvalues and their corresponding eigenvectors of input data. The network is composed of a set of forward linear projections and a set of linear lateral inhibition connections between them. The rules for both forward and lateral connections are not only mathematically rooted but also local which confers them biological plausibility. Stationary points of the rules and their stability are analyzed. Simulations are given to verify, the validity and effectiveness of the proposed method.