Christian Robert’s talk on ABC for Model Choice and open floor on Bayesian computation

This is the Storify of my tweets from the afternoon sessions of Christian Robert’s visit to QUT. After he gave his talk on ABC and model choice, we had an open floor talk about Bayesian computation with presentations by Chris Drovandi, Tony Pettitt, Chris Strickland and Clair Alston. There was a fire evacuation there so we ended up skipping two speakers at the end (who would’ve given a five minute talk each).

  1. samclifford
    Getting ready for round 2 of #xian #amsi talk. Approximate Bayesian Computation.

    Mon, Aug 20 2012 20:59:14
  2. samclifford
    Latent variable models with large latent components may not be computationally feasible under MCMC. #xian #amsi ABC to the rescue

    Mon, Aug 20 2012 21:10:25
  3. samclifford
    Coalescent tree, for example, has an intractable likelihood. #xian #amsi

    Mon, Aug 20 2012 21:10:55
  4. samclifford
    Accept an amount of approximation where we look at closeness of data, y, and simulated data, z. #xian #amsi

    Mon, Aug 20 2012 21:12:27
  5. samclifford
    Closeness of y and z evaluated trough summary statistics which may not be sufficient statistics, e.g. quantiles #xian #amsi

    Mon, Aug 20 2012 21:14:21
  6. samclifford
    ABC more of a nonparametric approach to Bayesian inference than merely a computational trick. #xian #amsi

    Mon, Aug 20 2012 21:17:09
  7. samclifford
    Model choice is an important question because we ultimately have to choose a model amongst many which gives best performance #xian #amsi

    Mon, Aug 20 2012 21:18:09
  8. samclifford
    Prior on models in addition to priors on parameters. Want posterior inference on likelihood of models. #xian #amsi

    Mon, Aug 20 2012 21:19:11
  9. samclifford
    Frequency of acceptance of proposals from model m gives an indication of which is better model

    Mon, Aug 20 2012 21:20:55
  10. samclifford
    “@samclifford: Frequency of acceptance of proposals from model m gives an indication of which is better model” #xian #amsi

    Mon, Aug 20 2012 21:21:30
  11. samclifford
    Potts model usually contains too many terms to be numerically manageable. Can use ABC to infer best neighbourhood structure. #xian #amsi

    Mon, Aug 20 2012 21:23:04
  12. samclifford
    Sufficient statistics seem to be the big challenge with ABC. #xian #amsi

    Mon, Aug 20 2012 21:27:13
  13. samclifford
    Bayes Factor takes some effort to recover in ABC due to posterior inference being based on summary stats. #xian #amsi

    Mon, Aug 20 2012 21:32:38
  14. samclifford
    Someone forgot to turn the air
    con on. Hopefully we all wake up a bit now that it’s on. #xian #amsi

    Mon, Aug 20 2012 21:33:07
  15. samclifford
    “When is a Bayes factor based on an insufficient statistic consistent?” #xian #amsi

    Mon, Aug 20 2012 21:39:46
  16. samclifford
    I remember seeing Judith Rousseau show this example of ABC. N(theta1,1) vs Laplace(theta2, 1/sqrt(2)). #xian #amsi

    Mon, Aug 20 2012 21:44:38
  17. samclifford
    Choose summary stats that have a different mean under the competing models. Choose ancillary statistics. #xian #amsi

    Mon, Aug 20 2012 21:55:21
  18. samclifford
    Conclusions for #xian #amsi. Be careful when choosing summary stats. #xian #amsi http://twitpic.com/alz36p

    Mon, Aug 20 2012 22:00:48
  19. samclifford
    Question: if MCMC is feasible but going to take a long time would you use ABC? Answer: it’s a good tool for first analysis #xian #amsi

    Mon, Aug 20 2012 22:03:16
  20. samclifford
    Part 3 of the #xian #amsi session is a somewhat open floor for Bayesian computation

    Mon, Aug 20 2012 22:31:49
  21. samclifford
    Chris Drovandi – Bayesian experimental design. Algorithms and computational challenges. #xian #amsi

    Mon, Aug 20 2012 22:32:39
  22. samclifford
    Back upstairs now. Drovandi talking about Sequential Monte Carlo with particle methods. #xian #amsi

    Mon, Aug 20 2012 22:57:14
  23. samclifford
    Chris Strickland talking change detection in multivariate time series and space time data sets #amsi #xian

    Mon, Aug 20 2012 23:33:40
  24. samclifford
    Strickland finishes with a demo of change points in some land clearing data #amsi #xian

    Mon, Aug 20 2012 23:51:25
  25. samclifford
    Clair Alston discussing pyMCMC, a python module for Bayesian inference. #amsi #xian

    Mon, Aug 20 2012 23:52:11
  26. samclifford
    Question: will pyMCMC reach beyond academia? Answer: it’s just the standard models in python, we are writing this for animal sci #xian #amsi

    Tue, Aug 21 2012 00:07:17
  27. samclifford
    That’s the end of the #xian #amsi sessions. Quite a good day.

    Tue, Aug 21 2012 00:08:52
Advertisements

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s