Daily Note (7th August 2026 & weekend)
Lots of fun little things scattered amongst the tedium of admin and forms. Building a desktop with over 256 gigs of ram is surprisingly expensive (even if i went into it expecting high costs, what are people doing!?).
Restarted work on the all catalog project, specifically looking at crossmatching objects across the different catalogs. need to think a bit more about differing astrometry, but for now the main issue is how to do it in a way that fits in memory (using the pre-built routines in LSDB HATS for now) which works for small sections as test cases, but might crash even for single healpix pixels(?) with ~81 catalogs unless we do something smart.
Had a session 0 for the project with MJ, where we talked about possible directions we could go in - extending the J-HIVE dataset, creating a catalog of weird/small populations, training a diffusion model to do latent-space to image generation, or doing value-added label propagation (e.g. for spec-z across the dataset). for the dataset i'm currently thinking of SDSS/Legacy Survey/Regalade with labels from MaNGA. There's so many directions this project could go in, so it would be best to start with a literature survey and then decide how to go from there.
The discussion also brought up a bit of hermeneutics1 in how we approach a project, whether we start with a well-motivated problem and find the simplest, most appropriate method that can solve it, or whether we start with a method (in this case, our framework to ingest galaxy data and create a latent space) and find an appropriate problem to apply it to (one where a simpler approach would not be able to reach a solution due to scale or complexity). As physical scientists, we often think of the former as the valid way to approach a problem, because the issues with the latter are often more obvious (the method might not capture all the subtleties of the problem, taking a hammer-to-nail approach can sometimes lead to causal inferences only compatible with the hammer's interpretation of the phenomenon, etc.), but often as method developers (especially in AI/ML) we seek current open problems that might be a good fit to the method we're developing. While not 'traditional' in the sense we often think of, this kind of training helps students look at the 'shape' of a problem2 in a way that is otherwise difficult to teach. As a corollary, it can also lead to interdisciplinary applications, bringing over techniques from problems that have similar shapes. This is probably the subject of a longer essay, but it is interesting to think about.
Back to filling out paperwork and looking for a desktop.
Highlights: Ms. Coach had a fun show at the Rat on Sunday, opening for a show produced by the Sweet Treat team. Npr tells me that the two-prov set at the end was particularly fun, with a night-at-the-museum-esque interlude.
Paul Ricoeur’s Hermeneutics between Epistemology and Ontology seems like an interesting related read.↩