Keir Mierle (Google) came into town for a week of hacking. He is on vacation; it pleases me that this is his idea of vacation! We spent the day working on the objective—that is, the scalar function we are trying to optimize—for the tweak problem, the fitting of fine-scale astrometric distortions in images. In the end we came close to what Astrometry.net uses in its verify step, the step at which it decides whether a quad-inspired hypothesis about astrometric calibration is correct (or not). The new thing is that model complexity must be penalized: At equal performance we prefer a simpler model; we need good evidence to instantiate new model freedoms. If we can get consensus with Lang, we will hack tomorrow.
Wish I was there too.
ReplyDelete