andrewphoto

Andrew Gordon Wilson

I am a Postdoctoral Research Fellow in the Sailing Lab of the Machine Learning Department at Carnegie Mellon University.  I am interested in developing expressive and scalable machine learning models, particularly for kernel learning and deep learning.  I have expertise in probabilistic modelling, Gaussian processes, Bayesian nonparametrics, kernel methods, neural networks, scalable algorithms, and automatic machine learning.  My work has been applied to time series, image, and video extrapolation, geostatistics, gene expression and natural sound modelling, kernel discovery, Bayesian optimisation, econometrics, cognitive science, NMR spectroscopy, PET imaging, and general relativity. 

Outside of work, I am a classical pianist who particularly enjoys Glenn Gould's playing of Bach.

I can be reached at andrewgw@cs.cmu.edu, and on Twitter @andrewgwils.
Recent Highlights

NIPS 2015 Workshop on Large Scale Representation Learning

Code and tutorials using kernel methods for large scale representation learning

Deep Kernel Learning
Andrew Gordon Wilson*, Zhiting Hu*, Ruslan Salakhutdinov, and Eric P. Xing
To appear in Artificial Intelligence and Statistics (AISTATS), 2016
[arXiv, PDF, BibTeX]

Thoughts on Massively Scalable Gaussian Processes
Andrew Gordon Wilson, Christoph Dann, and Hannes Nickisch
arXiv pre-print, 2015
(See KISS-GP and Deep Kernel Learning for more empirical demonstrations).
[arXiv, PDF, BibTeX, Music]

The human kernel
Andrew Gordon Wilson, Christoph Dann, Christopher G. Lucas, and Eric P. Xing
Neural Information Processing Systems (NIPS), 2015
Spotlight
[arXiv, PDF, Supplement, BibTeX]

Kernel interpolation for scalable structured Gaussian processes (KISS-GP)
Andrew Gordon Wilson and Hannes Nickisch
International Conference on Machine Learning (ICML), 2015
Oral Presentation
[PDF, Supplement, arXiv, BibTeX, Theme Song, Video Lecture]

A video to watch for a succinct introduction to some of my research interests:
Video lecture on KISS-GP (Scalable Gaussian Processes), Lille, France, July 2015

Thesis

In January 2014 I completed my PhD dissertation, "Covariance Kernels for Fast Automatic Pattern Discovery and Extrapolation with Gaussian processes" (news story), in the Machine Learning Group at the University of Cambridge, where I am a member of Trinity College

Kernel methods, such as Gaussian processes, have great potential for developing intelligent systems, since the kernel flexibly and interpretably controls the generalisation properties of these methods.  The predictive performance of a kernel method is in general extremely sensitive to the choice of kernel.  However, it is standard practice to use a simple RBF (aka Gaussian, or squared exponential) kernel, which is limited to smoothing and interpolation.  This thesis argues for the importance of developing new kernels, introduces new kernels for automatic pattern extrapolation (with a view towards feature extraction, representation learning, and automatic kernel selection), and discusses how to best scale flexible kernel learning approaches, in order to extract rich structure from large  multidimensional datasets.

Covariance kernels for fast automatic pattern discovery and extrapolation with Gaussian processes
Andrew Gordon Wilson
PhD Thesis, January 2014.
[PDF, BibTeX]
Papers

Deep Kernel Learning
Andrew Gordon Wilson*, Zhiting Hu*, Ruslan Salakhutdinov, and Eric P. Xing
To appear in Artificial Intelligence and Statistics (AISTATS), 2016
[arXiv, PDF, BibTeX]

Thoughts on Massively Scalable Gaussian Processes
Andrew Gordon Wilson, Christoph Dann, and Hannes Nickisch
arXiv pre-print, 2015
(See KISS-GP and Deep Kernel Learning for more empirical demonstrations).
[arXiv, PDF, BibTeX, Music]

Scalable Gaussian Processes for Characterizing Multidimensional Change Surfaces

William Herlands, Andrew Gordon Wilson, Seth Flaxman, Daniel Neill, Wilbert van Panhuis, and Eric P. Xing
To appear in Artificial Intelligence and Statistics (AISTATS), 2016
[arXiv, PDF, BibTeX]
Coming soon!

