Bayes@Lund 2019

Tuesday 7 May 2019, 08:30 → 17:00 Europe/Stockholm

Palaestra, Lund University, Sweden

Thank you all who participated in Bayes@Lund 2019 and who made it such a successful and enjoyable event! :)

All best,
The Organizers
Rasmus Bååth, Alex Holmes, and Ullrika Sahlin

Below you will find the original webpage of Bayes@Lund 2019.

---

You are welcome to participate in the sixth edition of Bayes@Lund!

The purpose of this conference is to bring together researchers and professionals working with or interested in Bayesian methods. Bayes@Lund aims at being accessible to researchers with little experience of Bayesian methods while still being relevant to experienced practitioners. The focus is on how Bayesian methods are used in research and in the industry, what advantages Bayesian methods have over classical alternatives, and how the use and teaching of Bayesian methods can be encouraged. (see last year's conference for what to expect).

The conference will take place at Lund University, Sweden on the 7th of May 2019 starting at 9.00 and ending at 17.00. It will include contributed talks and invited presentations. Please register for the conference here.

The Program

The program is now finalized! For a list of all the speakers, and abstracts for all talks, do check out the book of abstracts:

                       

Slides and info

Some of the speakers have agreed on sharing slides and information regarding their presentations which you'll find here:

• Visualisation for refining and communicating Bayesian analyses -- Robert Grant
  • More info about Robert's book Data Visualization: charts, maps and interactive graphics: http://robertgrantstats.co.uk/dataviz-book
• Extending Bayes to Make Optimal Decisions -- Jonas Kristoffer Lindeløv
  • Slides: https://github.com/bayesat/bayesat.github.io/blob/master/lund2019/slides/jonas_lindelov_bayes_at_lund_2019.pdf
  • Notebook with code and explanation (highly reccomended!): https://lindeloev.github.io/utility-theory/
• Signal detection theory as bridge for Bayesian statistics and modelling -- Gerit Pfuhl
  • Slides: https://github.com/bayesat/bayesat.github.io/blob/master/lund2019/slides/gerit_pfuhl_bayes_at_lund_2019.pdf
• Rich-man's Monte Carlo: Uncertainty Analysis in Excel -- Dmytro Perepolkin
  • Code: https://github.com/dmi3kno/PMMC_BayesAtLund19 
  • Slides: http://bit.ly/PMMC_BayesAtLund19
• Bayesian vs. Frequentism for experimentalists -- Jakob Lavröd
  • Slides: https://github.com/bayesat/bayesat.github.io/raw/master/lund2019/slides/jakob_lavrod_bayes_at_lund_2019.pptx
• Bayesian Deep Learning Applications in Biomedicine -- Nikolay Oskolkov
  • Slides: https://github.com/bayesat/bayesat.github.io/blob/master/lund2019/slides/nikolay_oskolkov_bayes_at_lund_2019.ppt
• Bayesian inference of conformational ensembles from small-angle scattering data -- Wojciech Potrzebowski
  • Slides: https://github.com/bayesat/bayesat.github.io/blob/master/lund2019/slides/wojciech_potrzebowski_bayes_at_lund_2019.pptx
• Automatic Learning of Summary Statistics for Approximate Bayesian Computation Using Deep Learning -- Samuel Wiqvist
  • Slides: https://github.com/bayesat/bayesat.github.io/blob/master/lund2019/slides/samuel_wiqvist_bayes_at_lund_2019.pdf
• What cause successful learning in Bayesian methods? -- George Moroz
  • Slides: https://github.com/bayesat/bayesat.github.io/blob/master/lund2019/slides/george_moroz_19.05.07_Lund_Bayes.pdf

 

Invited Speaker: Maggie Lieu

Maggie is an astrophysics research fellow working at the European Space Agency in Madrid. Her main research involves modelling the mass distribution of clusters of galaxies to understand the nature of dark matter and dark energy in our Universe. Maggie's talk is Hierarchical models and their applications in astronomy; how hierarchical models can be a powerful tool for inference.

