"Efficiently sampling functions from Gaussian process posteriors"


I came across this new method of approximate sampling from Gaussian Process posteriors:

There are also some notes in the twitter thread I found it in:

My question is (I haven’t yet dug that deep into the paper itself) - would this be interesting enough to add to PyMC3?


Funny enough I was just discussing this with @michaelosthege. He had work on similar topic (particularly the Fourier base approx)

I just open sourced an implementation of Random Fourier Feature approximations for sampling continuous functions from GPs (https://arxiv.org/abs/1511.05467).
I used numba to make it faster and wrote 1D and 2D examples with PyMC3: https://github.com/michaelosthege/pyrffI intend to add a Thompson sampling example notebook in the near future too.