Artificial Artificial Intelligence Tries To Track Down Steve Fossett
from the finding-a-needle-in-a-haystack dept
As the best thinkers are discussing the future of AI at the Singularity Summit, perhaps we shouldn’t just be looking at ways to create better computer artificial intelligence, but at ways to more efficiently make use of human intelligence that’s available all the time. For example, take a look at the technologies being used to harness human cognitive abilities to help in the search for Steve Fossett’s plane in the vast desert of Nevada. Searchers are leveraging Amazon’s Mechanical Turk community to quickly scan through Google Earth satellite imagery to flag areas where the plane might have crashed or eliminate barren areas where rescue pilots shouldn’t focus on. At the time of this post, there are still over 100,000 blocks to be looked over with each block representing a 278×278 sq ft. area. At what appears to be a scanning rate of about 10-20 square grids/sec by the Turk community, the entire area of interest could theoretically be searched in less than 5 hours — assuming that the system isn’t showing the same photo to multiple people (as appears to be the case). Contrast that to the (super)computing resources required to process an equivalent image of more than 8 billion pixels (img dimension(256×256) x (numOfImgs)100k), on top of the difficult task of defining the object-of-interest to a computer (what does a plane wreck look like to a computer?).
In this instance, using simple coordination mechanisms, human intelligence becomes an economic way to solve a hard problem — which is exactly the rationale behind Amazon’s Mechanical Turk. We have seen projects using other hooks and mechanisms to leverage human intelligence, like Recaptcha for OCRing books, and the ESP game for tagging images — and even the Techdirt Insight Community, which is bringing you this post. On a grander scale, Amazon’s Mechanical Turk is trying to be a platform for “artificial artificial intelligence”, though so far the success stories for MTurk have been minor. Even in this case, it’s not entirely clear how useful it is (or how they got the latest satellite imagery ready to go for this task). In fact, while there are fairly stunning reports that, in searching for Fossett, the remains of eight other plane crashes have been discovered — it doesn’t sound like any of them were found via Mechanical Turk and Google Earth. Still, with all this talk about mashing up web services and better artificial intelligence, perhaps it’s time we start thinking about more effective and efficient ways of leveraging human intelligence?
Filed Under: artificial intelligence, mechanical turk, steve fossett
Companies: amazon
Comments on “Artificial Artificial Intelligence Tries To Track Down Steve Fossett”
m-turk search for Fossett's plane
The search for Steve Fossett’s plane on m-turk, using small emailed satellite images is a suprisingly useful tool for a daunting task such as covering 10,000sq. miles of images.
After reviewing 400 of these small image squares on m-turk, my eyes did go a little buggy. I did, however, come across 15 or 20 hits, making me feel the effort was worthwhile.
While most of the “hits” were probably nothing, a few had real promise…Its a great cause, if not for Fossett, then to bring closure for the families of the numerous other lost airmen that suffered from the lack of notoriety or wealth this adventurer earned so much of. (or had the mis-fortune of crashing before the era of Google-Earth and M-Turk)
Re: m-turk search for Fossett's plane
I think this has a lot of promise for future web-based assistance in rescue recovery too. People can continue searching at night, during bad weather and never have to return to refuel.
They’ve made improvements to what is obviously a hastily created web page; today it takes less page navigation to review an image than yesterday, but it still fails to mention air-to-ground emergency signals, which Mr. Fossett would know to use if he was able. Web-bound searchers should know to look for an X or V or any other of the traditional signals that Fossett would create out of any available materials.
OK, some rich guy is lost and they end up finding 8 other crashes? If I were a relative of people killed in those crashes, I would be wondering why they could find them while looking for a rich guy but couldn’t find them when they crashed.
In response to your question about how the satellite images were prepared for mTurking: Wired reported that when Jim Gray dissappeared, satellites were re-routed to the search area, photos were taken, and then those photos were prepared and uploaded to mTurk. The same could have been done in this case.
Spam Detection
On the concept of using machines and Internet to harness the ability of humans:
I have been using a program that does something like this for Spam detection. Instead of using a machine to decide if a mail is spam or not, users of the software click on a button embedded in Outlook that flags spam back to the community. If enough humans say a message is spam, it gets filtered out for the rest of the community.
Turns out, people are far more accurate than machines for this task — although not perfect. People click on legit newsletters calling them spam, even though they could just unsub. So I’ve had some false positives.
Overall, it works very well, and I’ve been using it for a year or so. I have no commercial relationship with these guys, and I won’t even give you the link that gives me a “referral bonus”. It’s called Cloudmark, so if you’re interested in the concept, you can find it yourself.