 

Invited Speaker: Robert Grant

Robert Grant is a medical statistician, turned freelance trainer, coach and writer in Bayesian models and data visualisation. His book Data Visualisation: charts maps and interactive graphics is published by CRC Press. His talk Visualisation for refining and communicating Bayesian analyses will review relevant general principles of effective visualisation, recent work on Bayesian workflow, and the role of interactive graphics.

Pre-conference Bayesian tutorial

Are you interested in Bayesian statistics and want to get up to speed? Then join the pre-conference Bayesian tutorial. This 3h tutorial will be given by Rasmus Bååth and will go through the fundamentals of Bayesian statistics using R. It will be based on the online course of the same name and requires no prior knowledge of Bayesian statistics but basic knowledge of the R programming language.

The tutorial is free of charge and takes place on the 6th of May 14.00 - 17.00 at Lund University, Sweden. Please register here.

bayes_at_lund_2019_program.pdf

Timetable.pdf

08:30 → 09:00 Bayes@Lund 2019 registration

09:00 → 09:05 Welcome

Speaker: Alexander Holmes (European Spallation Source ERIC)

09:05 → 09:55 Bayes@Lund 2019: Keynote

09:05 Hierarchical models and their applications in astronomy

Maggie is an astrophysics research fellow working at the European Space Agency in Madrid. Her main research involves modelling the mass distribution of clusters of galaxies to understand the nature of dark matter and dark energy in our Universe. Maggie will be talking about how hierarchical models can be a powerful tool for inference.

Speaker: Maggie Lieu (ESA)

09:55 → 10:00 Break

10:00 → 10:40 Bayes@Lund 2019: Introducing Bayes for learning

10:00 What cause successful learning in Bayesian methods?

I would like to share the comparison of three groups of students that I taught Bayesian methods this year:

• mixed group (psychology, biology...) with good background in R and frequentist statistics;
• linguistic group (3d year) with medium background in R and frequentist statistics;
• further education group in Computer Linguistic with beginner R and no statistical background.

I expected first and second groups to be more successful than the third one, but the shocking result was that the third and the first groups were more successful than the second one. I will try to explain the obtained result.

Speaker: George Moroz (National Research University Higher School of Economics)

10:20 Introducing Bayesian Stats through Signal Detection Theory

To avoid starting with a formula a detour via utilized Signal detection theory (uSDT) familiarizes psychology undergraduates with some of the basic concepts in Bayesian statistics. uSDT includes payoffs (utility functions), base rates, and varies similarities, illustrated on perceptual decision processes. Payoffs and base rates influences bias, assisting students in understanding models and priors in Bayesian statistics.

Speaker: Gerit Pfuhl (UiT)

10:40 → 11:00 Coffee break

11:00 → 12:00 Bayes@Lund 2019: Bayesian inference – a tale of high flexibility

11:00 Prior thoughts on mixed-membership models in linguistics

Bayesian mixed-membership models are popular in linguistics, as they explicitly model contact between languages (Reesink et al 2009, Syrjänen et al 2016). Most linguistic applications use the biological Structure program (Pritchard et al 2000) with default presets, fixing the concentration parameter of the population-level Dirichlet prior over allele frequency (treated as an analog for the language-level prior over features) at 1. We show, using a crosslinguistic typological database, that there are linguistically meaningful consequences for the choice of this hyperparameter (either fixed at different values, or inferred from the data) using a series of posterior predictive checks designed for mixed-membership models (Mimno et al 2015).

Speaker: Chundra Cathcart

Bayes@Lund2019.pdf

11:20 A Bayesian method to localize lost gamma sources

Radioactive sources can sometimes be lost or misplaced despite the existing rigorous safety rules. Lost sources must be found as soon as possible to avoid inflicting harm to the public. Regardless of the type of equipment used it is desirable to use as much information as possible from the measurements to draw conclusions about the activity and location of a detected source. Using Bayesian inference it is possible to obtain a probability distribution for the position and activity of an unshielded gamma source in one pass with a mobile gamma spectrometry vehicle.
The aim of this research was to investigate the feasibility of a Bayesian algorithm for mobile gamma spectrometry,test its accuracy in determining the location and activity.

