Data processing and feature screening in function approximation: An application to neural networks (Q597242)

The paper deals with the scattered data approximation using the principal component analysis (PCA) in application to neural networks described by the equation \[ \text{ in}_i (t)=\sum_{j=1}^p w_{ij}x_j(t)+\text{ bias}_i, \] where \(x_j\) are the components of the training vector and \(\text{ in}_i (t)\) is the input value of unit \(i\) at time \(t\). The variables derived by PCA are ordered by the size of their variation so that some variables with smaller variations are replaced with their mean values. For the errors \(\Delta x_j(t)=x_j(t)-x_j(t-1)\) one has \[ \text{ in}_i (t)=\text{ in}_i (t-1)+\sum_j w_{ij}\Delta x_j(t). \] For the variables replaced by their means \(\Delta x_j(t)=0\), which can be used to achieve a considerable reduction in computation. In conclusion, some performance data are given for completing methods on a real world application.