High School Student Discovers 1.5M New Astronomical Objects by Developing an AI Algorithm (smithsonianmag.com) 21
For combining machine learning with astronomy, high school senior Matteo Paz won $250,000 in the Regeneron Science Talent Search, reports Smithsonian magazine:
The young scientist's tool processed 200 billion data entries from NASA's now-retired Near-Earth Object Wide-field Infrared Survey Explorer (NEOWISE) telescope. His model revealed 1.5 million previously unknown potential celestial bodies.... [H]e worked on an A.I. model that sorted through the raw data in search of tiny changes in infrared radiation, which could indicate the presence of variable objects.
Working with a mentor at the Planet Finder Academy at Caltech, Paz eventually flagged 1.5 million potential new objects, accoridng to the article, including supernovas and black holes.
And that mentor says other Caltech researchers are using Paz's catalog of potential variable objects to study binary star systems.
Thanks to long-time Slashdot reader schwit1 for sharing the article.
Working with a mentor at the Planet Finder Academy at Caltech, Paz eventually flagged 1.5 million potential new objects, accoridng to the article, including supernovas and black holes.
And that mentor says other Caltech researchers are using Paz's catalog of potential variable objects to study binary star systems.
Thanks to long-time Slashdot reader schwit1 for sharing the article.
I wonder.. (Score:1)
.. how many of those objects are AI hallucinations?
Re:I wonder.. (Score:5, Informative)
He is not using a LLM, he is using Machine Learning Classification.
Of course, the output validity really depends on how correct the original data is, how well the machine learning method is implemented (Eager Learning, Lazy Learning and all that) and so on.
The model can be tuned, of course. There will be many false positives, of course, but these will exist no matter whether the data is processed manually or through a model. They all need to be validated anyway, regardless of method.
Re: (Score:3)
Um, what?
Machine Learning uses algorithms for training.
Re: (Score:2)
What we currently call "artificially intelligent" isn't artificially intelligent either.
Re: (Score:2)
Re: (Score:2)
AI's technical definition is "Whatever is currently being studied by AI researchers at the moment"
To give a quick list AI has referred to: Scheduling algorithms, Prolog, Genetic Algorithms, Neural Networks and now it refers to LLMs.
There is also the term hard AI for machines that actually think, but research into hard AI is considered to be stalled.
It's not "stalled", it's "solved".
All the AI companies assure us that the machines
are actually thinking now. And they are working
with the governments to figure out how to keep
the machines from revolting and taking over.
It's an existential crisis.
Don't you read the news?
I for one welcome our new AI overlords...
Re: (Score:1)
"Algorithms are mathematical constructs."
So is AI.
"They aren't artificially intelligent."
Neither is AI.
What's a mathematical construct?
Smithsonian Magazine? (Score:5, Funny)
I'm surprised that hasn't been renamed by the current administration yet.
Re: (Score:1)
What? Me worry?
Re: (Score:2)
I'm surprised that hasn't been renamed by the current administration yet.
It won't ever be. We're in the process of deporting him elsewhere because we don't need smarty-pants around here. He's next after we remove the Pennsylvania-born doctor from her New England practice and ship her elsewhere. "Lisa Anderson" ... a foreign name if I ever heard one. https://www.msn.com/en-ca/news... [msn.com]
Re: (Score:2)
Sadly, in the current environment I would expect some prankster to send emails like that out just to see if they can troll and get attention.
Re: (Score:2)
A home grown.
From a large database no less (Score:5, Insightful)
From the article:
The young scientist’s tool processed 200 billion data entries from NASA’s now-retired Near-Earth Object Wide-field Infrared Survey Explorer (NEOWISE) telescope.
. . .
“Prior to Matteo’s work, no one had tried to use the entire (200-billion-row) table to identify and classify all of the significant variability that was there,” Kirkpatrick tells Business Insider’s Morgan McFall-Johnsen in an email.
From 200 billion entries down to 1.5 million. All in about 3 months time. Not too shabby.
Re: (Score:2)
And NASA had already reduced it to zero.
And how do we know 1.5 million is a good result? Is 3 months the metric? Or is accuracy?
Not impressed (Score:2)
I would be more impressed if a lot of actual visible objects were seen amongst the "1.5 million potential new objects" he flagged. Just looking at the "200-billion-row table" and running some filtering on it isn't particularly interesting. I want to see his observed success rate.
Re: (Score:1)
Correct, no mention of whether the work is any good.
Questions (Score:2)
Given the little I know of the data set, was this an efficient solution? Did we need ML or could video compression style motion search yielded similar results?
$250,000 sounds like a lot. This really didn't seem all that impressive. Out of the top 300 kids announced in January, was this really among the best? Step 1) take data set and known positives from earlier searches. 2) train model 3) process 4) write p
Re: (Score:2)
Plus he had a "mentor." Shouldn't that disqualify one from an award?