Bayesian nonparametric kernel learning
Junier Oliva*, Avinava Dubey*, Andrew Gordon Wilson, Barnabas Poczos, Jeff Schneider, and Eric P. Xing 
To appear in Artificial Intelligence and Statistics (AISTATS), 2016
[PDF, BibTeX]
Coming soon!

The human kernel
Andrew Gordon Wilson, Christoph Dann, Christopher G. Lucas, and Eric P. Xing
Neural Information Processing Systems (NIPS), 2015
Spotlight
[arXiv, PDF, Supplement, BibTeX]

Kernel interpolation for scalable structured Gaussian processes (KISS-GP)
Andrew Gordon Wilson and Hannes Nickisch
International Conference on Machine Learning (ICML), 2015
Oral Presentation
[PDF, Supplement, arXiv, BibTeX, Theme Song, Video Lecture]

Fast kronecker inference in Gaussian processes with non-Gaussian likelihoods
Seth Flaxman, Andrew Gordon Wilson, Daniel Neill, Hannes Nickisch, and Alexander J. Smola
International Conference on Machine Learning (ICML), 2015
Oral Presentation
[PDF, Supplement, BibTeX, Code, Video Lecture]

À la carte - learning fast kernels
Zichao Yang, Alexander J. Smola, Le Song, and Andrew Gordon Wilson
Artificial Intelligence and Statistics (AISTATS), 2015
Oral Presentation
[PDF, BibTeX]

Fast kernel learning for multidimensional pattern extrapolation
Andrew Gordon Wilson*, Elad Gilboa*, Arye Nehorai, and John P. Cunningham
Advances in Neural Information Processing Systems (NIPS) 2014
[PDF, BibTeX, Code, Slides]

Variational inference for latent variable modelling of correlation structure
Mark van der Wilk, Andrew Gordon Wilson, Carl Edward Rasmussen
NIPS Workshop on Advances in Variational Inference, 2014
[PDF, BibTeX]

A Bayesian method to quantifying chemical composition using NMR: application to porous media systems
Yuting Wu, Daniel J. Holland, Mick D. Mantle, Andrew Gordon Wilson, Sebastian Nowozin, Andrew Blake, and Lynn F. Gladden
European Signal Processing Conference (EUSIPCO), 2014
[PDF]

Bayesian inference for NMR spectroscopy with applications to chemical quantification
Andrew Gordon Wilson, Yuting Wu, Daniel J. Holland, Sebastian Nowozin, Mick D. Mantle, Lynn F. Gladden, and Andrew Blake
In Submission
. February 14, 2014
[arXiv, PDF, BibTeX]

Covariance kernels for fast automatic pattern discovery and extrapolation with Gaussian processes
Andrew Gordon Wilson
PhD Thesis, January 2014
[PDF, BibTeX]

Student-t
processes as alternatives to Gaussian processes
Amar Shah, Andrew Gordon Wilson, and Zoubin Ghahramani
Artificial Intelligence and Statistics, 2014
[arXiv, PDF, Supplementary, BibTeX]

The change point kernel
Andrew Gordon Wilson
Technical Report (Note), University of Cambridge.
November 2013.
[PDF, BibTeX]

GPatt: Fast multidimensional pattern extrapolation with Gaussian processes
Andrew Gordon Wilson, Elad Gilboa, Arye Nehorai, and John P. Cunningham
October 21, 2013.   In Submission.
[arXiv, PDF, BibTeX, Resources and Tutorial]

Bayesian optimization using Student-t processes
Amar Shah, Andrew Gordon Wilson, and Zoubin Ghahramani
NIPS Workshop on Bayesian Optimisation, 2013.
[PDF, BibTeX]

