Summary Notebooks on Derivative Gaussian Processes with Tinygp

[edwarddramirez]

Hi everyone! I have created a repo of some code that builds on the Derivative Observations & Pytree Data tutorial and implements other variants of derivative GPs in tinygp. Hope it can be useful to someone!

GitHub Repo: https://github.com/edwarddramirez/derivative-tinygp

Results from Notebooks:

1D Function and its derivative sampled from a Gaussian Process with a sinusoidal mean function. The true derivative of the GP sample is estimated using finite differences.

A horizontal slice of a 2D function and its partial derivative with respect to x (denoted by black line in the rightmost plot). The true partial derivative is estimated using finite differences.

SVI fit of Poisson model to toy data using a log-GP prior on the rate function. The fit is improved when we add a penalty term that enforces GP samples to have zero derivative at x-values denoted by the dashed red lines.