Active learning the weights of a RBF network

Abstract

We describe a principled strategy to sample functions optimally for function approximation tasks. The strategy works within a Bayesian framework and uses ideas from optimal experiment design to evaluate the potential utility of new data points. We consider an application of this general framework for active learning the weight coefficients of a Gaussian radial basis function (RBF) network. We also derive some sufficiency conditions on the learning problem for which there are analytical solution to the data sampling procedure.