Warped Gaussian Processes

by Edward Snelson, Carl Edward Rasmussen and Zoubin Ghahramani

This paper exploits the concept ``warped Gaussian process,'' referring to the observed data which after a warping function applied to is a GP. Well this won't make the problem difficult. Since we can compute the posterior GP as before and add a inverse warping to the posterior in order to get the posterior distribution of observed data. The problem is how it is possible we know the warping function a priori given the data? The author argue that this method is for preprocessing data (so the warping function is known). So if the data do not present any normality, it is possible to find a warping function so that the mapped data will be better modeled with GP.

On the whole, this paper is not very intriguing.