https://arxiv.org/abs/2006.16649
Sparse Gaussian Processes with Spherical Harmonic Features
Vincent Dutordoir, Nicolas Durrande, James Hensman
“Gaussian processes (GPs) provide a flexible framework for modelling unknown functions: they are robust to overfitting, offer good predictive uncertainty estimates and allow us to incorporate prior assumptions into the model…. In arbitrary dimensions, our method leads to diagonal covariance matrices which makes it faster than VFF as we fully bypass the need to compute expensive matrix inverses. Compared to both sparse GPs and VFF, our approximation scheme suffers less from the curse of dimensionality.”