linearizedGP 1.0

daniel steinberg — Fri, 28 Nov 2014 07:02:54 -0000

This code contains python 2.7 and 3.x implementations of the Extended and Unscented Gaussian Processes (E/UGPs) for solving inverse problems. These are similar to regular GPs, but the latent function, f, can optionally have an extra nonlinear relationship to the observations, y, in the likelihood,

Normal GP likelihood: y ~ N(f, s^2 I_N) or for a single observation, n, y_n = f_n + e.
E/UGP likelihood: y ~ N(g(f), s^2 I_N) or for a single observation, n, y_n = g(f_n) + e.

Where g(.) is an arbitrary scalar function (maps R to R). The posterior Gaussian process parameters are learned using a variational objective with Gauss-Newton style linearization (of g) and mean finding.

More information can be found in our NIPS 2014 paper here:

http://papers.nips.cc/paper/5455-extended-and-unscented-gaussian-processes

mloss.org linearizedGP

linearizedGP 1.0