Slashdot Mirror
DARPA's Cortically-Coupled Computer Vision System
BluePariah writes "Wired News has an article on a 'cortically coupled computer vision' system being developed at Columbia University and funded by the ever-curious folks at DARPA. Essentially, it uses the extremely powerful visual recognition ability of the human brain and couples it with a computer's raw processing power to allow a user wearing an EEG cap to filter through scores of digital images at high-speed and pick out something of interest. This has applications in military intelligence, face-recognition, anti-terrorism, and hunting down replicants."
It's 10 times faster (RTFA).
...and you've eaten your pen. simply stunning.
Replicants is a reference to Blade Runner. A move by Ridley Scott.
The IMDB link: http://www.imdb.com/title/tt0083658/
The move is based on the work of Phillip K. Dick. It also stars Harrison Ford in his least favorite role.
We could have a whole class of people created in test tubes, deprived of meaningful human contact and trained just to look at thousands of images per minute, all day every day.
Why use test tube babies when you can just use Slashdotters?
The enemies of Democracy are
One of the basic tasks our visual system is much, much better at executing than computers is visual search. The basic 'experiment' is that you are asked either a question like "Is there a red car in this picture?" (natural images) or "Are all the lines the same orientation?" (more traditional psychophysics). Then images are displayed, and our response time is recorded. Early experiments in the visual search paradigm appeared to show that there was two classes of search stimuli: those that 'pop-out' and those that require incremental search. The difference is that in pop-out conditions, increasing the number of elements in the image does not increase search time, while in incremental it does at XXms/element... and generally it takes about twice as long for us to respond if there is no positive element.
One main theory on how our brain does this, Feature Integration Theory by Anne Treisman (or similar but more recent, Guided Search by Jeremy Wolfe), which many computer vision algorithms try to copy, asserts that there are various feature maps for certain quantities like color, orientation, depth, spatial scale, etc. These are combined into a saliency map which is a weighted average of the feature maps. Things pop-out when the target has high salience compare to the background, for example it's easy to find the red T in a background of blue T's, but not so easy to find the red L in a background of red T's and blue L's.
Now, it appears from the article, and what little they say on the Lab webpage, that they are trying to measure EEG responses (which are quite crude) during rapid serial search tasks, in order to prime a computer vision object recognition system, which is then only run on those images human's appear to find sufficintly salient when they see them. This saves the time of a person actually having to search and make a decision about an image, while utilizing the visual systems incredibly powerful early 'pre-attentive' form & object binding resources.
If there is a sufficiently high signal from the EEG to do that after say, 100ms display times, then I think this could be useful for certain types of search task. However, due to the time courses present in most visual search experiments, the fact that it's not totally apparent how efficient certain parts of our saliency system actually are (check our Jeremy Wolfe's reviews for more data), I'm totally unconvinced that this type of system will give you a sufficent signal to noise ratio to be worth using for anything. This is especially true because of another perceptual phenomenon in search, which is that your error rate basically shoots up exponentially as the probability of a positive goes down. This is to say, in an experiment where a normal observer would have a 99% accuracy rate with 50% of the images containing the target, this drops to 60% accuracy for 10% target positive, and only 30% accuracy at 1% target positive (numbers fudged, but ballpark, since I'm too lazy to look them up). If this has its roots in insufficient priming in early vision, for example, then this entire scheme flops just as badly as using a human for tasks like finding the bomb in the x-ray image of the suitcase... and we haven't even started to get into issues of the person not actually looking at the image because they're bored, etc.
As it is, DARPA is spending a mere 758k, which is chump change for them, and there's a decent chance that it'll work in certain specific but useful circumstances which may warrant the research.