GP regularization effect

luisroque · January 9, 2021, 3:04pm

I completely agree and would do that if it was computationally doable. The problem is that I am using GPs as priors in a hierarchical model and running it with datasets with 500+ time series with 300+ points each.

Running NUTs with a simple model like this and just one time series takes several hours because of the GP complexity.

Any suggestion on how to improve performance and ensure that the posterior space is free of these pathological shapes?

Topic		Replies	Views
Gaussian Process Regression level 1 inference: Re-producing Mauna Loa CO2 Example with PyMC3 Questions	10	2244	August 16, 2017
Multidimensional gaussian process Questions	55	10241	October 22, 2020
Multiple (uncertain) function observations of the same Gaussian process Questions	15	6568	October 16, 2017
MarginalSparse with low sigma as a workaround for sparse Latent GP Questions	6	787	April 25, 2018
Marginal likelihood implementation and measurement uncertainty Questions	12	910	November 2, 2019

GP regularization effect

Related topics