Jeremias Knoblauch

-----

NOTE:

THE SITE BELOW WILL NO LONGER BE UPDATED;

BUT YOU CAN FIND ME ON MY PERSONAL WEBSITE.

-----

News

I am now the Biometrika Fellow for 2021 (based at UCL)
I will begin my EPSRC fellowship and an assistant professorship at UCL in June 2022.

About My Time at Warwick

I was a postgraduate researcher in statistics working at the boundary of computer science as part of the Oxford-Warwick Statistics Programme (OxWaSP) together with Theodoros Damoulas and Chenlei Leng. I was also part of the Warwick Machine Learning Group, a visiting researcher at the Alan Turing Institute and a Facebook Fellow.

My research interests are focused on scalable spatio-temporal inference procedures for data generating mechanisms in high dimensions that are ill-behaved or difficult to describe. This encompasses modelling and doing inference for non-stationary data streams that may have changing behaviours across time as well as space. The algorithms and inferential procedures developed as part of this research will be used within the framework of the Clean Air London project at the Turing Institute to support London's Major's office in taking well-informed and data-driven policy decisions. For a more detailed look, here is my CV.

Research Interests

Modelling changepoints in a Bayesian way is elegant and computationally efficient (see Adams & MacKay, 2007; Fearnhead & Liu, 2005). I am working to extend this into a spatio-temporal context and to enable scalable robust inference on multivariate data.

Knoblauch, J. & Damoulas, T. (2018). Spatio-temporal Bayesian On-line Changepoint Detection with Model Selection, International Conference on Machine Learning (ICML) 2018.
(10-minute talk recording summarizing this paper here)
Knoblauch, J., Jewson, J. & Damoulas, T. (2018). Doubly Robust Bayesian Inference for Non-Stationary Streaming Data using β-Divergences, Neural Information Processing Systems (NeurIPS) 2018.
(3-minute video summarizing this paper here)

When dealing with on-line large scale data streams in a Bayesian way, scalable inference methods are key. In particular, apart from being slow at run time, sampling-based approaches also require a memory-consuming particle-based representation of the distributions in question. At the other end of the spectrum, standard variational inference methods are fast, but provide insufficient uncertainty quantification in noisy, ill-behaved data streams. To remedy this issue, my collaborators and I have recently developed a generalization of variational inference method that allows for customized uncertainty quantification that is as conservative and robust as the application in question requires it to be.

Knoblauch, J., Jewson, J. & Damoulas, T. (2019) Generalized Variational Inference: Three arguments for deriving new posteriors (arXiv:1904.02063)
Knoblauch, J. (2019) Frequentist Consistency of Generalized Variational Inference (arXiv:1912.04946)
Knoblauch, J. (2019) Robust Deep Gaussian Processes (arXiv:1904.02303)
Knoblauch, J. & Lara, V. (equal contributions) Robust Bayesian Inference for Discrete Outcomes with the Total Variation Distance (arXiv:2010.13456)

I also have various interests beyond these topics, including Gaussian Processes and Continual Learning.

Knoblauch, J., Husain, H. & Diethe, T. (2020) Optimal Continual Learning has Perfect Memory and is NP-hard (arXiv:2006.05188)
Maronas, J., Hamelijnck, O., Knoblauch, J., & Damoulas, T. Transforming Gaussian Processes With Normalizing Flows (arXiv:2011.01596)

Selected Presentations

Optimization-centric Generalizations of Bayesian Inference, Seminar Series Talk @ Imperial College Statistics Seminar (30/09/21)
Optimization-centric Generalizations of Bayesian Inference, Seminar Series Talk @ Nottingham University School of Mathematical Sciences (30/09/21)
Optimization-centric Generalizations of Bayesian Inference, Algorithms Seminar Series Talk @ Warwick University Statistics Department (19/03/21)
Optimization-centric Generalizations of Bayesian Inference, Seminar Series Talk @ Leeds University Statistics Department (13/11/20)
Optimal Continual Learning has Perfect Memory and is NP-hard, Seminar Series @ ContinualAI Seminar Series (06/11/20)
Optimization-centric Generalizations of Bayesian Inference, Invited Talk @ Facebook Research Core Data Science, London (05/11/20)
Optimization-centric Generalizations of Bayesian Inference, Seminar Series Talk @ Rough Path Interest Group, Alan Turing Institute (30/09/20)
Generalized Variational Inference, Seminar Series Talk @ CSML, Lancaster University (07/11/19)
Generalized Bayesian Inference Procedures for Robust Changepoint Detection in the Presence of Outliers, Seminar Series Talk @ DNSE, Lancaster University (07/11/19)
Generalized Variational Inference, Machine Learning Seminar Series @ Sheffield University (13/09/19)
Generalized Variational Inference, Tech Talk @ Google Brain, Mountain View (13/09/19)
Generalized Variational Inference with Applications to Time Series, Tech Talk @ Facebook Research, Menlo Park (13/09/19)
Generalized Variational Inference Workshop, One-day seminar @ Alan Turing Institute (19/07/2019)
Generalized Variational Inference, OxCSML Seminar @ University of Oxford (28/06/2019)
Generalized Variational Inference, New York University, Courant Institute of Mathematical Sciences (23/04/2019)
Generalized Variational Inference, Statistics Seminar @ Cornell University (22/04/2019)
Generalized Variational Inference, Data Science Seminar @ Columbia University (19/04/2019)
Doubly Robust Bayesian Inference for Non-Stationary Streaming Data using β-Divergences, Neural Information Processing Systems (NeurIPS), Montreal (02/12/2018-08/12/2018) [poster presentation]
Doubly Robust Bayesian Inference for Non-Stationary Streaming Data using β-Divergences, Facebook's PhD London Tech Talk (25/10/2018) [poster presentation + best poster award]
Spatio-temporal Bayesian On-line Changepoint Detection with Model Selection, International Conference on Machine Learning (ICML), Stockholm (10/07/2018-15/07/2018) [Talk and poster presentations]
Spatio-temporal Bayesian On-line Changepoint Detection with Model Selection, Statistics Seminar of the Glasgow School of Mathematics & Statistics, University of Glasgow (06/07/2018)
Bayesian Analysis for Non-Stationary Streaming Data, Seminar Series of the CDT in Data Science, University of Edinburgh (04/07/2018)
Bayesian On-line Changepoint Detection and Model Selection in high-dimensional data, Workshop on Computational Strategies for Large-Scale Statistical Data Analysis by the International Centre for Mathematical Sciences, Edinburgh (05/07/2018)

Reviewing Activities

NeurIPS 2019-2020, ICML 2020, AISTATS 2021, ICLR 2021, IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), Journal of the Royal Statistical Society, Series B (JRSS-B)

Software

I have written a substantial amount of software in Python accompanying my research, through which we have been nominated as Turing Reproducible Research Champions 2018 by the Alan Turing Institute.

Bayesian On-line Changepoint Detection with Model Selection (ICML '18)
Robust Bayesian On-line Changepoint Detection with Model Selection (NeurIPS '18)
Generalized Variational Inference (unpublished; code will come online upon publication)

Office: 3.14, Mathematical Sciences Building (MSB)

Contact: j.knoblauch@warwick.ac.uk

Personal webpage: https://jeremiasknoblauch.github.io/