Slashdot Mirror
AI Astronomer Aids Effort To Analyze Galaxies
kkleiner writes "Scientists are teaching an artificial intelligence how to classify galaxies imaged by telescopes like the Hubble. Manda Banerji at the University of Cambridge, along with researchers at University College London, Johns Hopkins, and elsewhere, has succeeded in getting the program to agree with human analysis at an impressive rate of more than 90%. Banerji used data from Galaxy Zoo, a massive online project that has used more than 250,000 volunteers to analyze more than 60 million galaxies. The new automated astronomer will help with even larger analytical projects on the horizon, taking care of trivial classifications and leaving the tough cases to humans."
I wonder how this program compares to a human doing the same job If given the same "training" I wonder how many humans would get a 90% agreement rate looking at the same data.
I would think with a name like Al Astronmer your career choices would be limiting. I guess I was right.
If it is making a "guess as to the most likely classification", it sounds like there is a measure of confidence. Perhaps the system is capable of presenting questionable cases to human experts.
Insert self-referential sig here.
The paper discussing this work is http://arxiv.org/abs/0908.2033. They appear to be using a pretty standard neural network approach (disclaimer: I don't have much background in neural nets at all. I'm just going off of how they were described in the last class I took that discussed them.) This is part of a very general pattern where programs have done a lot of work that we would think could only be done by people. Other examples include the computerized proof of the Robbins conjecturehttp://en.wikipedia.org/wiki/Robbins_conjecture. TFA lists a few examples as well which are in more applied areas.
Let's say the system of scientific paradigms and their rise and fall is about finding outliers and surprising results that cannot easily be explained by current models.
If this is the case, then using a statistical system to classify observations has the danger that these "outliers" simply get classified in existing categories and whatever abnormality they represent thereby ignored.
An "AI researcher" must therefore have very explicit programming to set anything with even the slightest degree of abnormality aside for human evaluation. If it's set to classify anything and everything according to preset rules, it's actually mostly destructive to good science.
I wonder if it has any Tauntauns!
Careful What You Wish For....
R. Daneel Olivaw
"Kill 'em all and let Root sort 'em out"
If this technologies works for classifying galaxies, perhaps next we could put it to work classifying porn on the web!
I've abandoned my search for truth; now I'm just looking for some useful delusions.
So Galaxy Zoo doesn't need me anymore? That is the one activity where I was contributing to science to benefit all mankind.
Oh well, I guess I'll go back to trying to beat Mario 64 or something equally pointless....
PRO:
Using neural networks allows for graceful degradation when classifying galaxies by indicating to what degree it believes this galaxy is similar to other galaxies of this type (that it has been trained on). A threshold can be set so that if confidence falls below this threshold, the image is flagged for human intervention.
CONS:
Neural nets are largely black boxes. They use learned statistical relationships to classify images, but they're unable to provide an explanation as to why they made the decision that they did.
Maybe I'm missing something, but what exactly is the point of going to all this effort to classify far-away galaxies? I can understand astronomers wanting to examine closer galaxies and see how they work and interact and all, but surely all the galaxies that are close enough for us to be able to see that much detail have already been known for some time, and are classified, and studied in far more detail than just classifying what kind of galaxy they are (spiral, barred spiral, elliptical, etc.). What good does it do to classify millions of galaxies that you can barely see?
Wouldn't it be more useful to expend this effort on observing objects that are closer to us, and which we can see in much better detail, such as various stars and nebulae within our own galaxy? Or in trying to find more evidence of exoplanets and determine if there's life on them?
After all, it might be possible for us to send ships to nearby stars (like Alpha Centauri) eventually, and examine exoplanets there first-hand. Such a voyage may take a few decades, but it's doable. But unless we come up with a way of creating wormholes or something else that allows us to teleport vast distances instantly, there's no way we'll be able to visit far-off galaxies, ever. They're simply too far away. Even in uber-optimistic Star Trek, with warp speed travel, humans still never left the Milky Way galaxy (except for one episode of TNG with the Traveler, and even that was a nearby galaxy), and stayed mostly within one sector. The whole underlying plot of Voyager was that they were on the other side of the Milky Way, and it'd take hundreds of years to get back at top warp speed.
Pictures showing galaxies that are billions of light-years away make nice posters, but it seems totally pointless to put too much effort into these things, when there's so much we don't know about the stuff inside our own galaxy.
I'm surprised they're just now getting around to this. It's a straightforward pattern classification problem, and there is a huge set of training examples to be used for training a neural network or other Learning Classifier System technologies.
Sheesh, evil *and* a jerk. -- Jade
Artificial Agent Aids Astral Analysis
It breaks my pluginses, my precious!
This makes me very happy on one level and very sad on another.
At the amateur end, the advances in technology have meant that what use to be done by a professional with mind blowingly expensive equipment or what was not at all possible because it hadn't been invented can now be done by a dedicated amateur with a reasonable but largish hobby budget. For the amount of money some spend on recreational vehicles and holiday homes an amateur can now do spectroscopy, deep imaging, even adaptive optics. It's not open to everyone - you need to have good circumstances - a job that both pays well and puts somewhere within driving distance from less light polluted skies. But it can be done..
On the other hand the technology has meant at the professional end what was cutting edge a few decades ago is now obsolete and not an area of interest. What use to be done on an individual basis is being taken over by surveys etc.
What this means is that there are only a handful of ways in which an amateur can contribute real science. Mostly this revolves around tasks that are either considered not important enough to dedicate professional resources to, or areas that aren't easily automated or taken over by sky surveys. Stuff like variable star observing and galaxy zoo. Now those areas are dwindling too as the automation gets better. The amateurs have done a wonderful job especially with variable star observing - with records extending back hundreds of years - this is data that professionals did not have the time to gather themselves nor the technology to gather in bulk....until now. With projects like Pan-STARRS coming online, how long will this be a useful way to contribute? The records will improve but the opportunity to contribute will dwindle.
Also there's the nagging feeling that automation, while good for most things, can't completely replace human curiosity. For the Galaxy Zoo project, I wonder if this method would detect anomalous objects like Hanny's Voorwerp. That was only discovered because a schoolteacher bothered to ask "what the heck is that smudge" instead of simply dismissing it as a photographic error. This led to Galaxy Zoo 2 including a button to report such objects.
So overall I think we'll make great progress - much greater than human only efforts - but I do wonder what discoveries we'll miss.
These posts express my own personal views, not those of my employer
How many galaxies are there in the Milky Way?
Todos mis movimientos están friamente calculados
Sorry to burst the bubble, but automatic classification of galaxies from sky survey data using machine learning techniques was accomplished in the early '90s by the SKICAT system developed at JPL and Caltech. http://adsabs.harvard.edu//abs/1995PASP..107.1243W is a good overview of the system and its accomplishments as of 1995.
There's no sense in being precise when you don't even know what you're talking about. -- John von Neumann