The author starts off in a bit of a mood and needs a good martini and Laplace approximation to fix it but is out of martinis. The post, though started in February 2023, got delayed due to visas and jobs distraction. It delves into the Laplace approximation, a simple Gaussian distribution that approximates target distribution mode. The post explores computation of Laplace approximation in JAX and compares it to MCMC, finding similar results. In speeding up the computation, the author applies various techniques, including factoring out linear transformations and observing diagonal Hessians for efficiency. The post concludes with possible automation of these techniques for easier implementation.
https://dansblog.netlify.app/posts/2024-05-08-laplace/laplace