It serves as a nice reminder to me that some of the most elegant solutions in machine learning and statistics aren't always the newest deep learning architectures. Sometimes, good old Bayesian inference with smart sampling strategies can create something really cool!

This visualises how modern statistical inference actually works. RJMCMC explores spaces where we don't even know how many parameters we need, perfect for art, where we don't know in advance how many strokes capture an image.

- Proposals are accepted/rejected based on how well they reconstruct your image
- The final painting emerges from thousands of probabilistic decisions
- It's not deterministic, run it twice -> get two different artworks

Upload an image, and the algorithms "paints", one probabilistic brushstroke at a time. Each brushstroke represents a "birth" or "death" jump in the model space:
- The algorithm proposes adding new strokes or removing existing ones

Haha I actually switched to Claude this week so terse bluesky posts incoming.....

AHH cool! I'll have a play with this, doesn't seem active currently tho

Cheers Sam! I've not seen this have you got a link? They are fitting ODIN models with monty these days right?

Yeah it's actually great, converting the c++ to JavaScript is actually not too bad with a little help from AI !

Watch RJMC explore different model dimensions in real-time, use sample data or upload your own CSV.
No installation, just open and experiment. Great for teaching/learning Bayesian model selection!

Package/vignette: dchodge.github.io/rjmc/article...
#statistics #bayesian #MCMC #datascience

Sure, wanna drop me an email to sort out deets?

"2.1. Overview of inference framework" in the methods gives an overview. But basically if you infer an infection you also need to infer an infection time (an extra parameter), no infection then infection time isn't in the framework anymore. Hence need to jump between different dimensions

Thanks to everyone who worked on this: @jameshay.bsky.social, Sheikh Jarju, Dawda Jobe, Rhys Wenlock, @adamjkucharski.bsky.social, and @thushan-desilva.bsky.social!

It uses reversible jump MCMC to infer missed infections, to help understanding I made a little widget to show you how the fitting process works for simulated data: seroanalytics.org/serojump-widget

✨ What we did:
- Made a Bayesian model to infer who was infected, when, and how their antibody levels changed
- Validated on both simulations and real-world SARS-CoV-2 data from The Gambia.
- Showed that serojump detects more infections (including sub-threshold ones) and provides richer insights

Thanks to every who worked on this! @jameshay.bsky.social, Sheikh Jarju, Dawda Jobe, Rhys Wnelock, @adamjkucharski.bsky.social and @thushan-desilva.bsky.social

serojump was designed to be a flexible and pathogen-agnostic solution that can be applied to a wide range of pathogens.

Heres an interactive widget to help with understanding of them reversible jump mcmc methods: lnkd.in/eWGJ39PG

What we did:
- Made a Bayesian model to infer who was infected, when, and how antibody levels changed over time.
- Validated on both simulations and real-world SARS-CoV-2 data from The Gambia.
- Showed that serojump detects more infections (including sub-threshold ones) and provides richer insights

Key features:
- WebAssembly-powered performance (10-50x faster than JS)
- Adaptive MCMC for Bayesian inference
- Vaccine intervention analysis with waning immunity
- Real-time convergence diagnostics
- Export data and plots for further analysis