Gaussian Process Learning: A Divide-and-Conquer Approach

The Gaussian Process (GP) model is used widely in many hard machine learning tasks. In practice, it faces the challenge from scalability concerns. In this manuscript, we proposed a domain decomposition method in GP learning. It is shown that the GP model itself has the inherent capability of being trained through divide-and-conquer. Given a large GP learning problem, it can be divided into smaller problems. By solving the smaller problems and merging the solutions, it is guaranteed to reach the solution to the original problem. We further verified the efficiency and the effectiveness of the algorithm through experiments.