Scalable Log Determinants for Gaussian Process Kernel Learning