The Generalised Gaussian Process Convolution Model

This thesis formulates the Generalised Gaussian Process Convolution Model (GGPCM), which is a generalisation of the Gaussian Process Convolution Model presented by Tobar et al. [2015b]. The GGPCM provides a theoretical framework for nonparametric kernel models of multidimensional signals defined on multidimensional input spaces. We show that the GGPCM generalises and connects existing work; most notably, we derive a dual formulation of the cross-spectral mixture kernel presented by Ulrich et al. [2015]. Finally, we use the GGPCM to develop the Deep Kernel Model, which presents a new network structure for unsupervised learning.

Name dc:type.qualificationname

Master of Philosophy (MPhil)

Level dc:type.qualificationlevel

Masters

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2016

Author dc:creator

• Bruinsma, Wessel

Advisor dc:contributor.advisor

• Turner, Richard

• machine learning
• Gaussian process
• kernel
• nonparametric kernel
• multi-task learning

Licence

• https://apollo8-f-pro.lib.cam.ac.uk/bitstreams/800f5236-a9c6-4c71-8053-7eeb35eab557/download
• https://www.rioxx.net/licenses/all-rights-reserved/

Language dc:language

en

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.20389

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/273357