Script for Chainsail walkthrough video

[simeoncarstens]

To make Chainsail more accessible to potential users without needing them to actually try it out, we agreed on producing a walkthrough video. Let's discuss the script for that video in this issue.

Here's a first suggestion:

1. Quick introduction of the speaker, Tweag, and a pointer to the MCMC blog post series
2. Super brief discussion of multimodal probability distributions:
   • where do they occur in nature
   • show example (eog bimodal.png)
     • label switching, e.g. clustering, Bayesian neural networks
     • latent variable models, e.g., probabilistic PCA
     • Bayesian hierarchical models
     • ambiguous data (biomolecular structure determination)
   • what are current issues with sampling them
   • how does Chainsail help
     • implement Replica Exchange with automated parameter tuning
3. Demonstration of failing sampling on 2D Gaussian mixture example. This offers also the possibility for a shameless Nix plug
4. Super brief explanation of general Chainsail workflow:
   • write probability distribution
   • start job on website
5. Explain chainsail-helpers PDF interface
6. Zip & upload PDF:

   $ zip mixture.zip probability.py

7. Start job on Chainsail website, with short rundown of most important parameters
8. Wait for first plots to show up, explain dashboard
9. Fast-forward until job is finished
10. Download results, run postprocessing script (from chainsail-helpers)

    $ mkdir downloaded_results
    $ cd downloaded_results
    $ cp ~/Downloads/results.zip ./
    $ unzip results.zip
    $ concatenate-samples production_run chainsail_samples.npy

11. Demonstrate difference in sampling using comparison script

    $ cd ..
    $ python compare.py sc_samples.npy downloaded_results/chainsail_samples.npy

12. Finish up with call to action, support@chainsail.io email address, links to blog posts and pointer to @Etjean's soft k-means example (video / notebook)

I will produce a (unpolished) test video of this on the basis of which we can then discuss how well this script works and then likely amend it.

[Etjean]

2. Super brief discussion of multimodal probability distributions:
   • where do they occur in nature
     • mismatch between likelihhod and prior

Are you sure about this one? IIRC we said in the last Chainsail meeting that since it's the product likelihood*prior, and not the sum likelihood+prior, that should not produce a multimodal.

Other than that, I think the script is great and feels natural.

[simeoncarstens]

Walkthrough video has been ready since the August 2022 release and is available here.