Development and Logistical Notes from MParT Meeting 03/10/2023

[dannys4]

• It goes without saying, but anyone who is interested is welcome to come to MParT meetings! Out of abundance of caution, I won't post the zoom link on a public website, but please reach out to me (contact me at my github handle@mit.edu) or other MParT maintainers for the zoom link.
• Going forward we'll try meeting twice a month, where the second friday of the month is "show and tell" on transport maps and the fourth friday is a "developer meeting" for more mundane code development.
• Phase 1 of MParT was getting the JOSS paper out the door. Phase 2 was implementing enough with distributions etc. that we could get rudimentary map training and ATM for more "industry-friendly" transport maps. Phase 3 now is creating a more "research-friendly" MParT, largely dependent on moving around code to help develop more flexible multivariate expansions. This will hopefully allow you to "easily" create some parameterization in a binding language and use it in a map.
• At a higher level, we need to make MParT less rigid w.r.t. implementing new parameterizations. For that, we need to bring back Tensor Product Parameterizations, to be able to specify the complexity in each dimension (particularly when the basis on the diagonal is different than off the diagonal).
• In particular, it would be nice to have integrated RBFs and wavelets as options for the basis for diagonal variables.
• I'm petitioning to rename MonotoneComponent as RectifiedComponent, which allows me to implement a SeparableComponent. Hopefully, at some point, we can then train separable components as seen in e.g. Appendix B of Max Ramgraber's Paper, which will hopefully be significantly quicker since we only need to explicitly solve for the diagonal elements, in a form that is accelerated by matrix multiplication and vectorized operations.