A comparison of two eigen-networks
F. Palmieri, Jie Zhu · 2002
The authors compare two linear networks which project adaptively the input data points on their principal components. They rederive Sanger's algorithm as the result of a constrained optimization problem and compare it to the cascaded network suggested by P. Foldiak (1989). It is shown how the two approaches are asymptotically equivalent. The cascaded network does not require any backpropagation, seems to be faster, and perhaps could be more easily implemented in real hardware.>