Speaker: Antanas Bukartas (Lund University)

11:40 Spatio-Temporal Reconstructions of Global CO2-Fluxes using Gaussian Markov Random Fields

Atmosperic invere modelling is a method for constraining Earth surface fluxes (sinks and sources) of green house gases using measurements of athmosperic concentrations. The (linear) link between atmospheric concentration and fluxes are provided by an atmospheric transport model. Since the number of unknown surfaces fluxes is much larger than the number of observed atmospheric concentrations, the inverse problem is ill-conditioned. Requiring further assumption on the fluxes, leading to a Bayesian model. Hitorically, fluxes are discretized to a grid and modelled by a multivariate Gaussian distribution. Instead, we define the flux on a continuous spatial domain, with fluxes modelled as Gaussian Markov Random Fields, including both spatial and temporal dependence.

Speaker: Unn Dahlén

12:00 → 12:45 Lunch & Mingle

12:45 → 13:50 Bayes@Lund 2019: Keynote & contributed talk

12:45 Visualisation for refining and communicating Bayesian analyses

Introductory statistics classes teach the value of basic plots to diagnose problems with regression models. Like most of us, I found these very boring and spent a long time avoiding them. Now that I focus on complex Bayesian models, I find visualisation of variables, predictions and residuals more useful than ever. Analysts need to track down problems where the model is not fitting the data well, and also where the sampling algorithm is performing poorly, where the priors may be a bad choice, and where the likelihood may be a bad choice. In addition to this, Bayesian models fitted with simulation are fertile ground for visual communication of findings. In this talk, I will review relevant general principles of effective visualisation, recent work on Bayesian workflow, and the role that interactive graphics have to play.

Speaker: Robert Grant (BayesCamp)

13:30 Bayesian vs. Frequentism for experimentalists

A common rebuttal to Bayesian methods is that they are appropriate for large and complex problems (containing prior information and hidden variables), and most undergraduate teaching is often based upon frequentist methods. Starting from the context of the practical experimentalist, we explore the difference between the Bayesian and Frequentist methodology and highlights the advantages of teaching the Bayesian perspective already from the start. Many of the examples and the material come from the introductory course given for the students enrolled in the high school giftedness program “International Young Physicists’ Tournament”, an attempt at introducing Bayesian thinking into experimental practice.

Speaker: Jakob Lavröd (IYPT Sweden)

13:50 → 14:00 Break

14:00 → 15:00 Bayes@Lund 2019: Bayesian decisions

14:00 Extending Bayes to Make Optimal Decisions

Utility Theory allow you to make optimal decisions in the face of uncertainty. For example, what bidding price would maximize your earnings, taking the chance of failure into account? Utility Theory latches nicely onto Bayesian Inference. Once you have a posterior distribution, you need only a few more lines of code to apply a utility function (aka loss function) and identify the decision that optimizes said utility. This approach scales well to more complex models and decisions. We will use R and rstanarm/brms for Bayesian inference and hand-code the utility. An R notebook with worked examples will accompany the tutorial.

Speaker: Jonas Kristoffer Lindeløv (Department of Communication and Psychology, Aalborg University, Denmark)

14:20 Rich-man's Monte Carlo: Uncertainty Analysis in Excel

Adoption of Uncertainty Analysis in modern business environment is often challenging due to gaps in relevant skills and tooling (especially among decision makers). At the same time Excel is ubiquitous and can be used to build decision maker's intuition about uncertainties. This talk will introduce typical business problem faced by businesses and organizations on daily basis and showcase a fully transparent formula-only Excel model with no dependency on external libraries or macros, that can enlighten and inform about effect of uncertainties on business outcome(s).

In this talk we will discuss framing, expert knowledge elicitation, modeling, visualization and communication of results to facilitate discussion about uncertainties and ultimately aid the decision making.

