Convolutional kernels for function approximation and generalization in learning from data

Computational units induced by convolutional kernels belong together with biologically inspired perceptrons to the most widespread types of units used in neurocomputing. Radial convolutional kernels with variable widths form RBF (radial-basis-function) networks and with fixed widths are used in SVM (support vector machine) algorithm. In the talk, we will show how properties of Fourier transforms of convolution kernels influence their suitability for function approximation and generalization in learning from data.