The implementation of partial least squares with artificial neural network architecture

The widely used multivariate analysis method, partial least squares (PLS) regression is mapped to the general multilayer architecture of artificial neural networks. This architecture can be viewed as a parallel implementation of PLS method in the weight matrix of input-to-hidden layer. The nature of the PLS approach is comparable to the well-known back-propagation (BP) method, which also utilizes the input-output pair for error correction. This novel concept provides a way to view the statistical meaning of the extracted feature in BP method. Apart from the traditional views of principal component, which results from the autocorrelation of input patterns, this is the first time a different statistical description of the resultant weight matrix been proposed.