Abstract

I will introduce Bézier Gaussian Processes which is a variational class of polynomial GPs. With their polynomial bases being the Bernstein polynomials, we can make them have similar behaviour to some of the most common stationary kernels. Most important I will introduce what we call the Bézier buttress, which makes the framework scale to both high number of observations and a high number of input features. The scaling in observations stems from the not having to compute matrix inverses and determinants. The scaling in features stems from only a small increase in compute by adding more "summarising points", in this framework named control points. Lastly, I will go into some future directions for Bézier GPs.

Notes

• Reference: Bézier Gaussian Processes for Tall and Wide Data

• Personal website can be found here .