How many are we expecting Supernovae detection are we expecting to detect ? What percentage of these are the neural network able to successfully detect ?
> astronomers use supernovae to measure distances, which is important for cosmologists to study, for instance, the expansion of the universe and dark energy.
AFAIK supernovae datasets are usually obtained as a survey, resulting in a statistical sample which can be used to compute cosmological observables. Here it seems that there is a new sample bias introduced by the neural network classifier. Can this bias be accurately quantified?
Automated detection of "transients" (anything that changes in the sky) will become increasingly important as large surveys vacuum up ever more data.
LSST is supposed to get going in 2022, imaging the entire visible sky every few days from the Atacama desert in Chile.[1] It's supposed to take a 3.2 Gigapixel image every 30 seconds or so, pretty much non-stop for 10 years. There will be tons of interesting transients hidden in that stream of data, and the key is flagging them as quickly as possible, so that other telescopes can do targeted follow-up. LSST itself is really the discovery machine, but it doesn't give much information about the transients it discovers. It just tells you that the transients are there, and that you should take a look.
Unless we're going to have a room full of people going through 120 3.2 Gigapixel images an hour, comparing them with reference images of the same part of the sky, we need algorithms to flag interesting transients. The algorithms should give us their guess of what sort of transient we're looking at (Type Ia supernova, Type IIb supernova, microlensing event, etc.) with "this is nothing like anything I've been trained on" possibly being the most interesting answer.
One of the big worries is that LSST will discover so many transients that there won't be enough resources to follow them all up. There probably aren't enough spectrographs installed worldwide to get a good spectrum of every transient. That makes it critical to automatically ranki transients by novelty or importance.
Most English words are imported from other languages, but we conjugate them according to English grammar, not the origin language. Except for pluralizing latin nouns.
14 comments
[ 2.4 ms ] story [ 39.9 ms ] threadAFAIK supernovae datasets are usually obtained as a survey, resulting in a statistical sample which can be used to compute cosmological observables. Here it seems that there is a new sample bias introduced by the neural network classifier. Can this bias be accurately quantified?
LSST is supposed to get going in 2022, imaging the entire visible sky every few days from the Atacama desert in Chile.[1] It's supposed to take a 3.2 Gigapixel image every 30 seconds or so, pretty much non-stop for 10 years. There will be tons of interesting transients hidden in that stream of data, and the key is flagging them as quickly as possible, so that other telescopes can do targeted follow-up. LSST itself is really the discovery machine, but it doesn't give much information about the transients it discovers. It just tells you that the transients are there, and that you should take a look.
Unless we're going to have a room full of people going through 120 3.2 Gigapixel images an hour, comparing them with reference images of the same part of the sky, we need algorithms to flag interesting transients. The algorithms should give us their guess of what sort of transient we're looking at (Type Ia supernova, Type IIb supernova, microlensing event, etc.) with "this is nothing like anything I've been trained on" possibly being the most interesting answer.
One of the big worries is that LSST will discover so many transients that there won't be enough resources to follow them all up. There probably aren't enough spectrographs installed worldwide to get a good spectrum of every transient. That makes it critical to automatically ranki transients by novelty or importance.
1. https://en.wikipedia.org/wiki/Large_Synoptic_Survey_Telescop...
I think we should phase out these pseudo latin plurals from English. They're just annoying and were a bad idea when introduced in the 1700s.
Most English words are imported from other languages, but we conjugate them according to English grammar, not the origin language. Except for pluralizing latin nouns.
... I'll let myself out, sorry