Gaussian process kernels for pattern discovery and extrapolation
Andrew Gordon Wilson and Ryan Prescott Adams
International Conference on Machine Learning (ICML), 2013.
Oral Presentation
[arXiv, PDF, Correction, Supplementary, BibTeX, Slides, Resources and Tutorial, Video Lecture]

Modelling input varying correlations between multiple responses
Andrew Gordon Wilson and Zoubin Ghahramani
European Conference on Machine Learning (ECML), 2012
Nectar Track  for "significant machine learning results"
Oral Presentation

[PDF, BibTeX]

A process over all stationary covariance kernels
Andrew Gordon Wilson
Technical Report, University of Cambridge.
June 2012.
[PDF, BibTeX]

Gaussian process regression networks
Andrew Gordon Wilson, David A. Knowles, and Zoubin Ghahramani
International Conference on Machine Learning (ICML), 2012.
Oral Presentation
[PDF, BibTeX, Slides, Supplementary, Video Lecture]

Generalised Wishart processes
Andrew Gordon Wilson and Zoubin Ghahramani
Uncertainty in Artificial Intelligence (UAI), 2011.
Best Student Paper Award
[PDF, BibTeX]

Copula processes
Andrew Gordon Wilson and Zoubin Ghahramani
Advances in Neural Information Processing Systems (NIPS), 2010.
Spotlight

[PDF, BibTeX, Slides, Video Lecture]

Talks


Scalable Gaussian processes for scientific discovery
University of Cambridge, UK, July 2015

Kernel interpolation for scalable structured Gaussian processes (KISS-GP)
ICML, Lille, France, July 2015

Massively scalable Gaussian processes
New York University, NYC, June 2015

Kernel methods for large scale representation learning
Oxford University, Oxford, UK, November 2014
University College London, London, UK, November 2014
Neural Information Processing Systems (NIPS), Montreal, Canada, December 2014

Building kernel methods for large scale representation learning
Machine Learning Summer School (MLSS), Pittsburgh, USA, July 2014

Kernels for automatic pattern discovery and extrapolation
International Conference on Machine Learning (ICML), Atlanta, USA, June 2013

The automated Bayesian nonparametric statistician
Information Engineering Conference, University of Cambridge, Cambridge, UK, June 2013

Gaussian processes for pattern discovery
Research Seminar, Sheffield Translational Institute for Neuroscience,
University of Sheffield, Sheffield, UK, March 2013

Models of input dependent covariances
Xerox Research Seminar, Grenoble, France, November 2012

A machine learning approach to NMR spectroscopy
Microsoft Research Cambridge, UK, September 2012

Modelling input dependent correlations between multiple responses

ECML Nectar Track, Bristol, UK, September 2012
Information Engineering Conference, University of Cambridge, Cambridge, UK, June 2012


Gaussian process regression networks

ICML, Edinburgh, UK, June 2012

Bayesian nonparametric density estimation
Machine Learning Group, University of Cambridge, UK, May 2012

Bayesian nonparametric modelling of dependent covariances

Harvard University, April 2012
University of California, Berkeley, May 2012

Generalised Wishart processes
International Joint Conference for Artificial Intelligence, Award Winning Paper Track, Barcelona, July 2011
Uncertainty in Artificial Intelligence, Barcelona, July 2011

Copula and Wishart processes for modelling dependent uncertainty and dynamic correlations
(Poster) Bayesian Nonparametrics Workshop, Veracruz, Mexico, June 2011

Copula and Wishart processes for multivariate volatility
Rimini Bayesian Econometrics Workshop, Rimini, Italy, June 2011

Latent Gaussian process models
Latent Gaussian Models Workshop, Zurich, Switzerland, February 2011
ETH Zurich, Zurich, Switzerland, February 2011

Poisson processes
Machine Learning Group, University of Cambridge, UK, November 2010

Modelling changing uncertainty with copula processes
University College London, London, UK, October 2010

Time series
Machine Learning Group, University of Cambridge, UK, May 2010