5.2 Pattern association algorithm

The training data are $\bf x_{i}^{}$ $\in$ IR^d with i = 1,..., n. In training, the kernel matrix $\bf\tilde{\bf K}$, its eigenvectors $\bf a^{l}_{}$ (section 2.4), and (if used) the reduced set {$\bf y_{i}^{}$,$\beta_{i}^{l}$} (appendix B.2) are computed. From these values, the potential E of a point $\bf z$ $\in$ IR^d can be obtained. In recall, a pattern is completed by a gradient descent in the potential field E($\bf z$).