Measuring the quality of MCMC output

Pitch

Markov chain Monte Carlo (MCMC) methods are now employed in nearly all fields of science. Along with a constant development of new methods, research on MCMC has led to steady improvements in our theoretical understanding of these techniques. However, most theoretical tools translate only vaguely into guidelines for practitioners, for example regarding the tuning of parameters and the choice of iterations to perform. How to address questions of “quality control” for existing MCMC algorithms has seemed to attract less attention than the development of new algorithms.

This event puts a spotlight on the question of assessing the quality of MCMC samples. It brings together 1) people who work on the development of methods addressing that question, and 2) people who encounter challenging sampling problems in applications. The goal is to bridge methodology and practice, so that methodologists can update their benchmarks and practitioners can update their toolbox. The workshop is intended to highlight advances, put people in touch, share experiences, identify challenges, and start collaborations among participants.

Dates and format

The workshop will be fully online. This website will collect relevant material, including posters submitted by participants and a link to a YouTube channel for pre-recorded video presentations. This content will be freely available to all and we hope that this website will remain a useful portal for material on this topic after the workshop.

You can submit a 20 minute video recording of a research presentation and/or a poster by indicating the title of these items in the registration form and emailing the organisers with the appropriate materials by the deadline of September 29th. For videos, these materials should be an mp4 file or a YouTube link, in addition to a PDF file for slides. Poster submissions should be PDF files.

The workshop event itself will be made of three live sessions, on different dates and times so as to accommodate different time zones. The sessions will be interactive, using gather.town, and available to registered participants only. The live sessions will include virtual poster sessions, panel discussions and informal chats among participants. Details will be provided here prior to the event . The registered participants will be invited to familiarize themselves with the available material available on this website beforehand, to make the most of the live sessions.

The live sessions will be on 11am-2pm UTC on Wednesday 6th October, 1pm-4pm UTC on Thursday 14th October and 3pm-6pm UTC on Friday 22nd October.

Confirmed participants

The following individuals are confirmed participants and advisors for the workshop.

Registration

Registration is free but compulsory, so as to limit the number of participants to an approximate maximum of 100.

We are primarily welcoming participants whose current research aligns with the topic of the workshop, for example on

  • MCMC in challenging applications, such as distributions on graphs, on trees, on high-dimensional discrete or continuous spaces,
  • asymptotic variance / effective sample size estimation, stopping criteria,
  • developing new tools to theoretically validate MCMC outputs, such as concentration inequalities,
  • assessing convergence empirically using coupling techniques, Stein's method, etc
  • and many more topics! Feel free to reach out if you are not sure whether your research aligns with the workshop's theme.

Please fill out the registration form here .

Logo Competition

We're looking for a logo for our "Measuring the Quality of MCMC Output" workshop. Please contact us (the organisers) if you have any creations that you'd like to put forward!

Organizers and Sponsors

The event is primarily organized by Leah South, Queensland University of Technology (QUT), and Pierre E. Jacob, Harvard University, with the support of the BayesComp section of ISBA. We can be contacted at l1.south (at) qut.edu.au and pierre.jacob.work (at) gmail.com.

Free registration is courtesy of the QUT Centre for Data Science.