Ideas for libraries (so that people can use this stuff quickly)
- Implement a discrete Poisson distribution using RELAX/REBAR and/or the TRE.
- Reimplement the CUDA code for unbiased MCMC so that it’s PyTorch based, if that’s possible. That way we can let people define arbitrary likelihoods and still get coupled kernels. Maybe use CuPy as a glue language, though if the PyTorch story for external modules gets better then this shouldn’t be an issue.
Ideas for the blog:
- Use a serverless framework (AWS Lambda or Google Cloud Functions I guess) to implement comments on the blog, with KaTeX and Markdown support.
- Factor this out into a separate library that anyone can use
- Make it fast again:
- Rewrite my Math 221 final paper and add that section on algebraic statistics, and post it
- Write about High & Low