DailyDirt: AI Image Recognition Is Still A Bit Buggy

from the urls-we-dig-up dept

Mon, Feb 2nd 2015 5:00pm — Michael Ho

As artificial intelligence gets more and more advanced, the differences between how computers recognize patterns and how humans do may become harder and harder to discern. However, it's obvious there are differences -- which might matter significantly if we're going to put these image recognition algorithms in control of autonomous cars or military threat detection systems. Check out a few of these image processing algorithms.

You can fool some of the people some of the time, but you can fool some AI all of the time. Apparently, it's possible to create self-training software that can fool state-of-the-art image recognition algorithms with images specifically evolved to generate false positive recognition (and that look nothing like the objects that were supposedly represented). It's almost like making a kind of Rorschach test for AI. [url]
Several tech companies are developing advanced image recognition systems: Baidu, Google, IBM, Yahoo, Facebook, Twitter, Dropbox, etc. Baidu's Deep Image team has recently claimed to be the top-ranked system, beating out the performance of Google's team in the 2014 ImageNet computer vision competition. [url]
When detectives in a show look at a photo and say "enhance" on a small part of an image, there really isn't any magic technology that can reconstruct reflections off a disco ball... or not yet, at least. SparkleVision is an image reconstruction algorithm that can unscramble images from some kinds of distorted reflections. Complex image processing is getting a lot better, but it's not quite as good as Hollywood makes it look. [url]

If you'd like to read more awesome and interesting stuff, check out this unrelated (but not entirely random!) Techdirt post via StumbleUpon.

Thank you for reading this Techdirt post. With so many things competing for everyone’s attention these days, we really appreciate you giving us your time. We work hard every day to put quality content out there for our community.

Techdirt is one of the few remaining truly independent media outlets. We do not have a giant corporation behind us, and we rely heavily on our community to support us, in an age when advertisers are increasingly uninterested in sponsoring small, independent sites — especially a site like ours that is unwilling to pull punches in its reporting and analysis.

While other websites have resorted to paywalls, registration requirements, and increasingly annoying/intrusive advertising, we have always kept Techdirt open and available to anyone. But in order to continue doing so, we need your support. We offer a variety of ways for our readers to support us, from direct donations to special subscriptions and cool merchandise — and every little bit helps. Thank you.

–The Techdirt Team

Filed Under: ai, artificial intelligence, autonomous cars, image processing, image recognition, imagenet computer vision competition, neural networks, pattern recognition, self-learning algorithms, sparklevision
Companies: baidu, dropbox, facebook, google, ibm, twitter, yahoo

8 Comments

If you liked this post, you may also be interested in...

Reader Comments

Subscribe: RSS

View by: Time | Thread

Anonymous Coward, 2 Feb 2015 @ 6:37pm

Computer vision is never going to be the same as human vision. Just like a computer will never write a play like Shakespeare.
[ link to this | view in chronology ]
- MrTroy (profile), 2 Feb 2015 @ 10:17pm
  
  Re:
  [citation needed]
  
  Never is a long time...
  [ link to this | view in chronology ]
- Ninja (profile), 3 Feb 2015 @ 1:06am
  
  Re:
  Ask a person in 1930 and they'd say that man was never gonna hit the moon, much less land a robot (a robot?!!!!!) in Mars. Give it time and it will happen.
  [ link to this | view in chronology ]
  - JoeCool (profile), 3 Feb 2015 @ 9:03am
    
    Re: Re:
    By 1930, science fiction was already well into stories of going to the moon or even Mars. Remember that Jules Verne wrote from 1863 to 1905, and wrote "From the Earth to the Moon" in 1865. So I'd say at least some people in 1930 would have said it was inevitable that man go to the moon, or put probes on Mars.
    [ link to this | view in chronology ]
Anonymous Coward, 2 Feb 2015 @ 8:24pm

Ha ha ha...
You can fool most of the people most of the time actually. Politics is very much proof of that.

However, resolving the problems with AI will not be easy. At best we would probably have to start every AI off at the level of a childish mind, kinda like humans are at birth. This would allow an AI to develop itself and learn what things mean. The problem with this will be the same as with humans. It opens the possibility that the AI will decide to hate humans as part of its learning. The only thing that might save us is a rule like Asimov's laws for example. However, if it is truly an AI, its ability to learn and undo any safeguards we place in it, will quickly grow beyond our abilities to manage.
[ link to this | view in chronology ]
CharlieBrown, 3 Feb 2015 @ 12:04am

http://xkcd.com/1444/
[ link to this | view in chronology ]
Ninja (profile), 3 Feb 2015 @ 1:09am

I like the fool other AI software mentioned. We should start running it collectively to fuck up the surveillance crew. As for real life we have masks and cover for the face. As long as we fight for the right of hiding them.
[ link to this | view in chronology ]
nasch (profile), 3 Feb 2015 @ 8:55am

SparkleVision
SparkleVision is an image reconstruction algorithm that can unscramble images from some kinds of distorted reflections.

This is not at all like enhancing a grainy smudge of a reflection into a crystal clear photo like you see in the movies, though it is pretty cool. SparkleVision has all the information it needs, and assembles it into an image. In the Hollywood version, they're coming up with information that isn't even in the original image.

And for your amusement: http://imagizer.imageshack.us/a/img705/4124/zoomcsiinvestigation.jpg
[ link to this | view in chronology ]