I'm convinced that the reason for all the Chaos these days is because Eris was snubbed. The rules for what a planet is were changed purely to keep her out of the club. It seems even more egregious than the snub that led to the Trojan War.
As I understand the terminology, exoplanets are planets, so the headline is even technically correct. It's like if a headline said something about rectangles and you corrected it to squares.
To be clear, there have been prior studies which have discovered exoplanets with machine learning. One example (cited in this paper) was from Chris Shallue at Google and Andrew Vanderberg at CfA, who discovered the first eight planet system in 2017 by training convolutional neural networks on light curve data. [1]
The authors explain their contribution to be:
> To date these [previous studies] have all focused on identifying [false positives] or ranking candidates within a survey. We build on past work by focusing on separating true planets from [false positives], rather than just planetary candidates, and in doing so probabilistically to allow planet validation.
To translate, these previous studies focused on identifying exoplanet candidates. In the early days of exoplanet research, these candidates were then followed up with more detailed observations to confirm that the candidate was a true planet.
With the advent of Kepler, there are now way too many candidates to get follow-up observations on all of them. So now exoplanets are "validated" by calculating the probability of a false positive. I am not in this field, but it seems that the most common algorithm to do this is called VESPA [2]. The contribution these authors make is to use a machine learning system for the validation process, i.e., calculating the probability that the candidate is a false positive.
>there are now way too many candidates to get follow-up observations on all of them. So now exoplanets are "validated" by calculating the probability of a false positive
I don't math good, but this sounds like a "dial that goes to 11".
If you have a model to identify false positives, why not incorporate it into the original model to eliminate those positives? Or, if you really need observations, how can tweaking the model eliminate that?
It's not my field so I don't completely understand it, but I think it helps to have one method to propose candidates, and an independent technique to validate candidates. (Though I would welcome correction from someone who works on exoplanets.)
11 comments
[ 2.8 ms ] story [ 40.4 ms ] threadThe authors explain their contribution to be:
> To date these [previous studies] have all focused on identifying [false positives] or ranking candidates within a survey. We build on past work by focusing on separating true planets from [false positives], rather than just planetary candidates, and in doing so probabilistically to allow planet validation.
To translate, these previous studies focused on identifying exoplanet candidates. In the early days of exoplanet research, these candidates were then followed up with more detailed observations to confirm that the candidate was a true planet.
With the advent of Kepler, there are now way too many candidates to get follow-up observations on all of them. So now exoplanets are "validated" by calculating the probability of a false positive. I am not in this field, but it seems that the most common algorithm to do this is called VESPA [2]. The contribution these authors make is to use a machine learning system for the validation process, i.e., calculating the probability that the candidate is a false positive.
[1]: https://arxiv.org/abs/1712.05044
[2]: https://github.com/timothydmorton/VESPA
I don't math good, but this sounds like a "dial that goes to 11".
If you have a model to identify false positives, why not incorporate it into the original model to eliminate those positives? Or, if you really need observations, how can tweaking the model eliminate that?