At MPIA Galaxy Coffee, Daniel Lenz (JPL) spoke about foregrounds and component separation in CMB and LSS experiments. He emphasized (and I agree completely) that the dominant problem for next-generation ("Stage-4" in the unpleasant terminology of cosmologists) cosmology experiments—be they CMB, LSS, or line intensity mapping—is component separation or foreground inferences. He showed some nice results using generalized linear models of optical data for Milky-Way dust inferences. Afterwards I pitched him my ideas about latent variable models (all vapor ware right now).
Late in the day, Christina Eilers (MPIA) and I met to discuss why our project to fit for both labels and spectral model in a new version of The Cannon didn't work. I have various theories, most of which relate to some unholy mix of the curse of dimensionality (such that optimization of a model is a bad idea) and model wrongness (such that the model is trying to use the freedom it has inappropriately). But I am seriously confused. We worked through all the possible directions and realized that we need to re-group with our full team to decide what to do next. I assigned myself two things: The first is to look at marginalization of The Cannon internals (that is, what marginalizations might be analytic?). The second is to look at the machine-learning literature on the difference between optimizing a model for prediction accuracy as opposed to optimizing it for model accuracy (or likelihood).