Speaker: Dmytro Perepolkin (Lund University)

14:40 Bayesian Deep Learning Applications in Biomedicine

Next Generation Sequencing technologies gave rise to manifolds of Biomedical Big Data which is particularly manifested in the area of single cell transcriptomics where millions of cells are sequenced. Deep Learning (DL) is an ideal framework for analyzing large amounts of data and building predictive models for Clinical Diagnostics within the concept of Precision Medicine. Bayesian DL adds an important level of patient safety providing uncertainties to the biomedical predictions. Here, using single-cell transcriptomics data I demonstrate how Bayesian DL improves the accuracy of discovering novel cell sub-populations and dramatically outperforms classical methods when handling unknown cell sub-types.

Speaker: Nikolay Oskolkov (Lund University, Department of Biology)

15:00 → 15:30 Coffee and cake

15:30 → 16:50 Bayes@Lund 2019: Bayesian methods for the close to unseen

15:30 How to deal with a noisy zero – a simple Bayesian treatment for small angle neutron scattering

Neutron scattering measurements are an ideal case for Bayesian analysis – statistics are limited, measurement time is expensive and there is often relevant background information.

I will present an example of small angle neutron scattering from superconducting vortex lattices. Most of the signal detected is irrelevant, and contributes nothing but noise to the final result. A Bayesian treatment of the results, allows a huge enhancement of the signal to noise ratio, and treats missing data sensibly.

I will further argue that Bayesian analysis offers a huge amount to the scattering community and would benefit significantly from more systematic support from institutions such as the ESS.

A. T. Holmes, Phys. Rev. B 90, 024514

Speaker: Alexander Holmes (European Spallation Source ERIC)

15:50 Bayesian inference of conformational ensembles from small-angle scattering data

Small-angle scattering (SAS) uses x-ray or neutron scattering at small angles to investigate the structure of materials at the scale about 1-100nm. SAS is uniquely suited to study the conformational ensembles adopted by multidomain proteins. However, analysis is complicated by the limited information content in SAS data and care must be taken to avoid constructing overly complex ensemble models and fitting to noise in the experimental data. To address these challenges, we developed a method based on Bayesian statistics that infers conformational ensembles from a structural library generated by all-atom Monte Carlo simulations. The method involves a fast model selection based on variational Bayesian inference that maximizes the model evidence, followed by a complete Bayesian inference of population weights.

Speaker: Wojciech Potrzebowski (European Spallation Source, ERIC)

16:10 Bayesian determination of the effect of a deep eutectic solvent on the structure of lipid monolayers

We present a unique insight from Bayesian-driven modelling for a series of lipid monolayers at the air-deep eutectic solvent (DES) interface using reflectometry measurements.
A chemically-consistent modelling approach shows that the lipid monolayers at the air-DES interface are similar to those on water, while removing the need for water-specific constraints.
Furthermore, the use of Markov-chain Monte Carlo sampling enables the quantification of inverse uncertainties and parameter correlations in the modelling approach.
Finally, we discuss limitations present in the use of Bayesian methods for reflectometry analysis, and outline future work that will be conducted to overcome these.

Speaker: Andrew McCluskey (Diamond Light Source & University of Bath)

16:30 Automatic Learning of Summary Statistics for Approximate Bayesian Computation Using Deep Learning

Learning summary statistics is a fundamental problem in Approximate Bayesian Computation (ABC). The problem of learning summary statistics is in fact the main challenge when applying ABC in practice, and affects the resulting inference considerably. Deep learning methods have previously been used to learn summary statistics for ABC. Here we introduce a novel deep learning architecture (Partially Exchangeable Networks, PENs), with the purpose of automatically learning summary statistics for ABC. Our case studies show that our methodology outperforms other popular methods, resulting in more accurate ABC inference for Markovian data.

Joint work with Pierre-Alexandre Mattei, Umberto Picchini and Jes Frellsen.

Speaker: Samuel Wiqvist (Lund University)

16:50 → 17:00 Closing remarks

Speakers: Rasmus Bååth (Lund University; King Entertainment), Ullrika Sahlin (Lund University Centre of Environmental and Climate Research)