Kevin Murphy

@lawmurray@fosstodon.org I agree that paperspace is easy to use and GCP is hard.

Dynamax subsumes existing libraries, such as hmmlearn, pykalman, filterpy, etc. We are also working on various extensions to this library, such as for time-series forecasting, which we will release in the future. Big thanks to my colleagues @scottlinderman@sigmoid.social, @petergchang@sigmoid.social, @gileshd@sigmoid.social, karalleyna, mxinglongli, et al. and support from #GSOC.

For state estimation, we support various efficient, deterministic methods, such as forwards-backwards, Kalman filtering, sigma point filtering, etc. For parameter estimation, we support MLE via EM or SGD (using optax), and Bayes via Gibbs or (collapsed) HMC (using blackjax).

I am delighted to announce the release of “dynamax”, an open source library for dynamic state space models in JAX. It supports inference and learning in HMMs, Linear Gaussian SSMs, as well as non-linear and non-Gaussian SSMs. See https://github.com/probml/dynamax for the code.

It seems that the etiquette for this site is for your first post to be a self introduction. Oops. Anyway, for anyone who wants to know if they should follow me, I'll say: I don't post much, but when I do, it's usually about Bayesian machine learning, or my ML textbooks (https://probml.github.io/pml-book/), or my JAX software (stay tuned for a major update on that front). So if that's your jam, let's chat!

I just joined, after hearing buzz about this on twitter. Could Elon's acquisition of twitter be the "JMLR moment" for ML social media? (JMLR was created in response to the extortionate prices that Springer used to charge for "Machine Learning Journal"; now JMLR is arguably the top journal in the ML field, and it is entirely community led.)