Help a Mars Rover's AI Learn to Tell Rocks From Dirt (techcrunch.com) 18
Slashdot reader shirappu writes: For eight years now, the Mars Rover Curiosity has been exploring the surface of Mars. Even now, it's still exploring, and still getting upgrades. According to Tech Crunch, NASA is now looking to interested volunteers to help upgrade the rover's terrain-scanning AI systems by annotating image data of the planet itself.
"The problem is that while there are lots of ready-made data sets of images with faces, cats and cars labeled, there aren't many of the Martian surface annotated with different terrain types..." notes TechCrunch. "Improvements to the AI might let the rover tell not just where it can drive, but the likelihood of losing traction and other factors that could influence individual wheel placement."
shirappu continues: Volunteers go through a short tutorial after which they can label images to help the rover better understand the terrain on which it drives. The system is expected to be used in future planet rover robots, and the project marks an interesting example of open crowd-sourcing to improve machine learning systems, and how it is impacting technology even on other planets.
Click this link for the AI4Mars site link where people can volunteer.
"The problem is that while there are lots of ready-made data sets of images with faces, cats and cars labeled, there aren't many of the Martian surface annotated with different terrain types..." notes TechCrunch. "Improvements to the AI might let the rover tell not just where it can drive, but the likelihood of losing traction and other factors that could influence individual wheel placement."
shirappu continues: Volunteers go through a short tutorial after which they can label images to help the rover better understand the terrain on which it drives. The system is expected to be used in future planet rover robots, and the project marks an interesting example of open crowd-sourcing to improve machine learning systems, and how it is impacting technology even on other planets.
Click this link for the AI4Mars site link where people can volunteer.
Two more important things to distinguish (Score:1)
You gotta teach 'em right [youtu.be]
No need to do that! (Score:4, Insightful)
Yes (Score:2)
Collectively confer uniquely human abilities on machines.
What could possibly go wrong.
Re: (Score:2)
Exactly. Let the murder and mayhem to the experts. No need to have amateurs mucking up things.
Re: (Score:2)
Exactly. Let the murder and mayhem to the experts. No need to have amateurs mucking up things.
The "experts" are not capable of accomplishing this without our help. That is (part of) my point.
I get it, i'm being overly dramatic. But i feel that we should know everything about how a computer works, and a computer should no little to nothing about how we work. Or more specifically about how we think.
Re: (Score:2)
What could possibly go wrong.
Actually, I was thinking the same thing - but in the context of "asking on the Internet for volunteers with labeling stuff, what could possible go wrong".
Re: (Score:1)
Collectively confer uniquely human abilities on machines.
That's an odd thing to say.
Object detection and identification not uniquely human. Most every creature with eyes can do this ranging between far simpler and far better than humans can.
We have far less control over the Earths other life forms than the machines we make, if that's what you're worried about.
Yes I get it, object identification is a baby step down the path towards thinking.
But on the other hand we've already started down this road almost a hundred years back and there is no rebottling that.
I'd m
"Click this link for the AI4Mars site link" (Score:2)
Sooo, data mining? (Score:3, Insightful)
This exercise is leaning heavily on the 'artificial' and not much on the 'intelligence'. At some point the idea of data mining was rebranded 'AI'. There's nothing intelligent about this.
If a toddler tries to eat a bee, you better believe it's never going to try that again, and won't ever try to eat anything else insect-like. And you don't even have to show it thousands of pictures of insects so it 'learns' what they are. It's inherent in intelligence, which this rover does not have. Nor does anything else, for that matter.
Re:Sooo, data mining? (Score:4, Insightful)
If a toddler tries to eat a bee, you better believe it's never going to try that again, and won't ever try to eat anything else insect-like. And you don't even have to show it thousands of pictures of insects so it 'learns' what they are. It's inherent in intelligence, which this rover does not have.
Even a toddler has had years worth of dual high resolution video feeds, experiencing motion, ego-motion and interaction with objects in a 3D space. For an AI those pictures are literally all it knows about existence, the equivalent of a toddler that's been completely paralyzed since birth, blind on one eye and sedated while cared for so the only visual stimuli it's experienced ever is that slideshow. It's not an apples to apples comparison.
We've recently been making great strides in unsupervised pre-training. Basically we let the model study a ton of unlabeled data first, then it turns out we need only a few labeled examples/counterexamples to identify a class. The reason you need counters is that you don't know if the class you're trying to find is "insect" or "bee" so we also need some information on what's not bees.
Of course that's still only a fraction of the human mind, but it's hard sometimes to distinguish between reasoning ability and information compression. The more you're able to reason about the world, the more you can derive, the more compact your representation of the world is. If that was a scientist turning raw observations into a formula we'd call that intelligence, if it's a computer creating a compact representation through decomposing/deconstructing a complex result it's not.
Click on terrain type first (Score:2)
I messed up the first few images I submitted, as I was drying polygons then selecting the ground type I thought they were... the color of the line you are drawing corresponds to the terrain type in the side selector.
Just a helpful tip for those that want to try it out!
"AI" (Score:2)
Help Science captcha (Score:1)
I wonder how much work would it take to build a captcha system that helps science projects by building public, annotated datasets. I would love to see this being used in the wild instead the private datamining done by Google.
So, Where To Volunteer? (Score:2)
This is all verra nice, I even found the web site. But be damned if I could find any place in which to volunteer for the project. I've done lots of things like this (looking for Genghis Khan's grave in Mongolia, annotating great Lunar and Mars images, identifying whale tales, etc.). This would be interesting, one would think, if only one could figure how to get started.