Citation

BibTex format

@inproceedings{Calandra:2016:10.1109/IJCNN.2016.7727626,
author = {Calandra, R and Peters, J and Rasmussen, CE and Deisenroth, MP},
doi = {10.1109/IJCNN.2016.7727626},
publisher = {IEEE},
title = {Manifold Gaussian Processes for Regression},
url = {http://dx.doi.org/10.1109/IJCNN.2016.7727626},
year = {2016}
}

RIS format (EndNote, RefMan)

TY  - CPAPER
AB - Off-the-shelf Gaussian Process (GP) covariancefunctions encode smoothness assumptions on the structureof the function to be modeled. To model complex and nondifferentiablefunctions, these smoothness assumptions are oftentoo restrictive. One way to alleviate this limitation is to finda different representation of the data by introducing a featurespace. This feature space is often learned in an unsupervisedway, which might lead to data representations that are notuseful for the overall regression task. In this paper, we proposeManifold Gaussian Processes, a novel supervised method thatjointly learns a transformation of the data into a featurespace and a GP regression from the feature space to observedspace. The Manifold GP is a full GP and allows to learn datarepresentations, which are useful for the overall regressiontask. As a proof-of-concept, we evaluate our approach oncomplex non-smooth functions where standard GPs performpoorly, such as step functions and robotics tasks with contacts.
AU - Calandra,R
AU - Peters,J
AU - Rasmussen,CE
AU - Deisenroth,MP
DO - 10.1109/IJCNN.2016.7727626
PB - IEEE
PY - 2016///
SN - 2161-4407
TI - Manifold Gaussian Processes for Regression
UR - http://dx.doi.org/10.1109/IJCNN.2016.7727626
UR - http://hdl.handle.net/10044/1/30936
ER -
Email us: contact-ml@imperial.ac.uk