2018-11-21

information, interpretability, emulation, isocurvature

I started my day with a call to Adrian Price-Whelan (Princeton) to discuss my ideas around making decisions for observing using the EFDIG, which is my new acronym for expected future-discounted information gain. I want to make a real-time decision-making system based on this, but I don't want to spend tons of compute, for pragmatic reasons about comprehensibility and model-ability. I might be in trouble. In the middle of the call we had a funny idea about effectively marginalizing out short-period planets in a search for year-ish-period planets.

At lunch I discussed machine learning with Gabriella Contardo (Flatiron) and had a couple of duh moments: She pointed out that if you have a function or computation or simulation that can quickly go one way (from input to output) but cannot or cannot quickly go the other way (from output to input), then you have an ideal case for machine learning. Just generate data and train a model to go the other way. Duh! Machine learning to invert functions!

She pointed out that if you are trying to model a function that isn't one-to-one or many-to-one but rather many-to-many or one-to-many, in some sense, then vanilla machine-learning approaches won't be good: They are deterministic and single-valued, once trained. Vanilla methods, anyway. That was another duh for me. And yet I hadn't had these points emphasized so clearly and so sensibly.

Our conversation ventured into interpretability-land. I am all for generalizability—that's my jam—but recently I have been giving up on interpretability. Contardo isn't: Her feeling is that if she does her current project right (her project is to determine which light curves of stars in Kepler are in fact light curves of unresolved binaries), the features she obtains for light curves will be interpretable. Interesting!

Somewhere in the day I also had a nice chat with Stephen Feeney (Flatiron) about isocurvature perturbations in the initial conditions of the Universe and what they might do to Hubble Constant estimation. It looks like they might cause trouble! I wondered aloud about maximally adversarial isocurvature contributions.

No comments:

Post a Comment