Learning Gradients with Gaussian Processes

Abstract

The problems of variable selection and inference of statistical dependence have been addressed by modeling in the gradients learning framework based on the representer theorem. In this paper, we propose a new gradients learning algorithm in the Bayesian framework, called Gaussian Processes Gradient Learning (GPGL) model, which can achieve higher accuracy while returning the credible intervals of the estimated gradients that existing methods cannot provide. The simulation examples are used to verify the proposed algorithm, and its advantages can be seen from the experimental results.