George Wynne

I'm a post-doc at the Univeristy of Bristol in the School of Mathematics. Former PhD student at Imperial College

My main academic interests orbit around the relationship between data sets and models. This includes questions such as "If I have a data set and train a model, what is the smallest subset of data I could use to get a similar model?" and "If I have a data set and train a model, can I cook up a synthetic data set which would give me the same model?". I am interested in models which go beyond classical statistics and are used on more abstract spaces such as Hilbert spaces, a key example of this being models which are applied to functional data e.eg time series and surfaces. More specifically, I use tools in kernel-based discrepancies, coresets, Gaussian processes and statistical discrepancies. If you are interested in these areas please get in touch.

My email is: [[g.wynne]]](guess what symbol goes here)[[bristol]].[[ac]].[[uk]] without the braces

Bio

• Post-doc, University of Bristol 2022-Present
• PhD Student, Imperial College London 2018-2022 Supervised by Andrew Duncan and Mark Girolami.

News

• I have uploaded the file for the animations for my talks at LIKE23 Bayes Hilbert Spaces for Posterior Approximation and Kernel Mean Embeddings Meet Functional Data Analysis
• I gave two talks at LIKE23
• Our pre-print A Spectral Representation of Kernel Stein Discrepancy with Application to Goodness-of-Fit Tests for Measures on Infinite Dimensional Hilbert Spaces has been accepted to Bernoulli :)
• My new pre-print on the connection between Bayes Hilbert spaces and posterior approximation methods, in particular Bayesian coresets, is now on arxiv HERE
• I have submitted my thesis and started as a post-doc at the University of Bristol in November 2022.
• I gave a talk at the BIRS workshop "New Interfaces of Stochastic Analysis and Rough Paths"
• I will be giving a talk at Microsoft Research, New England in July 2022.
• Our pre-print "Variational Gaussian Processes: A Functional Analysis View" has been accepted to AISTATS 2022 :)
• I will be giving a seminar at Charles University Statistics department in December 2021.
• I will be giving a talk at the mini-symposium Incorporating structural information in kernel methods for prediction and design space exploration at SIAM UQ 2022 in Atlanta.
• Our pre-print "A Kernel Two-Sample Test For Functional Data" has been accepted to JMLR :)
• I'll be giving a talk at the Lifting Inference with Kernel Embeddings (LIKE22) workshop in January 2022 LINK
• I'm presenting a poster at Bernoulli World Congress in Probability and Statistics LINK
• I'll be giving a talk at Warwick SIAM Summer Conference 2021 on Friday 9th July LINK
• I've been awarded a top 10% reviewer at UAI 2021
• I will be giving a talk at IWFOS 2021 about applications of kernel mean embeddings to functional data LINK
• Our new pre-print "Statistical Depth Meets Machine Learning: Kernel Mean Embeddings and Depth in Functional Data Analysis" with Stanislav Nagy just hit arxiv https://arxiv.org/abs/2105.12778. In it we show that some common notions of statistical depth used in functional data analysis coincide with kernel mean embeddings and then leverage this to address some open problems surrounding statistical depth for functional data.
• Our pre-print "Convergence Guarantees for Gaussian Process Means with Misspecified Likelihoods and Smoothness" has been accepted to JMLR :)
• Our session "Probabilistic Numerical Integration" was accepted as a session at Bayesian Young Statisticians Meeting: Online 2020 LINK where I shall be talking about our paper on the impact of maximum likelihood parameter estimation on uncertainty quantification in Bayesian cubature this had to be cancelled too :(
• I'm giving a talk on (you guessed it) kernel based two-sample tests for functional data at the 1st edition of the school in Machine Learning and Dynamic Processes and Time Series Analysis at Scuola Normale Superiore, Pisa LINK
• A video on our pre-print A Kernel Two-Sample Test For Functional Data" was just uploaded onto the Fields Institute YouTube channel for the Second Symposium on Machine Learning and Dynamical Systems, video here
• Our preprint "A Kernel Two-Sample Test For Functional Data" was just put on arXiv https://arxiv.org/abs/2008.11095
• The paper "Maximum likelihood estimation and uncertainty quantification for Gaussian process approximation of deterministic functions" by Toni Karvonen, GW, Filip Tronarp, Chris J. Oates, Simo Särrkä has been accepted at SIAM/ASA Journal of Uncertainty Quantification

Papers

1. Title: Bayes Hilbert Spaces for Posterior Approximation
   Authors: GW*
   Type: Pre-print
   Link: https://arxiv.org/abs/2304.09053
2. Title: A Spectral Representation of Kernel Stein Discrepancy with Application to Goodness-of-Fit Tests for Measures on Infinite Dimensional Hilbert Spaces
   Authors: GW*, Mikolaj Kasprzak, Andrew B. Duncan
   Type: Bernoulli, to appear
   Link: https://arxiv.org/abs/2206.04552
3. Title: Variational Gaussian Processes: A Functional Analysis View
   Authors: GW*, Veit Wild*
   Type: AISTATS 2022
   Link: Proceedings
4. Title: Statistical Depth Meets Machine Learning: Kernel Mean Embeddings and Depth in Functional Data Analysis
   Authors: GW, Stanislav Nagy
   Type: Pre-print
   Link: https://arxiv.org/abs/2105.12778
5. Title: A Kernel Two-Sample Test For Functional Data
   Authors: GW, Andrew B. Duncan
   Type: JMLR 2022
   Link: Proceedings
6. Title: Maximum likelihood estimation and uncertainty quantification for Gaussian process approximation of deterministic functions
   Authors: Toni Karvonen, GW, Filip Tronarp, Chris J. Oates, Simo Särrkä
   Type: SIAM/ASA Journal on Uncertainty Quantification 8 (3), 926-958
   Link: https://arxiv.org/abs/2001.10965
7. Title: Convergence Guarantees for Gaussian Process Means with Misspecified Likelihoods and Smoothness
   Authors: GW, François-Xavier Briol, Mark Girolami
   Type: Journal of Machine Learning Research 22 (123), 1-40
   Link: https://arxiv.org/abs/2001.10818